Speculation: open models is what will kill Anthropic and OpenAI. Hyperscalers can run the models without a licensing fee. Apple can make them smaller and put them on the device.
The frontier models are an edge and a liability. They're astronomically expensive to train. Without them, their models will fade into obscurity. Their marketing depends on people believing the models are meaningfully different, as people have sweatily argued on this forum. Personally, I'm not convinced there's much of a difference between these models at this point. The harness is what takes these random and hallucinogenic models and make them into something deterministic and useful.
The outcome is plausible. Open weights models though look like a tactical more than a principled play by Chinese companies to overcome the disadvantage and difficulties to access western markets. Two issues:
1. If market conditions change they might decide to close down like Meta did.
2. If as you said models keep getting more expensive to train, is an open weights strategy financially sustainable?
I think (and have heard) the Chinese govt is also interested in trying to embarrass the US as well by showing off their capabilities (so there's a political angle, too)
All countries see AI as a geopolitical concern. An open source strategy is smart. It’s effective and also builds good will / soft power. Many see Chinese AI companies as the good guys and root for them. I’m pro open source and also celebrate but still aware this is most likely just the means to an end.
I used to work at Mozilla. I think we are missing a player in the market with a more principled open source approach.
And who might that be, I wonder? The company that seems more worried about executive bonuses than to adhere to its mission, for at least a decade now? Perhaps a certain one whose name starts with M and ends with ozilla?
A lot of things have to come together: talent, capital that is not looking to maximize returns, ambition to compete in the marketplace. At peak Mozilla it was clear to everyone that we needed an undeniably successful product to be able to influence the market. Advocacy alone is not very effective.
That's probably pretty likely, but if we're honest, are LLMs built and funded by a hostile Chinese authoritarian regime any more dangerous or harmful than LLMs built and funded by a hostile American authoritarian regime?
China absolutely does not have my best interests at heart, but America's technofascism is probably more immediately dangerous and harmful. Americans genuinely have more to fear from America than China at this point.
One big difference is that the US doesn't have much public record of using state espionage resources to steal industrial secrets to give domestic industry a leg up in the market, whereas CCP very much does.
And while I use and love the open weight models, it seems likely to be pretty hard to prove conclusively that they have no reinforcement-learning-trained proclivity towards curling specific URLs if the topic happens to be some very specific thing that the government is interested in, when used to drive agents.
So it depends who you are. If you're a company, you might have something to fear. If you're a private citizen, maybe not so much.
I don’t think US companies could ever fully trust models sponsored by foreign adversaries. How would you ever verify they are safe?
Provenance of training set is going to be critical. It’s largely ignored now, but the first time a malicious exfiltration occurs it will go off as if a bomb was detonated.
You're moving the goal post here. The parent never mentioned anything about the Chinese models themselves being dangerous or harmful; that's a totally different topic.
The point they were making was that the most successful open models- those coming out of China- are made my companies that are using those open models to get exposure in Western markets. The goal is to undercut the Western dominant players, not out of any particular "open source" philosophy, so it wouldn't be wise to expect them to continue providing open models long term.
I think the term 'techno' in 'technofascism' is doing the work here, because--just as you claim that historical fascism is the idolatry of the state--technofascism is the idolatry of technology and intelligence. Modern accelerationism, as espoused today by people like Marc Andreessen in his Techno-Optimist Manifesto, is really just another rehash of Italian Futurism, which was closely intertwined with Italian Fascism and one of its intellectual foundations. It was, essentially, a progressive ideology.
The idolatry of the state does not vanish, but rather gets transformed and re-imagined as a technology, e.g. network states, platform governance, companies that function as sovereign entities. The idolatry shifts from the nation-state to the infrastructure that replaces it. Also, you don't need to look far to answer your own question about corporatism, as you defined it yourself. Now, who's currently moving between Silicon Valley boardrooms and government offices?
Fascism is not merely idolatry of the state. That is simply its ideological mask, its method for currying the favor of its base. To say that this is what fascism is at a fundamental level is to mistake its superstructure for its base. But how is Fascism structured? How does society reproduce itself? Is there a fascist mode of production? No. The state mobilizes its population so as to target scapegoat(s) while simultaneously expanding in order to secure the spoils of war. Without scapegoats, there is no Fascism. And a state requires social cohesion to mobilize its base. Thus without cohesion there is no Fascism. At minimum, social cohesion requires the ability to sustain and motivate a critical mass of the population. Thus Fascism requires the continued maintenance of the capitalist mode of production in order to persist, but it simultaneously undermines capitalism due to A. the continued need for war machines and B. the need to punish a scapegoat, which both siphon value from the rest of the economy, thus dooming the entire enterprise. This is why fascism burns out and leads to doom for all parties involved wherever it has been tried. Any definition of fascism that ignores this is historically inaccurate and practically useless.
Open models are probably also comparatively astronomically expensive to train - just less so than the frontier models because they’re somewhat smaller, +/- the creators are more incentivised to focus on getting more from less compute because they’re have to, +/- they rely on distillation of the frontier models and this is more efficient.
But efficiencies aside; creation of open models still requires a lot of money and compute from a large organisation which is willing to accept zero return for that spend. This largesse is unlikely to continue forever; so the question is which will crack first, the frontier models’ business model or the fast followers’ generosity?
Yes, the problem with comparing open models to open source is that open source requires humans to volunteer their time. Open models requires humans to volunteer their money.
These two types of contributions have very different behavioral profiles, and it doesn't obviously follow that the historical success of getting people to collaborate socially on building software for fun and for the benefit of the community will translate in any meaningful way to the necessity of being able to raise enormous amounts of money to pay for enormous amounts of electricity.
The biggest hurdle is whether humans volunteer their expertise. Not time or money. We need top talent to make the open models. Sponsorship is plentiful. Open source volunteers are less critical with LLM doing the grunt work. Its about talent contributing to the open
This is not my experience working in the field the last 8 years. There is not a dearth of talented researchers, engineers, etc. who are willing to contribute to open models. Just look at the ecosystem generally, from academia to industry. So much research still happens in the open, and the open source community in ML is still massive and energetic.
Training a new frontier model just costs a lot of money and there are very few companies for whom it makes sense to do so, and even fewer orgs who are just going to gift those kinds of resources to an open source initiative.
> open source requires humans to volunteer their time
Your idealistic of open source may require that, but in practice a huge part of open source is commercial and a large chunk of that is low on collaboration (across vendor boundaries).
Technically open source requires some amount of monetary volunteering, it's just that the electricity to run a code editor and compile (most) open source code bases is within hobby budget for most people.
I don't think it needs to be framed purely as generosity. You just need a sufficiently self-interested actor that sees open ecosystems as a necessary part of reducing their own risk profile, relative to the alternative of complete reliance of a third-party business that can take an exorbitant cut and/or Sherlock them at any time.
Valve and SteamOS are a good example of what this idea looks like in practice. (Though they may also illustrate a third thing you need: a privately-run company, that has enough profit, and enough commitment from leadership to the company's vision, that they can make long-term bets without having to eventually bow to investors seeking short-term gains.)
> You just need a sufficiently self-interested actor that sees open ecosystems as a necessary part of reducing their own risk profile, relative to the alternative of complete reliance of a third-party business that can take an exorbitant cut and/or Sherlock them at any time.
This would be an argument for an organisation developing its own model; but not per se for releasing the weights openly.
The possible explanations (I'm aware of, which overlap somewhat) for spending large amounts of money on models then releasing them for free (i.e. the current Chinese approach) are soft power, marketing for a future paid model business (i.e. competing with the US models for customers and mindshare during the time you can't compete directly at the bleeding edge), and/or a geopolitical move to diminish the value of the US's frontier model companies.
I acutally see the reduced burden that comes with actually sharing resources.
To extend the previous analogy: Valve didn't make desktop Linux viable on their own. A lot of it is owed to another self-interested actor -- Google -- through the reduced need for dedicated desktop apps (largely pushed by Chrome), and various enhancements to wireless and power management that were necessary to make it a viable mobile platform (directly benefiting Android/ChromeOS, but then spreading out to laptops and mobile devices in general, including handhelds like the Steam Deck).
You can see it on a smaller scale in ecosystems like Android -- where handset makers regularly contribute features from their UI skins upstream, so that they no longer need to spend engineering resources maintaining distinct versions of theming engines/notification badges/multiwindow/various other stuff.
On a related note this is an argument for open source models, not just open weights. I think a lot of the diminishing returns relate to the opaque nature of most models.
I’m not exactly sure on the “how” but it only makes logical sense for (non-AI) companies to band together to fund the training of a shared model. Apple is a great example, AI is not their core business but they still require it.
The only thing that took us down a different path is the vast sums of VC funding pumped into the AI companies.
There's a reason we let companies specialize in some kind of service and buy it from them.
LLMs aren't looking like they'll be highly differentiated like software, so their market will probably be competitive. What negates the main reason Open Source software exists.
LLM training doesn't carry the same NIH risks that normal internal software bloat does. They are relatively simple to setup training for and analysis of accuracy/recall can be automated.
This leaves the price differential between a private third party and an internal initiative as barely more than the cost to train the model[1] - perhaps that's where we'll end up, a centrally trained model will represent an economy of scale that can leverage that difference into a margin it can profit off of but your business being purely profit driven by that training expenditure seems like a ridiculously thin margin.
So where does that leave the AI companies? If their LLMs are off the shelf-once built products they have a strong advantage for casual low usage but enterprise customers will have a huge cost incentive to roll their own - if the LLMs require continuous retraining and the frontier keeps moving then enterprise customers will find a packaged service more attractive and likely continue to subscribe for more accuracy but casual low usage will likely shift towards "good enough" models. It seems inevitable that they'll lose half the market and it seems difficult to discern their long term profitability[2].
1. Costs can, I think, reasonably be reduced to hardware depreciation and energy - if trends continue with cloud resource availability (it's possible this won't be the case if large compute providers start pulling resources offline to build a moat but I think they'd likely prefer the reliable compute income over model income which has several other competitive weaknesses). Hardware depreciation would normally be pretty negligible and equal across different training entities, right now we have a chip shortage but given the demand that can't last too long so I'd consider hardware to be fungible - and energy is entirely fungible - they're both hard to moat.
2. Outside of AGI, who knows if AGI will be or what even counts for it at this point - but I think if AGI isn't a doomsday scenario we fall back to one of the two above scenarios - either the frontier is ever moving and they can retain enterprise customers or the frontier seizes up and everyone can just use an off the shelf offering. In either scenario they don't have a lot of moat to deal with for their products unless they can restrict compute which is why Alphabet, AWS and MSFT are the only players I could see realistically coming out of this as an AI vendor winner and I'm not even certain if it'd be a good idea for them if it'd hamstring their cloud profitability.
A fair amount of ML/AI innovations came out of the market in general. Neural networks are a useful tool to solve a variety of problems... LLMs specifically were a more recent interesting market to develop but I've yet to see anything that could give a market player a real competitive advantage. It feels like we just invented a new hammer and now that we know how to build it it isn't that hard to build one yourself. The all purpose hammers are, of course, unreasonable to build - but those don't seem to be that useful. I don't really need Claude to be able to generate sonnets when I'm programming so I think specialization is the place we'll see genuine markets form.
The company was VC-funded as a search engine but by the time they made significant investments in AI (DeepMind etc) they'd been a publicly held company earning multiple billions a year from advertising for a decade.
Google bought DeepMind and their other major AI acquisitions. Public companies make corporate venture investments for very different reasons than LP-backed VCs. They do early-stage investments to search for emerging players they can buy as soon as possible or to gain market intelligence on trends. They do later stage investments to help grow future vendors or customers and sometimes to foster ecosystems that form their competitive moat.
But if they think it's important to their core business, corporations don't want to invest, they want to buy. Source: I used to be involved in corporate venture investing at a top 10 valley tech leader.
If your definition of VC is "Literally anything that requires a long term investment of money" then sure, but I think most people mean something different than that.
Historically speaking a lot of inventions have come about without things like VC investment. Either way, there’s probably little point in debating it, just because VC funded companies control the market now doesn’t mean they should indefinitely.
My (unverified) AI research claimed generally Chinese models are cheaper to train because Chinese data scientists are cheaper to hire and they're also under more pressure to optimize training cost due to limited hardware availability
Seemed believable but not sure where that's true
Chinese AI companies are generally smaller tho and the models they're releasing are also smaller (I think estimates put OpenAI and Anthropic SOTA into trillions of parameters)
In the US and Europe its extremely expensive indeed. In China it is much lower and expected to be much lower in the future when China can bake their own high end chips. You know China is focussed on optimizing production and manufacturing. Unlike the west that primarily focus on regulations that make things super expensive.
How does it work if people flock to open models but they're too expensive to train? What is the financial incentive to do so?
I seem to understand open models are mostly coming from China, and the benefit of training and releasing them for 'free' is a powerful geopolitical weapon against the Western/US economy that at this point depends on OpenAI & co. to succeed.
If we'd been sharing all along (as we should have been), we probably would have gotten even further along in the development of the tech.
Think of everything we could do if every researcher on the planet had first class access to the frontier. No academic fallback models. No crude API access. No limits, but direct access to the weights and the ability to lobotomize, splice, and dice.
We could pour intelligence from one container to the next without paying a tax or wearing a blindfold. All without spilling a drop.
> releasing them for 'free' is a powerful geopolitical weapon...
I agree that, currently, the Chinese govt is not only allowing but tacitly encouraging open weight model releases. However, I don't see it as an attack. I think it's more of a strategic delaying move to slow the revenue to frontier models while China works to catch up. This strategy will likely change over time.
> Will the West make open models illegal?
In the U.S. this seems highly unlikely due to the current administration's generally laissez-faire approach to tech as well as the U.S. constitution severely limiting the government's latitude to constrain economic activity.
As we saw with the temporary Mythos restriction, there are legal mechanisms to limit tech on certain grounds, but over time such limits are subject to close judicial and constitutional review. The Mythos embargo was also likely driven in part by the administration's anger at Anthropic for choosing to block the DoD from using their products for mass domestic surveillance and warfighting. I doubt we'll see any meaningful restrictions on OAI or other large companies. It'll be nearly 3 years before a different admin is in office and could enact serious limits and by then it will be too late for fundamental bans.
There are vested interests in most governments, such as intelligence agencies, law enforcement and the military, who would prefer to restrict some AI from broad use. As we saw with strong encryption, they'll only be able to delay and constrain, not stop, such a broadly useful dual-use tech. The geopolitical, economic, competitive and civil liberty interests are similar between strong encryption and AI, setting up a similar game theory dynamic. While it can be argued AI poses some potential danger, the specter of any such threat is abstract and not immediate.
On the other hand, the tech is obviously too economically essential and competitively vital to risk 'falling behind'. While there will certainly be attempts to ban, limit or constrain AI, the well-funded, highly organized commercial interests and civil libertarians will deploy lobbying, legal challenges and public opinion to ultimately prevail.
Aren't these the same guys who won't even let us have Chinese cars?
I'm not as confident as you that they will keep allowing us access to technology as strategic as AI models out of China and elsewhere that undercut US models in the market.
To everyone reading, download open models from anywhere as soon as they are released. You really have no guarantee at all that access to those models won't be cut off in the future with the stroke of a President's pen. Those downloads are your insurance policy. You'll always be able to access whatever you've already downloaded.
I would definitely pay a monthly subscription to help fund a non profit compete with Anthropic and OpenAI. I already pay subscriptions for myself and 2+ other people. It's a non brainer to be able to pay for the training of better models that I can then run myself for many more. I hope someone starts this, I think this model would work. I'd start it today if I had the team and initial capital to bootstrap the infr. I know VCs won't fund it, but we definitely will, enthusiastically and continually.
I'm a bit skeptical of the token cost/ROI for all models, but sunk costs are sunk.
It has the feel of self-improving super-intelligence or bust to me.
If you get that, the frontier model(s) run away with a faster exponential.
It's a bit like semi with Moore's Law with silicon, GaAs could never catch up. If you don't get it, the fast followers crush the high investment and there's no moat. Not like they can enforce copyright!
There's a world where frontier models run away with a faster exponential and still go bust due to being outcompeted on efficiency.
There's a point past which "intelligence" stops mattering as much, and IMO we're already there.
Consider which would be more useful (and profitable for its creator): a model that is 3x/5x/10x as "intelligent" as Mythos, for whatever your favorite yardstick of intelligence is? Or a model that is as "intelligent" as Opus 4.5, but can run at reasonable speed on a typical consumer laptop/cell phone?
not really a feeling; if you listen to ed zitron and strip out the vitriol, you still get the fact that the VCs are looking for some 5 trillion dollars in 5 years.
The onlly way that happens is if America turns into zimbawe.
I've driven in a LOT of the USA. Sure Chicago, NYC, DC, LA, LV... They're all built up and feel modern.
Try driving anywhere in the Midwest outside of the big cities. Dilapidated carcas buildings everywhere. Urban and rural blight. Only jobs are low paying service work. Its bleak. Like, really bad poverty as a disease bleak.
And its crazy watching it too. They're ignorant (involuntatily), poor, and trapped. And democrats only seem to care about special interest of the week, so these areas vote republican.
I don't have a solution btw. Just something I've seen growing in the last 25 years. And its getting worse, not better.
At risk of greatly oversimplifying American politics, it's truly impressive how good Democrats are at shooting themselves in the foot.
That's why I personally believe Sanders and Mamdani have found so much success with the working class; they keep themselves separate from the Culture War slugfest that mainline Democrats either voluntarily engage in or let Republicans drag them into.
IMO the vast majority of those "culture war" issues (LGBT freedoms, etc) are incredibly important, but to the average poor rural American it feels incredibly distant from their day to day. I can't put my finger on it exactly but Democrats have a tendency to message on those issues in ways that are either counter-productive or get soundbit saying something moronic. So when Fox News and whatnot say that Democrats are prioritizing other groups over them and message on it day after day, it's not hard to see why that propaganda becomes effective.
That's not to say that Sanders/Mamdani/etc don't message on social issues, they obviously do, but they are somehow effective at not alienating voters who may otherwise latch onto that in a negative way.
I don't have a good solution. Just my observations.
I'm not sure I understand what part of my post this is in reference to. I agree that the democratic party gets a _lot_ of money. Are you contrasting to something in my post, or agreeing?
Appreciate clarification, I could also just be dumb today
As a Canadian driving through Michigan/Ohio/Upstate New York and places look run down and like they peaked in the 60s-80s and its been down hill since.
When I was younger and it was a new thing it was quite a shock since American media obviously doesnt portray it that way. It is quite a contrast to even how the more run down parts of Canada look.
My prediction is that hardware costs will make open source models impractical for the foreseeable future.
Yes, tinkerers and enthusiasts will continue to make use of them, but frontier companies will maintain near total dominance because they will be the only ones with access to the hardware.
Meta is selling their now excess compute, other compute has been on the market for a while. The current hardware cost bubble is temporary, especially once people are forced to pay the real inference price instead of majorly subsidized subscriptions.
yep, all those coders paying $200-$500 per day to use claude once subsidy ends will be seriously rethinking how much they really want to vibe-code "rewriting X in rust". Helping people write word docs, recipes, and emails isn't going to justify $15K per month subscriptions either.
Not really. Models that can run on a $250 card may be able to produce a decent word doc sometimes, but the fidelity of such models is so poor that it's no longer very attractive for general use on those types of tasks.
The subscription is certainly subsidized but its no where near 100x cheaper than API prices.
Heck, most large enterprise moved to usage based billing and are still happily paying for it. They are force multipliers for your top talent, and when a top engineer is being paid $500k a year, doubling their output for $500/day is a no brainer.
Lowering the cost of hardware still won't solve the issue. HBM and DDR5 was never cheap, even before the shortages, so selling a full inference system is beyond the acceptable price range for most casual customers.
We're going to see Apple and Google compete over services and AI/OS integration instead, it will probably be years before your OEM takes local models seriously.
> But how many years do you guess? I personally do not think it will take even 10 years for the situation to be commonplace.
IMO it won't be possible for the foreseeable future. There's essentially zero possibility that phones will gain the hardware capacity to run today's Kimi, so the only other alternative is to squeeze the power of today's Kimi into something that can fit on a smartphone, which also seems fairly unlikely considering the current rate of progress.
A bold prediction. Phones have gigabytes today. There are famous laws of growth that put terabytes at just a few years away - perhaps 10 isn't too bad an estimate?
> I personally do not think it will take even 10 years for the situation to be commonplace.
Do you personally remember how far smartphones progressed in the past 10 years? It's not as long a time as you think it is, the limits of what a smartphone GPU is capable of did not substantially change in that time. Nor did the amount of onboard RAM that we include in the package. This is true even for Nvidia's ARM SOCs, frankly.
Apple, Microsoft and Google all eventually want to enforce OS-level lock-in for the most profitable AI services (eg. their own). It's much more attainable and profitable to use that lock-in to sell you exclusive service integration, the local AI revolution probably won't begin on their hardware.
Only if you narrowly define AI as talking with a chatbot. Today, on an iPhone, there are a number of features that only work due to some sort of a model running locally. OCR and intelligent object selection in a photos, and summarization of texts are local models running on the iPhone hardware, they're just not a chatbot. Yes, there's also cloud backing various features, but the idea of running models locally isn't foreign to Apple. On Google's side, the Pixel 10 Pro is powerful enough to run quantized local chatbot models locally today. Local translate is a model, and has been for a while. Both corporations are going to sell whatever customers are willing to pay for, in money or via ads, and if local models get good enough that customers actually are willing to pay for it, I have no doubt that it's Apple and Google will go that direction. It's Google that's releasing Gemma models for download, and they're a big enough organization that the left hand doesn't know what the right hand is doing, so the one conspiracy theory that they'll never do local models because they only want to profit off hosted models is too simplistic.
For the near future it seems that the new models will consume whatever improved hardware capacity we have. Competing with that is challenging, but I also think there will be strong economic incentives towards cheaper but adequate models on other providers.
I don't think we'll see home users being able to match even the low end clouds for a long time.
Longer term I think we'll see these uses of AI cluster into a few groups:
- maximal code / reasoning quality, at high prices (Fable)
- so cheap I don't care about utilization (Deepseek, and friends)
And also more niche ones:
- ultra fast with high quality answers (typically sub-SOTA). Cerebras / dedicated silicon type approaches, expensive.
- ultra fast with mostly-adequate answers, and an openness to retries, moving up to better models
I think the open models will dominate (not with individuals, but low cost providers) all except the top 1-2 of those categories, and there will be a continuous erosion on the big player's moats. The top categories are also where all the money is, but I'm not sure it can justify those investments long-term. I also think they will have to squeeze more money out of them to justify the investments, which will also drive people down the list.
>For the near future it seems that the new models will consume whatever improved hardware capacity we have.
But they're not. Meta, SpaceX, Microsoft, Amazon, they're all leasing out capacity to others. If they were truly constrained, we wouldn't see that happening.
You may be correct but the combination of local models when they are fast enough and work, combined with paying close to zero for deepseek v4 flash from US providers is pretty good. When you need it, glm5.2 is cheap to use and very good for working with larger projects.
Are frontier models actually astronomically expensive to train? GLM 5.2 was trained on ~30T tokens, so ~10^25 FLOPs. Say you get B300's for $5/h (pretty high), and you get 50% MFU; that's ~$15M. Of course there's also a bunch of risk that the training itself goes badly, post-training, etc. But still, compared to the inference spend after, it's not that crazy.
Completely agree. Once I can reliably get open models doing what I am on Fable ultra I imagine I will switch for good. I am fortunate to have access to a decent bit of local RAM, 192GB of DDR5 at an OK speed. It is not enough and costs are well past absurd. In a few years time I envisage a setup that is sub $10k which can accomplish such tasks. The pace so far has been breakneck. That is all I personally need. That may change, but until true AGI I do think there will be a ceiling to how much I will pay for something frontier if it is only marginally better.
This is easier to say as Fable is good (even SOTA). But people have been were saying this continuously for the current model and for now the improvement are still coming.
A better question is would you settle for o3 now or pay 20$ or 200$/month for fable ? Because o3 quality is available OSS.
It is like the new IPhone, in some sort. At some point come a feature many would like to have, despite diminishing returns.
We will see how long labs can keep up and what the scaling curve look like, but I would be more worried into losing sota status to Chinese companies than letting them take the open non-sota approach.
IMO we are there almost. I had every iphone until 14 pro and still use this today. I know what the top of the s curve feels like. From a pure model standpoint weighed against every day use cases for every day people (the right way to measure when comparing to something as ubiquitous as a phone), the models already have diminishing returns.
I think there is also the case were companies will simply use different tiers for different tasks.
While the engineering team might need a cutting edge model (with the associated costs), the marketing department will be fine by something that can grammar correct or turn a few bullet points into prose. Likewise you already don't need Fable for Ticket -> RAG -> Reply with Faq knowledge or escalate workflows
That's already the case with other very expensive software like CAD packages were oftentimes you have different feature sets enabled for different employees.
Speculation: Anthropic and and OpenAI become more like hedge funds, keeping their models to themselves, using their them and their compute for market/event prediction, internal tech development, and AI self improvement. I already wonder if that is what Google is doing. Why release your best models to help the competition? Their published models are good enough for 99% of consumers, and they can leverage their market dominance in other areas to put that consumer model in front of eyes. Then the decision to improve your internal models would both depend on whether you think they'd have value internally.
Yeah, I'm pretty sure AI is going to go insular within the largest companies. This will only be hastened by the growing national security concerns/awareness.
They would have done that long ago if it was that easy. The fact that they aren't tells us something about the actual business utility of leading models.
> open models is what will kill Anthropic and OpenAI.
Maybe killing Anthropic is a good thing? Anthropic believe that they are the moral god and they get to determine what we can do and can't do, and they get to tell us what model to use and what not to use. I find it very counterproductive.
If consumers ever have enough VRAM for it locally, sure, otherwise nowhere near close.
Nvidia is looking like they are ditching consumer markets in favor of enterprise GPUs since nobodys heard a peep about the next iteration of RTX cards. The 60xx series is postponed till 2028.
Nvidias playing a dangerous gamble, in my eyes I see all the frontier labs eventually just only buying Nvidia chips for training and building custom ASICs for a fraction of the cost, longer lifespan and cheaper to host.
This will eat their 5 year gravy train for GPUs vs the 10 to 15 for ASICs.
I think this is possibly true, but the other piece of the equation is tooling. Right now, Anthropic has by far the best tooling around (for both SWEs and non-technical users), and a huge ecosystem of integrated ISVs. I'm not suggesting it will happen, but if Anthropic decided to provide options of using models beyond Claude, they'd still have a significant moat.
I dont know. The tooling feels trivial. I swapped from claude to codex and back with zero friction. If they were very expensive to me personally I would probably just use them to build the harness and then plug Kimi into it etc.
The real moat aren't the models, but the tooling around the models that allow them to perform specific tasks/goals. That's what really sets apart frontier vs open. Open only has the model itself, closed have the tooling to enhance the model.
> The real moat aren't the models, but the tooling around the models that allow them to perform specific tasks/goals. That's what really sets apart frontier vs open. Open only has the model itself, closed have the tooling to enhance the model.
As these frontier companies have been boasting, writing software is now a negligible cost because the LLM can do it.
IOW, no, their software can't be a moat, because, according to their own arguments, you can use their LLM to trivially clone their software.
> you can use their LLM to trivially clone their software.
Perhaps, but what you can't clone is familiarity, polish, integration, and network effects. These companies desperately need a moat, and so for example all the new additions to Claude's web interface are becoming the products that the vast majority of people will use and be locked into (if it goes according to their plan).
I'm happier with my Opencode Go running with the pi harness than Claude Code at work currently. I'm not sure if this will be that big of a moat, at least in the software dev market.
unfortunately, for me, that's an anti-moat; their dark, inconcievable, alignment, random "no goblins" and inability to be reliable where their models are inevitable non-determinant means you're going to constantly run into the "this worked yesterday" problems, no matter how smart the models actually are; being filtered constantly through economics, political and ego-maniacal filters means everything you do with it will break at any given interval.
I agree with your speculation to be honest. And yet I’ve tried several local open weights model now and none gives the same quality of answers as Claude gives me on a regular Sonnet model. Mind you: I am “running” 48GB of RAM so I can’t try every model. Where does this difference come from? Can we actually get close locally?
You’re running 48GB now but imagine a future where everyone has 512GB RAM or 1TB RAM in their computers (it might sound like a lot but also 20 years ago we had 512MB PCs).
It’s not hard to imagine what 5-10 years of pressure to increase RAM will do to specs, on top of the normal tech improvements.
That’s worth bearing in mind when thinking about local models.
Plus, local models keep getting better and better; 2 years ago what you could get out of those 48GB of RAM was embarrassing compared to what’s doable today.
Something I've noticed is that local models are giving better answers these days than they did a year or two ago, even if the size (in parameters and in the amount of RAM used) hasn't increased. I'm not familiar enough with the technical side of model training to explain how they're doing this, but I think in another couple of years, models that use up 48 GB will be able to squeeze out even more incredible performance.
Though on the level of something like Sonnet 5... well, maybe not.
>open models is what will kill Anthropic and OpenAI
i doubt it. it cost money to train a model. we can see that with the price increase for Kimi3. Chinese AI companies is leaving a lot of money on the table for third party providers. you think they going to let go of those money that they can make. sooner or later they will want to collect. after all, none of Chinese open weight model is release by a none profit. its all for-profit companies that is releasing open weight model.
I can't take anybody seriously when they keep declaring open models is beating frontier models. What they don't understand is that besides the huge capex to train and run inference, the real gold is in the human response to the prompt results, this is what all the Chinese companies are making their open models dirt cheap and distilling american frontier models via scraping.
The idea that we can out-parameterize frontier models is a common misconception, the true moat that Anthropic and OpenAI is why Chinese model providers are open sourcing and making it dirt cheap to keep pace through its "proxy chain operators"
I worry about that but here are two positive things:
1. it would put us (USA) at a competitive disadvantage, and cooler heads will prevail in this fight
2. there are good US open models. I have the latest gemma4:27b with better tool support functioning at a high level in the pi coding harness. Thinking Machines seems to be on a good path, we will see what they and other US companies can do.
> at a competitive disadvantage, and cooler heads will prevail
I don't think this a safe assumption at all, given the the last few years of government policy where the "cooler heads" very much failed to stop such things. Either by partisan choice, or because they were silenced/fired for telling the Emperor that his clothes weren't there.
In particular, consider the tariff-taxes [0] against on US importers, and the ensuing breakdown between America and its oldest allies and trading-partners. Also relevant would be the President's recent pet-project war [1], which has shut down a major part of international shipping.
_______
[0] Unilaterally imposed by fiat, with idiotic "failed economics 101" math for the rates... and the majority in Congress still actively tried to keep it going, until the Supreme Court ruled it illegal.
[1] Simultaneously a war and not a war, both already-won and ongoing, depending on which lies need to be told today to harvest applause and grab cash. Just this week the White House submitted documents to Congress claiming what's going on is a completely separate and new, second war with Iran...
There is a massive difference when you zoom in close or take an angled perspective. You can manufacture uniqueness. The issue is when it comes to every day use for every day people there is no differentiation.
You can run the same harness on fable, opus, sonnet, and see a huge difference between them. It is true the harness is important, and openai has begun encryption its instructions to swarmed sub-agents instead of just encrypting the chain of thought, but the model is still important at this stage.
Referent of "the models are meaningfully different" reads as <top closed, top open> rather than <top closed, cheaper closed> to me, so I'm not sure why we'd be comparing Fable vs Opus/Sonnet or Sol vs Terra rather than the same against Kimi K3.
I was just responding to "The harness is what takes these random and hallucinogenic models and make them into something deterministic and useful."
You can compare Fable vs Sol vs Kimi in the same harness if you want too and there are meaningful big differences. I chose all Anthropic ones to be safe from the they were finetuned on different harnesses complaint that would be made from that comparison.
I still strongly believe Google Gemini has the best position for one simple reason: model maintenance. Accurate information is a moving target.
Open models are indeed very capable, but they will eventually become more specialized to the application to keep an edge. It makes perfect sense that the future shape of AI conforms to the landscape it was born out of.
You're saying it's important to have up-to-date facts stored in parametric knowledge? It seems to me like that's grown less and less important as agentic capabilities have grown. Even if a frontier model doesn't know something, if it's out there, it can easily find it through tool use.
But google can give gemini access to summaries of all the avaiable information google has, including what people look for right now and where and how long.
Eventually they will kill the hyperscalers too because of privacy issues. It's better for a company to pay an uprfont cost and then run everything on premise that uploading their entire codebase to a third party service.
The vast majority of companies are still putting most things in the cloud and will continue to do so and this far outnumbers the must-be-on-premise companies.
Sure there will be self-hosters but hosting AI models will always be more of a challenge than running scalable database on your own hardware and specialized hyperscalers will be here.
Would that require a watershed event to clearly establish the importance/risk of privacy though? For example, right now it seems like most big software companies w/ strong security process are comfortable uploading entire codebases to Israeli cybersecurity firms for vulnerability scanning compliance purposes
The thing about not much difference between models and the harness making them deterministic and useful is wrong. Also models have different strengths and weaknesses and some are better at almost everything by a large margin compared to others.
As for your speculation, I think it's hinging on some companies releasing models for free or no big differences between models. In a world with hyperscalers and companies training models you can quickly recreate Anthropic or OpenAI by having an hyperscaler ally with a model training company, train a good/a better model, and not release it.
You do realize OpenAI survived for years without LLMs, right? There is still more AI research they can do as a lab even if they stop experimenting with LLMs.
I don’t even know the names of any open source search engines, but the open source models perform decently on various benchmarks and in personal experience.
Was it ever even a claim that open source search engines were trying to outperform google, let alone kill it?
Open models are 4k TV (or maybe 1080p tv now and 4k TV soon) and SOTA frontier models are 8k TV. Can I or the average user tell the difference? Not really. Would they pay for that difference? Not a chance. Our entire economy is teetering on some future hope that this fragile and immaterial difference will pay off, when the reality is that LLMs are a race to the bottom and eventual razor thin margins. Maybe a tiny vocal subset of programmers can use it for work and make paying for it worth it to them, but that can't prop up an entire economy, especially when said programmers are phased out, jobless, and replaced by AI with each better iteration...
Except when we upgraded to 1080p from older TVs, they got bigger. Now with 4k they are getting bigger yet. More powerful models means new use cases that didn't make sense on the weaker models.
Even in the world where all models are basically equivalent (a thesis I don’t buy, but will grant you for arguments sake) - I believe there is much more to the AI business than just training and running models.
It’s a very new set of technologies, and understanding what is useful to customers and what isn’t is the whole game. Call it, product taste. There were a million cell phones before the iPhone took over the world. Why iPhone? Product taste. There are a million startups, and only a select few become unicorns. Why? Product taste.
>There were a million cell phones before the iPhone took over the world.
You have tripped yourself up there.
iPhone took over as it introduced something innovative over standard phones, but then Open Source (Android) matched the multi-touch and software differences and Apple's branding, lock-in and design etc have managed to keep it as a big player in wealthier countries. IPhone also came on the back of the massive iPod success.
ChatGPT launched the same innovation vs Google Search, but just like Android Opensource AI is moving fast now.
Android has 72.7% market share at present, Open Source AI will do the same unless the frontier labs can continue to do something new.
The frontier labs are saddled with enormous investor and other debts. How long they can keep innovating by spending so much on R&D and paying there staff very high wages remains to be seen.
Once investors cash out via an IPO, the companies are back down to earth and playing in the real world again.
Android has market share, but Apple makes all of the money! I find it really funny when people attribute Apple’s success to “oh, the only reason they succeed is design and marketing.” Yeah, I mean factually speaking design and marketing actually do matter a lot!
Us developer types like to pretend like specs are the only thing that matters? If you could have a 10x more powerful model you could only access running locally through your terminal, versus a weaker model through a clean web interface, normies will pick the web ui every single time. Product experience is simply everything, as much as we like to pretend like nitty technical decisions are the most important thing.
> Android has market share, but Apple makes all of the money!
So? The benefit of open source is that you don’t have to worry about making a ton of money. You just need to be viable.
Apple: premium product a minority is willing to pay for
Android: standard product the majority use
I’m sure there will continue to be iPhone equivalents in the AI world, premium bespoke models. But the vast majority of people will be happy with a cheaper offering.
Well I think a critical difference is that, unlike Apple, OpenAI and Anthropic have taken on so much VC funding that a 20-something % market share is not going to be enough for them. So open models could kill them, not because of the techonology but because of the way they're financed.
> If you could have a 10x more powerful model you could only access running locally through your terminal, versus a weaker model through a clean web interface, normies will pick the web ui every single time.
More like if you could have a 1.25x more powerful model that you could only access through some weird surveillance megacorps aggressive monetization scheme, or choose from 100 others running open models and accessible through 100 different interfaces pandering to every taste.
Normies will choose the megacorp every time, because that was the one in the tv commercial, and within six months will have left for one of the others in a rage.
The only corporate hope is that the government steps in to ban their competition.
While that may be technically true for a strict definition of “smartphone,” there’s no denying the iPhone redefined the concept in a way that its competitors were forced to copy to have any hope of keeping up. Nobody hears the word “smartphone” and thinks of a Blueberry or Treo anymore.
It preceded its release with the iPod craze, making it a lifestyle product rather than something marketed for its capabilities (which were very humble compared to other contemporary devices and only started to catch up in second gen).
That's a subjective question, so I'll give a subjective answer. The browser, for better or worse, was a lot less dumbed down for mobile than competitors, the stylus-less touch interface reduced UI friction and the odds that you'd lose a critical (if inexpensive) component, and the slew of contemporary iPod users could easily migrate their libraries over.
Exactly 4 months ago, the marketshare on openrouter was 60%-40% in favor of closed models. Now it's 63%-37% in favor of open models. On March 19th, the open models processed 888B tokens in aggregate, yesterday, they processed 4.19T tokens in aggregate. That's almost 5x in 4 months! I can't think of the right intensifier to describe this level of growth.
If you are looking for more details (as inferred by openrouter data), I built a dashboard that updates daily: https://dirac.run/labs-market-share
Definitely interesting seeing it grow there. I'm now curious to see if it's overall market usage growing for LLMs (with more OpenAI and Anthopic usage in a similar growth rate), or if it's more heavily weighted towards open models specifically.
With the price to performance of the latest open models, I think most cases for integration into applications would get a best bang for the buck from an open model.
This right here is going to be considered one of the first major signs of the downfall of closed models years from now.
And look, if you disagree with me PLEASE tell me why. What moat do these companies have? I genuinely want to know because looking at the spend for companies like OAI and Anthropic with no actual moat I can identify is actually driving me insane.
I think the frontier AI companies think you're missing key details:
- they can still discover entire new untapped markets for AI (that, potentially, only their models can unlock)
- they can find (novel, unique to them) ways to drive down the cost of running their models
- they can provide other ancillary value (e.g. write better harnesses) because of their expertise, and then charge for that value
I'm probably missing a few bullet points also. However, none of those are moats (or at least not yet) ... I'd consider them more like bets. The frontier AI companies are betting "the house" on them, and if they pay off they could, hypothetically, make them financially competitive.
I've been keeping my cash/stock ratio a bit higher than usual recently because I'm waiting for the bubble to burst. My "stock" side is mostly a company I work for that is riding the wave.
If we use the same logic as the AI boosters and just assume we're in exponential growth, that means in a year open models will be handling 523T tokens.
I'd be curious to learn more about how/why executives come to sign their names to this kind of writing. Maybe it feels like a natural iteration of the pre-existing experience of signing one's name to an assistant's manuscript, or a press release from staff in Comms? I don't know execs who do this kind of thing, so would be grateful for any observations others might be able to share (anonymously or otherwise).
Aside: I think it would be very good for the culture if HN automatically ran Pangram on any text-centric articles. Even perhaps just those which hit the front page? I have my own browser plugin, of course, but judging by comments here, at least some seem not to notice?
Agreed re: automatically running Pangram, maybe once the articles hit some vote threshold (or just, frontpage) in case cost is a concern.
I think aggregators running AI checks automatically seems quite good, rather than individual readers all doing it on their own sporadically. We recently implemented auto Pangram checks on Manifund, and I think LessWrong has done this for a while now.
This presentation is painful to read. It's an LLM's idea of a CTO presentation. I'm overwhelmed by charts, only slightly connected to the text around them. But no matter, it looks like a CTO slide deck. HIGH IMPACT.
Much better would be if the CTO of Mozilla had actually articulated their own analysis.
As someone that's generally for the proliferation of open models, I want to take this seriously, but it's really difficult when it was clearly written by AI.
I guess they fired whoever used to write copy for these things.
Edit: to be clear, I'm not trying to just dunk on them, I think it's actively hurting their own point to do this, and counter-productive when people can easily clock it - it makes some percent of the audience immediately tune out.
> Mozilla exists because one company tried to own the front door to the web, and an open community rose up to make sure it never could.
I'd say that the front door to the web is pretty much owned by Google and Apple at this point given Firefox current marketshare. And maybe that's enough, maybe a future where a low percentage of open models keep the rest of the system honest but that doesn't seem the argument of this article
Mozilla exists because one company owned the front door to the web, and another company abused their market position to push their free (as in beer) browser. Mozilla is the phoenix rising from the ashes of that first company.
Then another company came along, abused their market position to push their free browser, demolished Firefox's market share but keep handing them cash to avoid the appearance of a monopoly
At their current trajectory, absolutely. But if they focused their development efforts, cut down on careless spending especially to third parties, and looked for alternative sources of funding (it is presently not possible for users to donate to Firefox development, for example), I feel like they could survive without Google for at least a little while. Frankly, I think these things should be started on now, while they still have Google's backing.
It sure is nice to see that Mozilla is still doing all that they can to keep on top of current trends, except developing a decent privacy-focused web browser for developers and power users.
I’ve been using Firefox on mobile (and desktop) for years. I still don’t understand all the hate thrown at them. Whatever the downsides/ shortcomings people see, they are irrelevant to me, because I hate Google more & refuse to give them my data. They are good browsers & work just a well as Chrome 99.9% of the time for me
This used to be my attitude, but I do feel like Firefox has made good ground in 2026.
I've also become sympathetic to their AI strategy. They don't seem delusional in their approach. They're not building models or selling slop, they're building OSS compatibility layers.
I don't want to see a world of vertically locked AI, so if Mozilla truly can dust off the old playbook for OSS AI, we'll all be better off for it.
Yes, Firefox itself is a “general purpose” browser, and that’s probably for the best in terms of wide market appeal. Other developers have taken the engine and made power-user-focused browsers with it. Lately I’ve fallen in love with Zen (though after 2 months of use its pinned tabs features still confuse me a bit).
I am sad that there doesn’t seem to be any community whatsoever around _truly_ open models that are released with source data and training methodology, such that they could actually be reproduced given the resources. We’ve allowed the term “open” to be diluted to a shocking extent.
The design and layout made it harder to read than it needed to be.
Regardless, the inference costs dropping almost 50× is really amazing to see. And now Kimi K3 release has shown how open models are getting closer to the frontier level already. Open source AI is moving a lot faster than Anthropic and OpenAI would have expected lol.
When europes mismanagement of AI suddenly turns out to be the best thing that has ever happened. At best only few billions lost to training models that become obsolete next month.
I use opencode as my harness which is pretty ok. I daily drive it with no issues. It's not going to do any orchestration for you but it works.
I'm slowly moving to Pi but it's a very DIY system. Think neovim for agents. I think it has the potential to be where I end up for a while but it's a long road.
I’m not ready to celebrate the victory of open models just yet considering all the good ones are built by private, VC funded companies. How long will they continue to be charitable? And what’s their actual business model once the money stops and investors start demanding returns?
Gemini had a pop-up on my phone today, asking me if I wanted to "bring Gemini up to speed" by importing conversations from my other AI apps. This tells me that Google is threatened, or data hungry as always, or both. Open source AI, Anthropic, and OpenAI, knocking on doors.
It's their own fonts: Mozilla Headline and Mozilla Text.
No idea why they'd be using the display font for the abstract though, that kind of defeats the whole purpose. It's supposed to be quirky and bold, but used far more sparsingly.
It appears open models were used to create this slop.
That opening is so hard to understand what they are trying to say, from the font and how it's written. It took me several times rereading to even grasp.
Plus the article is filled with cryptic things like:
Open ships easy.
Open deploys hard.
What?! Is it a meta answer to "the state of open source AI" question?
I think it’s supposed to mean “open source is easily shipped, but open source is hard to deploy”? Or perhaps “deploys hard” is a figure of speach, as in “we are deploying this open source and we are deploying it /hard/“? I don’t know, it’s not good.
This is truly some proper slop. The "PRODUCTION RATE BY COMPANY SIZE" graph has bars that start offset from the text underneath them, which LOOKS like a mistake that happened due to word wrap, but if you visibly compare the 54% to the 55% bars they seem to have compensated for this?! I can't tell if his was on purpose or accident and it's impossible to take the data seriously!
This is on mobile in portrait. In landscape the text doesn't wrap or offset anything.
The issue is that all of the text is a quote, and that renders enormous. That’s probably fine for a tiny quote amongst more text, but here it is jarring.
querySelector/querySelectorAll() are great for plucking out deeply nested elements but if all you need is to find all elements of a certain class and the API gives you a tool to do exactly that, why not do that instead of reaching for the general-purpose Swiss army knife? Sure, the execution speed difference may be only measurable in microseconds, but it takes about the same amount of time to type so why not use the specific tool?
This new trend of content appearing while scrolling down is so terrible accessibility-wise, I do not understand how Mozilla of all institutions would do it.
Not every trend needs to be followed. Have some backbone. You receive donations to have that.
___
Apart from the website being - frankly - bullshit, the content is also - frankly - bullshit.
It's just on the frontpage because the title says "open source AI".
> This new trend of content appearing while scrolling down is so terrible accessibility-wise, I do not understand how Mozilla of all institutions would do it.
Could you explain what is wrong with the accessibility of this page? All the content is included in the html payload, so it is accessible to screen readers and text-based browsers; and as for the "reveal" effect, it seems to respect user's choice of "prefers reduced motion" and is disabled when that is user's preference.
> it seems to respect user's choice of "prefers reduced motion".
Cool, that I didn't check, because it is impossible to enable that setting, as it breaks _huge_ amounts of websites.
I'm not aware of a way to enable it selectively, but one could also just display the content at all times. It's a static page. It's static content. None of this makes any sense.
___
The idea behind that style of gradual reveal is probably some kind of storytelling, but the only story it tells is that mozilla is wasting donations on people with incorrect opinions that could be used on.. idk not building torment nexii?
Seems quite odd to use OpenRouter as “proof” that open weights models won. If you’re using OpenRouter, you’re already looking to bypass frontier models. To suggest there’s no longer a tradeoff simply isn’t true. But this isn’t the first time I thought Mozilla was a less than trustworthy source of information.
I think "If you’re using OpenRouter, you’re already looking to bypass frontier models." is false. Our company uses both Claude subscriptions and OpenRouter ... and a lot of what we use OpenRouter for is more Claude.
We do a little exploration with other models through it, but it's not at all accurate to say we use it because we are "already looking to bypass frontier models".
... or at least, no more than any other company that doesn't want to overpay for their tooling, but is basically happy (ATM) with the current state of Claude.
The UI is really hard on the eyes. Personally, I think the font size is way too big, and the animation timing feels off. If this is a benchmark page and not a product page, I feel like the information should be scannable at a glance. The UX is bad.
I'm unsure what it is about AI developers seemingly not having eyeballs. The Hermes Agent website is absolutely eye-searing and the application itself resembles some sort of weird "RETVRN" greek-styled travel agent website.
There isn't any open-source AI. There is Open AI (not to be confused with the closed company called OpenAI, which was unable to trademark its name). There's no open source AI both because the open source community doesn't have the resources to train a useful AI and because AI doesn't have source code.
I'm not sure what kind of point you're trying to make. There are projects to train competent modern LLMs in which the entire pipeline (data, training process, final weights) is all completely transparent, shared, and reproducible by anyone with the compute to try it out.
Or is your definition of "open source" mean that a small indie dev should be able to reproduce the entire pipeline? Because that would disqualify more than just LLMs, but also hardware platforms like Arduino where you need to pay for manufacturing to get the underlying stuff built... is Arduino "open source?"
I don't know of any useful model that would match the usual definition of "open source". That is, where everything needed to build the thing is provided, such as the training data and code. Not that it would be tremendously useful anyways considering how expensive training a model is.
I much prefer the "open weights" term. It is not open source in the sense that you only get the finished product, not the actual source, but it is still open in the sense that it is not only accessible as a service.
For an analogy, take Quake for instance. When it was launched, its game server was available as an executable, so you could run it your machine, but that didn't make it open source. Only much later it was released as true open source software.
Olmo 3? K2 V2? There are definitely LLMs with very compelling capabilities where the dataset, training process, and final weights are all open. There are also initiatives in the EU and various national government levels (e.g. Switzerland) to develop LLMs in an open and transparent way, not just "open weights."
mozilla only began babbling about open source ai when they got their teeth extracted for free via the entire community giving them the curbiest of curbstomps.
i still use firefox but hot damn did they utterly fail to read the room initially. the only other company i can think of in recent memory -- besides sony cutting out discs -- is logitech when their ceo began gibbering about a subscription mouse or microsoft and its copilot button(s).
It's frustrating because I truly believe they are well-meaning, honest actors. Every single "controversy" surrounding Mozilla can be directly attributed to miscommunication rather than malice. They do often struggle to read the room.
That said, the takeaways from this report are exciting, and I do feel that Mozilla now has the right lens in their assessment of OSS AI and their own approach to ensuring interoperability and setting open, modern standards.
I think Mozilla is chasing a past formula, but the projection isn't linear enough to remain consistent, and the critical parts of the outcome, utter centralization of the market dominance of the three C's, are left out of the equation.
We might get the consolation prize, a few nerds having competitive alternatives to applaud, but we will be left with the hidden costs: stagnation by bloated market leaders, consumers and businesses pouring trillions of dollars into the commercial offerings while open development wonders where money comes from, and the leakage of these imbalances into political and social spheres.
If we follow a Mozilla template and get to the peak of Mozilla's success at the web, look at what that really is. Facebook, Amazon, Google etc are orthogonal to that equation.
I think the fact that Mozilla survival model ultimately depends on Google's money means Google is keeping a corpse propped up just to have a defense argument that a browser competitor still exists, so they don't get hit with monopoly regulations.
Mozilla presents themselves as an "open community". But they are a commercial corporation bankrolled by Google, and with an oppressive organizational culture. I would interpret this post as being not just about an opportunity for their self-aggrandizement, but also Mozilla trying to whitewash its mass surveillance of users, likely used also to train some model or another, or perhaps just to feed their patronn (Alphabet)'s ad machine.
They tell us about how the farmers and native people and whateve are all happy with their chatbots and models. The major effects are a massive and ever-increasing energy use - in a time where we must cut back and economize to avoid further global warming; a massive diversion of investment capital - especially in the US; fantastic stock valuations for a few tech giants (gee, I wonder whether any of them is related to Mozilla somehow); and other effects one could survey, all more significant by far than the examples they bring.
Haven't been following the articles and snippets we get from these labs about training their models for a while. But I'm guessing the latest chinese models are way less based on distilling? If not, then your speed of progress is still limited by the two labs (which we are collectively, in various forms subsidizing).
There's nothing practical about open-source models yet that makes them even remotely comparable to closed frontier models.
All the hype around GLM, Qwen, now Kimi.... Are people really this naive that they believe these reports or, more worringly, are people NOT using these models and seeing the HUGE gap that still exists?
Take a task, any medium-sized task, decently scoped that you'd trust to give to Sonnet to finish without a hitch. Now give it to ANY open-source frontier model and watch them struggle and go in circles while failing tool calls and randomly assuming things.
Open-source is and has been amazing but its so hard to deploy reliably and at scale and there's still big problems in the underlying models with instruction following and tool calling that makes it basically unusable for production workloads at a decent price point...
> Take a task, any medium-sized task, decently scoped that you'd trust to give to Sonnet to finish without a hitch. Now give it to ANY open-source frontier model and watch them struggle and go in circles while failing tool calls and randomly assuming things.
Claude used to be much worse than it is now, just as bad the open weights models are. And the open weights were worse. The labs will also try to keep the lead, but at some point people start seeing real value from open models. Maybe you say they're not ready yet for medium tasks, but everyone sees the writing on the wall.
I hope you're right and I want you to be right, but, even seeing the current hype around local models, etc... and open-source models, I think the industry is currently under a big confusion where they see the benchmarks of things like Kimi, GLM, Qwen, they play with it via opencode, and they think like: "Wow this is pretty good, I want to deploy this". But they don't understand how the KV cache grows over time and can take almost as much memory as needed for a 30B param model, they dont understand that a quantized model WILL NOT be the same as a full precision one, and they surely don't see the engineering work needed to serve inference to even tens of customers at a decent quality and latency level.
The biggest moat of these giant labs and models is increasingly shifting towards deployment capabilities and (debatably) having better (proprietary) harnesses.
The models themselves can be impressive on benchmarks, but unless they can be served reliably to customers either at scale, hosted somewhere, or even on edge with predictable latency and memory usage, then frontier will always be leading.
It sounds like you're focusing on the problems of running local models, or running models yourself, but I don't think that many people seriously expect near term improvement on that, it's definitely more just hopeful thinking there. That's not what I meant to address, and I also am in more of a "wait and see" mode.
But at this point we do expect that open weights _hosted_ options become feasible for the tasks they're using the frontier models for. And because of the lack of "legal monopoly" (intellectual property of whatever kind), they're way cheaper, not mention more flexible.
The launch of the tinker platform from Thinking Machines is an example of the "more flexibility" part that people want (and they chose to make their model open weights, maybe because this is the angle they want to push).
At this point I think it's realistic enough that the ball is in OpenAI / Anthropic's court to figure out how to respond to this threat to their business model.
That said, I think it's concerning that there are apparently only a couple of providers of hosted open weights inference, due to the complexities of doing so (per Dax from OpenCode's tweets).
i'm... not sure? This assumes ~stagnation in task-possibility. We've had ~exponential progress for like 3+ years now; I'd have never dreamed the tooling I hammer daily would exist in my lifetime just.. 3? years ago. And it's improving daily.
Maybe Open will win, maybe Closed will keep pushing the envelope. The world here is raw enough i don't think anyone can make any significant claim other than 'holy shit this is useful and moving Fast'.
Speculation: open models is what will kill Anthropic and OpenAI. Hyperscalers can run the models without a licensing fee. Apple can make them smaller and put them on the device.
The frontier models are an edge and a liability. They're astronomically expensive to train. Without them, their models will fade into obscurity. Their marketing depends on people believing the models are meaningfully different, as people have sweatily argued on this forum. Personally, I'm not convinced there's much of a difference between these models at this point. The harness is what takes these random and hallucinogenic models and make them into something deterministic and useful.
The outcome is plausible. Open weights models though look like a tactical more than a principled play by Chinese companies to overcome the disadvantage and difficulties to access western markets. Two issues:
1. If market conditions change they might decide to close down like Meta did.
2. If as you said models keep getting more expensive to train, is an open weights strategy financially sustainable?
edit: typo
I think (and have heard) the Chinese govt is also interested in trying to embarrass the US as well by showing off their capabilities (so there's a political angle, too)
All countries see AI as a geopolitical concern. An open source strategy is smart. It’s effective and also builds good will / soft power. Many see Chinese AI companies as the good guys and root for them. I’m pro open source and also celebrate but still aware this is most likely just the means to an end.
I used to work at Mozilla. I think we are missing a player in the market with a more principled open source approach.
And who might that be, I wonder? The company that seems more worried about executive bonuses than to adhere to its mission, for at least a decade now? Perhaps a certain one whose name starts with M and ends with ozilla?
Purely random thought, of course.
I don’t know. thinking machines perhaps?
A lot of things have to come together: talent, capital that is not looking to maximize returns, ambition to compete in the marketplace. At peak Mozilla it was clear to everyone that we needed an undeniably successful product to be able to influence the market. Advocacy alone is not very effective.
Definitely. Sadly "capital that is not looking to maximize returns" is a very tall order in the last 20+ years.
That's probably pretty likely, but if we're honest, are LLMs built and funded by a hostile Chinese authoritarian regime any more dangerous or harmful than LLMs built and funded by a hostile American authoritarian regime?
China absolutely does not have my best interests at heart, but America's technofascism is probably more immediately dangerous and harmful. Americans genuinely have more to fear from America than China at this point.
One big difference is that the US doesn't have much public record of using state espionage resources to steal industrial secrets to give domestic industry a leg up in the market, whereas CCP very much does.
And while I use and love the open weight models, it seems likely to be pretty hard to prove conclusively that they have no reinforcement-learning-trained proclivity towards curling specific URLs if the topic happens to be some very specific thing that the government is interested in, when used to drive agents.
So it depends who you are. If you're a company, you might have something to fear. If you're a private citizen, maybe not so much.
I don’t think US companies could ever fully trust models sponsored by foreign adversaries. How would you ever verify they are safe?
Provenance of training set is going to be critical. It’s largely ignored now, but the first time a malicious exfiltration occurs it will go off as if a bomb was detonated.
Only way would be an actual open source model: code + training data
You're moving the goal post here. The parent never mentioned anything about the Chinese models themselves being dangerous or harmful; that's a totally different topic.
The point they were making was that the most successful open models- those coming out of China- are made my companies that are using those open models to get exposure in Western markets. The goal is to undercut the Western dominant players, not out of any particular "open source" philosophy, so it wouldn't be wise to expect them to continue providing open models long term.
Define technofascism lol.
Fascism was laid out by Mussolini in the 1920s - it amounts to the idolatry of the state.
"All within the state, nothing outside the state, nothing against the state." B Mussolini
Also defined as Corporatism: the union of state and corporate power.
Which country do you think is closer to Mussolini's model ?
Fascism was strongly influenced by Futurism and its optimism for a machine-powered future where technology will bring forth utopia and order.
It’s no wonder than Thiel & co. are rediscovering Futurism, and blind faith in the machine is basically what Silicon Valley is all about.
I think the term 'techno' in 'technofascism' is doing the work here, because--just as you claim that historical fascism is the idolatry of the state--technofascism is the idolatry of technology and intelligence. Modern accelerationism, as espoused today by people like Marc Andreessen in his Techno-Optimist Manifesto, is really just another rehash of Italian Futurism, which was closely intertwined with Italian Fascism and one of its intellectual foundations. It was, essentially, a progressive ideology. The idolatry of the state does not vanish, but rather gets transformed and re-imagined as a technology, e.g. network states, platform governance, companies that function as sovereign entities. The idolatry shifts from the nation-state to the infrastructure that replaces it. Also, you don't need to look far to answer your own question about corporatism, as you defined it yourself. Now, who's currently moving between Silicon Valley boardrooms and government offices?
Fascism is not merely idolatry of the state. That is simply its ideological mask, its method for currying the favor of its base. To say that this is what fascism is at a fundamental level is to mistake its superstructure for its base. But how is Fascism structured? How does society reproduce itself? Is there a fascist mode of production? No. The state mobilizes its population so as to target scapegoat(s) while simultaneously expanding in order to secure the spoils of war. Without scapegoats, there is no Fascism. And a state requires social cohesion to mobilize its base. Thus without cohesion there is no Fascism. At minimum, social cohesion requires the ability to sustain and motivate a critical mass of the population. Thus Fascism requires the continued maintenance of the capitalist mode of production in order to persist, but it simultaneously undermines capitalism due to A. the continued need for war machines and B. the need to punish a scapegoat, which both siphon value from the rest of the economy, thus dooming the entire enterprise. This is why fascism burns out and leads to doom for all parties involved wherever it has been tried. Any definition of fascism that ignores this is historically inaccurate and practically useless.
Open models are probably also comparatively astronomically expensive to train - just less so than the frontier models because they’re somewhat smaller, +/- the creators are more incentivised to focus on getting more from less compute because they’re have to, +/- they rely on distillation of the frontier models and this is more efficient.
But efficiencies aside; creation of open models still requires a lot of money and compute from a large organisation which is willing to accept zero return for that spend. This largesse is unlikely to continue forever; so the question is which will crack first, the frontier models’ business model or the fast followers’ generosity?
Yes, the problem with comparing open models to open source is that open source requires humans to volunteer their time. Open models requires humans to volunteer their money.
These two types of contributions have very different behavioral profiles, and it doesn't obviously follow that the historical success of getting people to collaborate socially on building software for fun and for the benefit of the community will translate in any meaningful way to the necessity of being able to raise enormous amounts of money to pay for enormous amounts of electricity.
The biggest hurdle is whether humans volunteer their expertise. Not time or money. We need top talent to make the open models. Sponsorship is plentiful. Open source volunteers are less critical with LLM doing the grunt work. Its about talent contributing to the open
This is not my experience working in the field the last 8 years. There is not a dearth of talented researchers, engineers, etc. who are willing to contribute to open models. Just look at the ecosystem generally, from academia to industry. So much research still happens in the open, and the open source community in ML is still massive and energetic.
Training a new frontier model just costs a lot of money and there are very few companies for whom it makes sense to do so, and even fewer orgs who are just going to gift those kinds of resources to an open source initiative.
> open source requires humans to volunteer their time
Your idealistic of open source may require that, but in practice a huge part of open source is commercial and a large chunk of that is low on collaboration (across vendor boundaries).
Technically open source requires some amount of monetary volunteering, it's just that the electricity to run a code editor and compile (most) open source code bases is within hobby budget for most people.
I don't think it needs to be framed purely as generosity. You just need a sufficiently self-interested actor that sees open ecosystems as a necessary part of reducing their own risk profile, relative to the alternative of complete reliance of a third-party business that can take an exorbitant cut and/or Sherlock them at any time.
Valve and SteamOS are a good example of what this idea looks like in practice. (Though they may also illustrate a third thing you need: a privately-run company, that has enough profit, and enough commitment from leadership to the company's vision, that they can make long-term bets without having to eventually bow to investors seeking short-term gains.)
> You just need a sufficiently self-interested actor that sees open ecosystems as a necessary part of reducing their own risk profile, relative to the alternative of complete reliance of a third-party business that can take an exorbitant cut and/or Sherlock them at any time.
This would be an argument for an organisation developing its own model; but not per se for releasing the weights openly.
The possible explanations (I'm aware of, which overlap somewhat) for spending large amounts of money on models then releasing them for free (i.e. the current Chinese approach) are soft power, marketing for a future paid model business (i.e. competing with the US models for customers and mindshare during the time you can't compete directly at the bleeding edge), and/or a geopolitical move to diminish the value of the US's frontier model companies.
I acutally see the reduced burden that comes with actually sharing resources.
To extend the previous analogy: Valve didn't make desktop Linux viable on their own. A lot of it is owed to another self-interested actor -- Google -- through the reduced need for dedicated desktop apps (largely pushed by Chrome), and various enhancements to wireless and power management that were necessary to make it a viable mobile platform (directly benefiting Android/ChromeOS, but then spreading out to laptops and mobile devices in general, including handhelds like the Steam Deck).
You can see it on a smaller scale in ecosystems like Android -- where handset makers regularly contribute features from their UI skins upstream, so that they no longer need to spend engineering resources maintaining distinct versions of theming engines/notification badges/multiwindow/various other stuff.
On a related note this is an argument for open source models, not just open weights. I think a lot of the diminishing returns relate to the opaque nature of most models.
I’m not exactly sure on the “how” but it only makes logical sense for (non-AI) companies to band together to fund the training of a shared model. Apple is a great example, AI is not their core business but they still require it.
The only thing that took us down a different path is the vast sums of VC funding pumped into the AI companies.
It probably doesn't.
There's a reason we let companies specialize in some kind of service and buy it from them.
LLMs aren't looking like they'll be highly differentiated like software, so their market will probably be competitive. What negates the main reason Open Source software exists.
LLM training doesn't carry the same NIH risks that normal internal software bloat does. They are relatively simple to setup training for and analysis of accuracy/recall can be automated.
This leaves the price differential between a private third party and an internal initiative as barely more than the cost to train the model[1] - perhaps that's where we'll end up, a centrally trained model will represent an economy of scale that can leverage that difference into a margin it can profit off of but your business being purely profit driven by that training expenditure seems like a ridiculously thin margin.
So where does that leave the AI companies? If their LLMs are off the shelf-once built products they have a strong advantage for casual low usage but enterprise customers will have a huge cost incentive to roll their own - if the LLMs require continuous retraining and the frontier keeps moving then enterprise customers will find a packaged service more attractive and likely continue to subscribe for more accuracy but casual low usage will likely shift towards "good enough" models. It seems inevitable that they'll lose half the market and it seems difficult to discern their long term profitability[2].
1. Costs can, I think, reasonably be reduced to hardware depreciation and energy - if trends continue with cloud resource availability (it's possible this won't be the case if large compute providers start pulling resources offline to build a moat but I think they'd likely prefer the reliable compute income over model income which has several other competitive weaknesses). Hardware depreciation would normally be pretty negligible and equal across different training entities, right now we have a chip shortage but given the demand that can't last too long so I'd consider hardware to be fungible - and energy is entirely fungible - they're both hard to moat.
2. Outside of AGI, who knows if AGI will be or what even counts for it at this point - but I think if AGI isn't a doomsday scenario we fall back to one of the two above scenarios - either the frontier is ever moving and they can retain enterprise customers or the frontier seizes up and everyone can just use an off the shelf offering. In either scenario they don't have a lot of moat to deal with for their products unless they can restrict compute which is why Alphabet, AWS and MSFT are the only players I could see realistically coming out of this as an AI vendor winner and I'm not even certain if it'd be a good idea for them if it'd hamstring their cloud profitability.
If not for VC-funded LLMs there wouldn't be any LLMs.
Most of the innovations needed for LLMs came from people at Google.
A fair amount of ML/AI innovations came out of the market in general. Neural networks are a useful tool to solve a variety of problems... LLMs specifically were a more recent interesting market to develop but I've yet to see anything that could give a market player a real competitive advantage. It feels like we just invented a new hammer and now that we know how to build it it isn't that hard to build one yourself. The all purpose hammers are, of course, unreasonable to build - but those don't seem to be that useful. I don't really need Claude to be able to generate sonnets when I'm programming so I think specialization is the place we'll see genuine markets form.
> came from people at Google
Who had to leave to build anything.
So VC-funded.
The company was VC-funded as a search engine but by the time they made significant investments in AI (DeepMind etc) they'd been a publicly held company earning multiple billions a year from advertising for a decade.
Google is a VC. Their side projects are VC projects.
Google bought DeepMind and their other major AI acquisitions. Public companies make corporate venture investments for very different reasons than LP-backed VCs. They do early-stage investments to search for emerging players they can buy as soon as possible or to gain market intelligence on trends. They do later stage investments to help grow future vendors or customers and sometimes to foster ecosystems that form their competitive moat.
But if they think it's important to their core business, corporations don't want to invest, they want to buy. Source: I used to be involved in corporate venture investing at a top 10 valley tech leader.
If your definition of VC is "Literally anything that requires a long term investment of money" then sure, but I think most people mean something different than that.
[citation needed]
Historically speaking a lot of inventions have come about without things like VC investment. Either way, there’s probably little point in debating it, just because VC funded companies control the market now doesn’t mean they should indefinitely.
My (unverified) AI research claimed generally Chinese models are cheaper to train because Chinese data scientists are cheaper to hire and they're also under more pressure to optimize training cost due to limited hardware availability
Seemed believable but not sure where that's true
Chinese AI companies are generally smaller tho and the models they're releasing are also smaller (I think estimates put OpenAI and Anthropic SOTA into trillions of parameters)
I can't imagine even with the crazy salaries at frontier labs, staff costs make much of a difference.
In the US and Europe its extremely expensive indeed. In China it is much lower and expected to be much lower in the future when China can bake their own high end chips. You know China is focussed on optimizing production and manufacturing. Unlike the west that primarily focus on regulations that make things super expensive.
How does it work if people flock to open models but they're too expensive to train? What is the financial incentive to do so?
I seem to understand open models are mostly coming from China, and the benefit of training and releasing them for 'free' is a powerful geopolitical weapon against the Western/US economy that at this point depends on OpenAI & co. to succeed.
Will the West make open models illegal?
> Will the West make open models illegal?
We better not.
> What is the financial incentive to do so?
If we'd been sharing all along (as we should have been), we probably would have gotten even further along in the development of the tech.
Think of everything we could do if every researcher on the planet had first class access to the frontier. No academic fallback models. No crude API access. No limits, but direct access to the weights and the ability to lobotomize, splice, and dice.
We could pour intelligence from one container to the next without paying a tax or wearing a blindfold. All without spilling a drop.
*Open* *Must* *Win*
> No X. No Y. No Z, but Q
If by West you mean the USA, maybe.
Other countries in the westen hemisphere, probably not.
"You wouldn't download an LLM"
You wouldn't crash the stock market by preferring Chinese models.
I absolutely would. Let the bubble burst.
> releasing them for 'free' is a powerful geopolitical weapon...
I agree that, currently, the Chinese govt is not only allowing but tacitly encouraging open weight model releases. However, I don't see it as an attack. I think it's more of a strategic delaying move to slow the revenue to frontier models while China works to catch up. This strategy will likely change over time.
> Will the West make open models illegal?
In the U.S. this seems highly unlikely due to the current administration's generally laissez-faire approach to tech as well as the U.S. constitution severely limiting the government's latitude to constrain economic activity.
As we saw with the temporary Mythos restriction, there are legal mechanisms to limit tech on certain grounds, but over time such limits are subject to close judicial and constitutional review. The Mythos embargo was also likely driven in part by the administration's anger at Anthropic for choosing to block the DoD from using their products for mass domestic surveillance and warfighting. I doubt we'll see any meaningful restrictions on OAI or other large companies. It'll be nearly 3 years before a different admin is in office and could enact serious limits and by then it will be too late for fundamental bans.
There are vested interests in most governments, such as intelligence agencies, law enforcement and the military, who would prefer to restrict some AI from broad use. As we saw with strong encryption, they'll only be able to delay and constrain, not stop, such a broadly useful dual-use tech. The geopolitical, economic, competitive and civil liberty interests are similar between strong encryption and AI, setting up a similar game theory dynamic. While it can be argued AI poses some potential danger, the specter of any such threat is abstract and not immediate.
On the other hand, the tech is obviously too economically essential and competitively vital to risk 'falling behind'. While there will certainly be attempts to ban, limit or constrain AI, the well-funded, highly organized commercial interests and civil libertarians will deploy lobbying, legal challenges and public opinion to ultimately prevail.
In the U.S. this seems highly unlikely
Aren't these the same guys who won't even let us have Chinese cars?
I'm not as confident as you that they will keep allowing us access to technology as strategic as AI models out of China and elsewhere that undercut US models in the market.
To everyone reading, download open models from anywhere as soon as they are released. You really have no guarantee at all that access to those models won't be cut off in the future with the stroke of a President's pen. Those downloads are your insurance policy. You'll always be able to access whatever you've already downloaded.
I would definitely pay a monthly subscription to help fund a non profit compete with Anthropic and OpenAI. I already pay subscriptions for myself and 2+ other people. It's a non brainer to be able to pay for the training of better models that I can then run myself for many more. I hope someone starts this, I think this model would work. I'd start it today if I had the team and initial capital to bootstrap the infr. I know VCs won't fund it, but we definitely will, enthusiastically and continually.
I'm a bit skeptical of the token cost/ROI for all models, but sunk costs are sunk.
It has the feel of self-improving super-intelligence or bust to me. If you get that, the frontier model(s) run away with a faster exponential. It's a bit like semi with Moore's Law with silicon, GaAs could never catch up. If you don't get it, the fast followers crush the high investment and there's no moat. Not like they can enforce copyright!
There's a world where frontier models run away with a faster exponential and still go bust due to being outcompeted on efficiency.
There's a point past which "intelligence" stops mattering as much, and IMO we're already there.
Consider which would be more useful (and profitable for its creator): a model that is 3x/5x/10x as "intelligent" as Mythos, for whatever your favorite yardstick of intelligence is? Or a model that is as "intelligent" as Opus 4.5, but can run at reasonable speed on a typical consumer laptop/cell phone?
If the former can develop the latter, then the former.
not really a feeling; if you listen to ed zitron and strip out the vitriol, you still get the fact that the VCs are looking for some 5 trillion dollars in 5 years.
The onlly way that happens is if America turns into zimbawe.
Who says it hasn't already?
I've driven in a LOT of the USA. Sure Chicago, NYC, DC, LA, LV... They're all built up and feel modern.
Try driving anywhere in the Midwest outside of the big cities. Dilapidated carcas buildings everywhere. Urban and rural blight. Only jobs are low paying service work. Its bleak. Like, really bad poverty as a disease bleak.
And its crazy watching it too. They're ignorant (involuntatily), poor, and trapped. And democrats only seem to care about special interest of the week, so these areas vote republican.
I don't have a solution btw. Just something I've seen growing in the last 25 years. And its getting worse, not better.
At risk of greatly oversimplifying American politics, it's truly impressive how good Democrats are at shooting themselves in the foot.
That's why I personally believe Sanders and Mamdani have found so much success with the working class; they keep themselves separate from the Culture War slugfest that mainline Democrats either voluntarily engage in or let Republicans drag them into.
IMO the vast majority of those "culture war" issues (LGBT freedoms, etc) are incredibly important, but to the average poor rural American it feels incredibly distant from their day to day. I can't put my finger on it exactly but Democrats have a tendency to message on those issues in ways that are either counter-productive or get soundbit saying something moronic. So when Fox News and whatnot say that Democrats are prioritizing other groups over them and message on it day after day, it's not hard to see why that propaganda becomes effective.
That's not to say that Sanders/Mamdani/etc don't message on social issues, they obviously do, but they are somehow effective at not alienating voters who may otherwise latch onto that in a negative way.
I don't have a good solution. Just my observations.
It’s very difficult to be for the people and also get hundreds of millions of dollars from wealthy donors. That is the democratic party’s plight.
I'm not sure I understand what part of my post this is in reference to. I agree that the democratic party gets a _lot_ of money. Are you contrasting to something in my post, or agreeing?
Appreciate clarification, I could also just be dumb today
As a Canadian driving through Michigan/Ohio/Upstate New York and places look run down and like they peaked in the 60s-80s and its been down hill since.
When I was younger and it was a new thing it was quite a shock since American media obviously doesnt portray it that way. It is quite a contrast to even how the more run down parts of Canada look.
>Who says it hasn't already?
Great.
My prediction is that hardware costs will make open source models impractical for the foreseeable future.
Yes, tinkerers and enthusiasts will continue to make use of them, but frontier companies will maintain near total dominance because they will be the only ones with access to the hardware.
Meta is selling their now excess compute, other compute has been on the market for a while. The current hardware cost bubble is temporary, especially once people are forced to pay the real inference price instead of majorly subsidized subscriptions.
yep, all those coders paying $200-$500 per day to use claude once subsidy ends will be seriously rethinking how much they really want to vibe-code "rewriting X in rust". Helping people write word docs, recipes, and emails isn't going to justify $15K per month subscriptions either.
> Helping people write word docs, recipes, and emails isn't going to justify $15K per month subscriptions either.
Those things can all be done today on a $250 used video card and pennies of electricity
Not really. Models that can run on a $250 card may be able to produce a decent word doc sometimes, but the fidelity of such models is so poor that it's no longer very attractive for general use on those types of tasks.
The subscription is certainly subsidized but its no where near 100x cheaper than API prices.
Heck, most large enterprise moved to usage based billing and are still happily paying for it. They are force multipliers for your top talent, and when a top engineer is being paid $500k a year, doubling their output for $500/day is a no brainer.
Lowering the cost of hardware still won't solve the issue. HBM and DDR5 was never cheap, even before the shortages, so selling a full inference system is beyond the acceptable price range for most casual customers.
We're going to see Apple and Google compete over services and AI/OS integration instead, it will probably be years before your OEM takes local models seriously.
Apple and Google (via smartphones) are in literally everyone's pocket.
Running KIMI on a phone is not possible today and I agree with you that it will "probably be years before..." it is.
But how many years do you guess? I personally do not think it will take even 10 years for the situation to be commonplace.
> But how many years do you guess? I personally do not think it will take even 10 years for the situation to be commonplace.
IMO it won't be possible for the foreseeable future. There's essentially zero possibility that phones will gain the hardware capacity to run today's Kimi, so the only other alternative is to squeeze the power of today's Kimi into something that can fit on a smartphone, which also seems fairly unlikely considering the current rate of progress.
A bold prediction. Phones have gigabytes today. There are famous laws of growth that put terabytes at just a few years away - perhaps 10 isn't too bad an estimate?
> I personally do not think it will take even 10 years for the situation to be commonplace.
Do you personally remember how far smartphones progressed in the past 10 years? It's not as long a time as you think it is, the limits of what a smartphone GPU is capable of did not substantially change in that time. Nor did the amount of onboard RAM that we include in the package. This is true even for Nvidia's ARM SOCs, frankly.
Apple, Microsoft and Google all eventually want to enforce OS-level lock-in for the most profitable AI services (eg. their own). It's much more attainable and profitable to use that lock-in to sell you exclusive service integration, the local AI revolution probably won't begin on their hardware.
Only if you narrowly define AI as talking with a chatbot. Today, on an iPhone, there are a number of features that only work due to some sort of a model running locally. OCR and intelligent object selection in a photos, and summarization of texts are local models running on the iPhone hardware, they're just not a chatbot. Yes, there's also cloud backing various features, but the idea of running models locally isn't foreign to Apple. On Google's side, the Pixel 10 Pro is powerful enough to run quantized local chatbot models locally today. Local translate is a model, and has been for a while. Both corporations are going to sell whatever customers are willing to pay for, in money or via ads, and if local models get good enough that customers actually are willing to pay for it, I have no doubt that it's Apple and Google will go that direction. It's Google that's releasing Gemma models for download, and they're a big enough organization that the left hand doesn't know what the right hand is doing, so the one conspiracy theory that they'll never do local models because they only want to profit off hosted models is too simplistic.
For the near future it seems that the new models will consume whatever improved hardware capacity we have. Competing with that is challenging, but I also think there will be strong economic incentives towards cheaper but adequate models on other providers.
I don't think we'll see home users being able to match even the low end clouds for a long time.
Longer term I think we'll see these uses of AI cluster into a few groups:
- maximal code / reasoning quality, at high prices (Fable)
- typical code / agents (sub-Opus, Terra)
- cheap but decent enough quality (think Deepseek / GLM / Luna)
- so cheap I don't care about utilization (Deepseek, and friends)
And also more niche ones:
- ultra fast with high quality answers (typically sub-SOTA). Cerebras / dedicated silicon type approaches, expensive.
- ultra fast with mostly-adequate answers, and an openness to retries, moving up to better models
I think the open models will dominate (not with individuals, but low cost providers) all except the top 1-2 of those categories, and there will be a continuous erosion on the big player's moats. The top categories are also where all the money is, but I'm not sure it can justify those investments long-term. I also think they will have to squeeze more money out of them to justify the investments, which will also drive people down the list.
Edit: clarifications.
>For the near future it seems that the new models will consume whatever improved hardware capacity we have.
But they're not. Meta, SpaceX, Microsoft, Amazon, they're all leasing out capacity to others. If they were truly constrained, we wouldn't see that happening.
You may be correct but the combination of local models when they are fast enough and work, combined with paying close to zero for deepseek v4 flash from US providers is pretty good. When you need it, glm5.2 is cheap to use and very good for working with larger projects.
There will be plenty of model providers with prices that undercut Anthropic/OpenAI's prices.
Doubtful for the same reasons. The frontier providers are working at a hardware scale that will make it impractical to undercut them.
I wouldn't rule out the possibility completely, but it won't be very common.
They also have to pay for staff / model training. Those are not cheap.
Are frontier models actually astronomically expensive to train? GLM 5.2 was trained on ~30T tokens, so ~10^25 FLOPs. Say you get B300's for $5/h (pretty high), and you get 50% MFU; that's ~$15M. Of course there's also a bunch of risk that the training itself goes badly, post-training, etc. But still, compared to the inference spend after, it's not that crazy.
Nobody spends 15M on inference only to check at the end that their money was wasted.
Completely agree. Once I can reliably get open models doing what I am on Fable ultra I imagine I will switch for good. I am fortunate to have access to a decent bit of local RAM, 192GB of DDR5 at an OK speed. It is not enough and costs are well past absurd. In a few years time I envisage a setup that is sub $10k which can accomplish such tasks. The pace so far has been breakneck. That is all I personally need. That may change, but until true AGI I do think there will be a ceiling to how much I will pay for something frontier if it is only marginally better.
This is easier to say as Fable is good (even SOTA). But people have been were saying this continuously for the current model and for now the improvement are still coming.
A better question is would you settle for o3 now or pay 20$ or 200$/month for fable ? Because o3 quality is available OSS.
It is like the new IPhone, in some sort. At some point come a feature many would like to have, despite diminishing returns.
We will see how long labs can keep up and what the scaling curve look like, but I would be more worried into losing sota status to Chinese companies than letting them take the open non-sota approach.
IMO we are there almost. I had every iphone until 14 pro and still use this today. I know what the top of the s curve feels like. From a pure model standpoint weighed against every day use cases for every day people (the right way to measure when comparing to something as ubiquitous as a phone), the models already have diminishing returns.
I think there is also the case were companies will simply use different tiers for different tasks.
While the engineering team might need a cutting edge model (with the associated costs), the marketing department will be fine by something that can grammar correct or turn a few bullet points into prose. Likewise you already don't need Fable for Ticket -> RAG -> Reply with Faq knowledge or escalate workflows
That's already the case with other very expensive software like CAD packages were oftentimes you have different feature sets enabled for different employees.
"that can grammar correct" ... Did you do that on purpose?
damn! how muhc you dropped for 192gb?
Speculation: Anthropic and and OpenAI become more like hedge funds, keeping their models to themselves, using their them and their compute for market/event prediction, internal tech development, and AI self improvement. I already wonder if that is what Google is doing. Why release your best models to help the competition? Their published models are good enough for 99% of consumers, and they can leverage their market dominance in other areas to put that consumer model in front of eyes. Then the decision to improve your internal models would both depend on whether you think they'd have value internally.
Yeah, I'm pretty sure AI is going to go insular within the largest companies. This will only be hastened by the growing national security concerns/awareness.
They would have done that long ago if it was that easy. The fact that they aren't tells us something about the actual business utility of leading models.
Except Deepseek going in opposite direction: from hedge fund to an ai model provider.
> open models is what will kill Anthropic and OpenAI.
Maybe killing Anthropic is a good thing? Anthropic believe that they are the moral god and they get to determine what we can do and can't do, and they get to tell us what model to use and what not to use. I find it very counterproductive.
If consumers ever have enough VRAM for it locally, sure, otherwise nowhere near close.
Nvidia is looking like they are ditching consumer markets in favor of enterprise GPUs since nobodys heard a peep about the next iteration of RTX cards. The 60xx series is postponed till 2028.
Nvidias playing a dangerous gamble, in my eyes I see all the frontier labs eventually just only buying Nvidia chips for training and building custom ASICs for a fraction of the cost, longer lifespan and cheaper to host.
This will eat their 5 year gravy train for GPUs vs the 10 to 15 for ASICs.
I think this is possibly true, but the other piece of the equation is tooling. Right now, Anthropic has by far the best tooling around (for both SWEs and non-technical users), and a huge ecosystem of integrated ISVs. I'm not suggesting it will happen, but if Anthropic decided to provide options of using models beyond Claude, they'd still have a significant moat.
I dont know. The tooling feels trivial. I swapped from claude to codex and back with zero friction. If they were very expensive to me personally I would probably just use them to build the harness and then plug Kimi into it etc.
The real moat aren't the models, but the tooling around the models that allow them to perform specific tasks/goals. That's what really sets apart frontier vs open. Open only has the model itself, closed have the tooling to enhance the model.
> The real moat aren't the models, but the tooling around the models that allow them to perform specific tasks/goals. That's what really sets apart frontier vs open. Open only has the model itself, closed have the tooling to enhance the model.
As these frontier companies have been boasting, writing software is now a negligible cost because the LLM can do it.
IOW, no, their software can't be a moat, because, according to their own arguments, you can use their LLM to trivially clone their software.
> you can use their LLM to trivially clone their software.
Perhaps, but what you can't clone is familiarity, polish, integration, and network effects. These companies desperately need a moat, and so for example all the new additions to Claude's web interface are becoming the products that the vast majority of people will use and be locked into (if it goes according to their plan).
If development is trivial and capability remains the differentiator, then "familiarity, polish, [and] integration" are non-issues.
I'm not sure what network effects even means in the context of llm selection.
I'm happier with my Opencode Go running with the pi harness than Claude Code at work currently. I'm not sure if this will be that big of a moat, at least in the software dev market.
unfortunately, for me, that's an anti-moat; their dark, inconcievable, alignment, random "no goblins" and inability to be reliable where their models are inevitable non-determinant means you're going to constantly run into the "this worked yesterday" problems, no matter how smart the models actually are; being filtered constantly through economics, political and ego-maniacal filters means everything you do with it will break at any given interval.
good luck.
I agree with your speculation to be honest. And yet I’ve tried several local open weights model now and none gives the same quality of answers as Claude gives me on a regular Sonnet model. Mind you: I am “running” 48GB of RAM so I can’t try every model. Where does this difference come from? Can we actually get close locally?
You’re running 48GB now but imagine a future where everyone has 512GB RAM or 1TB RAM in their computers (it might sound like a lot but also 20 years ago we had 512MB PCs).
It’s not hard to imagine what 5-10 years of pressure to increase RAM will do to specs, on top of the normal tech improvements.
That’s worth bearing in mind when thinking about local models.
Plus, local models keep getting better and better; 2 years ago what you could get out of those 48GB of RAM was embarrassing compared to what’s doable today.
We’re getting there. Just takes time.
Something I've noticed is that local models are giving better answers these days than they did a year or two ago, even if the size (in parameters and in the amount of RAM used) hasn't increased. I'm not familiar enough with the technical side of model training to explain how they're doing this, but I think in another couple of years, models that use up 48 GB will be able to squeeze out even more incredible performance.
Though on the level of something like Sonnet 5... well, maybe not.
>open models is what will kill Anthropic and OpenAI
i doubt it. it cost money to train a model. we can see that with the price increase for Kimi3. Chinese AI companies is leaving a lot of money on the table for third party providers. you think they going to let go of those money that they can make. sooner or later they will want to collect. after all, none of Chinese open weight model is release by a none profit. its all for-profit companies that is releasing open weight model.
I can't take anybody seriously when they keep declaring open models is beating frontier models. What they don't understand is that besides the huge capex to train and run inference, the real gold is in the human response to the prompt results, this is what all the Chinese companies are making their open models dirt cheap and distilling american frontier models via scraping.
The idea that we can out-parameterize frontier models is a common misconception, the true moat that Anthropic and OpenAI is why Chinese model providers are open sourcing and making it dirt cheap to keep pace through its "proxy chain operators"
https://x.com/HarshalsinghCN/status/2056626175959826692
> Apple can make them smaller and put them on the device
Someone can, but Apple has essentially admitted defeat and handed the reigns over to Google.
Have they?
https://www.cnbc.com/amp/2026/07/14/apple-prismml-ai-compres...
> handed the reigns over
Oh man, they gave them free reign?
How will anyone reign them in now?
For all intensive porpoises, this is like Babe Ruth, chomping at the bat!
They'll lobby to ban them, especially Chinese models, as Amodei is already doing.
I worry about that but here are two positive things:
1. it would put us (USA) at a competitive disadvantage, and cooler heads will prevail in this fight
2. there are good US open models. I have the latest gemma4:27b with better tool support functioning at a high level in the pi coding harness. Thinking Machines seems to be on a good path, we will see what they and other US companies can do.
Cooler heads, in this administration?
> at a competitive disadvantage, and cooler heads will prevail
I don't think this a safe assumption at all, given the the last few years of government policy where the "cooler heads" very much failed to stop such things. Either by partisan choice, or because they were silenced/fired for telling the Emperor that his clothes weren't there.
In particular, consider the tariff-taxes [0] against on US importers, and the ensuing breakdown between America and its oldest allies and trading-partners. Also relevant would be the President's recent pet-project war [1], which has shut down a major part of international shipping.
_______
[0] Unilaterally imposed by fiat, with idiotic "failed economics 101" math for the rates... and the majority in Congress still actively tried to keep it going, until the Supreme Court ruled it illegal.
[1] Simultaneously a war and not a war, both already-won and ongoing, depending on which lies need to be told today to harvest applause and grab cash. Just this week the White House submitted documents to Congress claiming what's going on is a completely separate and new, second war with Iran...
There is a massive difference when you zoom in close or take an angled perspective. You can manufacture uniqueness. The issue is when it comes to every day use for every day people there is no differentiation.
You can run the same harness on fable, opus, sonnet, and see a huge difference between them. It is true the harness is important, and openai has begun encryption its instructions to swarmed sub-agents instead of just encrypting the chain of thought, but the model is still important at this stage.
This will only delay the inevitable. Sitting on some magic prompts is hardly the moat they need.
Referent of "the models are meaningfully different" reads as <top closed, top open> rather than <top closed, cheaper closed> to me, so I'm not sure why we'd be comparing Fable vs Opus/Sonnet or Sol vs Terra rather than the same against Kimi K3.
I was just responding to "The harness is what takes these random and hallucinogenic models and make them into something deterministic and useful."
You can compare Fable vs Sol vs Kimi in the same harness if you want too and there are meaningful big differences. I chose all Anthropic ones to be safe from the they were finetuned on different harnesses complaint that would be made from that comparison.
Haven't tried Kimi K3 for now but there was a huge difference between GPT 5.6/Fable and GLM 5.2/Kimi K2.7 that were previous frontier open models.
I hope so, but I am not holding my breath. OpenAI and Anthropic have too much money riding on them. I'll be surprised if they die.
I still strongly believe Google Gemini has the best position for one simple reason: model maintenance. Accurate information is a moving target.
Open models are indeed very capable, but they will eventually become more specialized to the application to keep an edge. It makes perfect sense that the future shape of AI conforms to the landscape it was born out of.
You're saying it's important to have up-to-date facts stored in parametric knowledge? It seems to me like that's grown less and less important as agentic capabilities have grown. Even if a frontier model doesn't know something, if it's out there, it can easily find it through tool use.
Grok has the biggest advantage in current events knowledge because it's integrated with X, which enough people still use even though it isn't Twitter.
Hm, when compared to all the information people with android devices share with google, or those with gmail, ..
Google wouldn't give Gemini direct access to everyone's Gmail. It would carry far too much risk of being exposed. X can, because tweets are public.
But google can give gemini access to summaries of all the avaiable information google has, including what people look for right now and where and how long.
Eventually they will kill the hyperscalers too because of privacy issues. It's better for a company to pay an uprfont cost and then run everything on premise that uploading their entire codebase to a third party service.
The vast majority of companies are still putting most things in the cloud and will continue to do so and this far outnumbers the must-be-on-premise companies.
Sure there will be self-hosters but hosting AI models will always be more of a challenge than running scalable database on your own hardware and specialized hyperscalers will be here.
Would that require a watershed event to clearly establish the importance/risk of privacy though? For example, right now it seems like most big software companies w/ strong security process are comfortable uploading entire codebases to Israeli cybersecurity firms for vulnerability scanning compliance purposes
The frontier models are obviously superior. The question is if progress slows down.
in what context, and at what ROI?
Because... I have use-cases where this is true, and use-cases where this falls flat on its face.
I don't actually think it's obvious (at all, really) without defining what "superior" means.
In the same way that I don't think it's obvious that a plane is superior to a car, or a boat, or a bike.
They each do things the others don't, and excel in different spaces.
The thing about not much difference between models and the harness making them deterministic and useful is wrong. Also models have different strengths and weaknesses and some are better at almost everything by a large margin compared to others.
As for your speculation, I think it's hinging on some companies releasing models for free or no big differences between models. In a world with hyperscalers and companies training models you can quickly recreate Anthropic or OpenAI by having an hyperscaler ally with a model training company, train a good/a better model, and not release it.
> Their marketing depends on people believing the models are meaningfully different, as people have sweatily argued on this forum.
They are noticeably different. Benchmarks, anecdotes, all say the same thing.
Now, is a ~6 month lead actually worth 1 gajillion dollars? Maybe not.
You do realize OpenAI survived for years without LLMs, right? There is still more AI research they can do as a lab even if they stop experimenting with LLMs.
I think they didnt have the amount of debt /negative cashflow as they do now.
That is on OpenAI Group PBC. If that bankrupts it doesn't bankrupt OpenAI, Inc.
Just like opensource search engines killed google
oh wait
I don’t even know the names of any open source search engines, but the open source models perform decently on various benchmarks and in personal experience.
Was it ever even a claim that open source search engines were trying to outperform google, let alone kill it?
Yacy tried in the 2000s. I'm sure some magazines made headlines posing the question whether yacy is a google-killer
Marginalia already gives me better results for many queries than Google. Because Google has sunk so low.
Open models are 4k TV (or maybe 1080p tv now and 4k TV soon) and SOTA frontier models are 8k TV. Can I or the average user tell the difference? Not really. Would they pay for that difference? Not a chance. Our entire economy is teetering on some future hope that this fragile and immaterial difference will pay off, when the reality is that LLMs are a race to the bottom and eventual razor thin margins. Maybe a tiny vocal subset of programmers can use it for work and make paying for it worth it to them, but that can't prop up an entire economy, especially when said programmers are phased out, jobless, and replaced by AI with each better iteration...
Except when we upgraded to 1080p from older TVs, they got bigger. Now with 4k they are getting bigger yet. More powerful models means new use cases that didn't make sense on the weaker models.
Even in the world where all models are basically equivalent (a thesis I don’t buy, but will grant you for arguments sake) - I believe there is much more to the AI business than just training and running models.
It’s a very new set of technologies, and understanding what is useful to customers and what isn’t is the whole game. Call it, product taste. There were a million cell phones before the iPhone took over the world. Why iPhone? Product taste. There are a million startups, and only a select few become unicorns. Why? Product taste.
>There were a million cell phones before the iPhone took over the world.
You have tripped yourself up there.
iPhone took over as it introduced something innovative over standard phones, but then Open Source (Android) matched the multi-touch and software differences and Apple's branding, lock-in and design etc have managed to keep it as a big player in wealthier countries. IPhone also came on the back of the massive iPod success.
ChatGPT launched the same innovation vs Google Search, but just like Android Opensource AI is moving fast now.
Android has 72.7% market share at present, Open Source AI will do the same unless the frontier labs can continue to do something new.
The frontier labs are saddled with enormous investor and other debts. How long they can keep innovating by spending so much on R&D and paying there staff very high wages remains to be seen.
Once investors cash out via an IPO, the companies are back down to earth and playing in the real world again.
Android has market share, but Apple makes all of the money! I find it really funny when people attribute Apple’s success to “oh, the only reason they succeed is design and marketing.” Yeah, I mean factually speaking design and marketing actually do matter a lot!
Us developer types like to pretend like specs are the only thing that matters? If you could have a 10x more powerful model you could only access running locally through your terminal, versus a weaker model through a clean web interface, normies will pick the web ui every single time. Product experience is simply everything, as much as we like to pretend like nitty technical decisions are the most important thing.
> Android has market share, but Apple makes all of the money!
So? The benefit of open source is that you don’t have to worry about making a ton of money. You just need to be viable.
Apple: premium product a minority is willing to pay for
Android: standard product the majority use
I’m sure there will continue to be iPhone equivalents in the AI world, premium bespoke models. But the vast majority of people will be happy with a cheaper offering.
The original comment was “open models are what kill OpenAI and Anthropic”, which to me is as silly as saying “Android is what killed Apple”
Well I think a critical difference is that, unlike Apple, OpenAI and Anthropic have taken on so much VC funding that a 20-something % market share is not going to be enough for them. So open models could kill them, not because of the techonology but because of the way they're financed.
> If you could have a 10x more powerful model you could only access running locally through your terminal, versus a weaker model through a clean web interface, normies will pick the web ui every single time.
More like if you could have a 1.25x more powerful model that you could only access through some weird surveillance megacorps aggressive monetization scheme, or choose from 100 others running open models and accessible through 100 different interfaces pandering to every taste.
Normies will choose the megacorp every time, because that was the one in the tv commercial, and within six months will have left for one of the others in a rage.
The only corporate hope is that the government steps in to ban their competition.
There were many smartphones before both iOS and Android.
While that may be technically true for a strict definition of “smartphone,” there’s no denying the iPhone redefined the concept in a way that its competitors were forced to copy to have any hope of keeping up. Nobody hears the word “smartphone” and thinks of a Blueberry or Treo anymore.
What exactly did the iPhone do better?
It preceded its release with the iPod craze, making it a lifestyle product rather than something marketed for its capabilities (which were very humble compared to other contemporary devices and only started to catch up in second gen).
Perfected multi-touch touchscreen
Before that we had touchscreen but they sucked.
---
I really don't think multitouch is what made or broke the original iPhone.
That's a subjective question, so I'll give a subjective answer. The browser, for better or worse, was a lot less dumbed down for mobile than competitors, the stylus-less touch interface reduced UI friction and the odds that you'd lose a critical (if inexpensive) component, and the slew of contemporary iPod users could easily migrate their libraries over.
Exactly 4 months ago, the marketshare on openrouter was 60%-40% in favor of closed models. Now it's 63%-37% in favor of open models. On March 19th, the open models processed 888B tokens in aggregate, yesterday, they processed 4.19T tokens in aggregate. That's almost 5x in 4 months! I can't think of the right intensifier to describe this level of growth.
If you are looking for more details (as inferred by openrouter data), I built a dashboard that updates daily: https://dirac.run/labs-market-share
Correct me if I'm wrong, but OpenRouter doesn't tell the whole story.
If I'm using OpenAI, Anthropic, or Google models, I'm probably using their API directly, so OpenRouter won't have those stats to compare to.
All that said, it is very exciting to see open model usage grow via OpenRouter.
It doesn’t show the whole market of course, but the trend lines are interesting
Definitely interesting seeing it grow there. I'm now curious to see if it's overall market usage growing for LLMs (with more OpenAI and Anthopic usage in a similar growth rate), or if it's more heavily weighted towards open models specifically.
With the price to performance of the latest open models, I think most cases for integration into applications would get a best bang for the buck from an open model.
would love to see a statistic by model, and maybe some sort of classification to get a sense of how good the model is and how much it costs.
edit: this exists https://artificialanalysis.ai/
Why would I use open router with claude? No thanks
Interesting to see, but I also think more and more usage is moving towards subscriptions, which isn't captured by this metric.
(both my personal and corporate use has done exactly this)
This right here is going to be considered one of the first major signs of the downfall of closed models years from now.
And look, if you disagree with me PLEASE tell me why. What moat do these companies have? I genuinely want to know because looking at the spend for companies like OAI and Anthropic with no actual moat I can identify is actually driving me insane.
I think the frontier AI companies think you're missing key details:
- they can still discover entire new untapped markets for AI (that, potentially, only their models can unlock)
- they can find (novel, unique to them) ways to drive down the cost of running their models
- they can provide other ancillary value (e.g. write better harnesses) because of their expertise, and then charge for that value
I'm probably missing a few bullet points also. However, none of those are moats (or at least not yet) ... I'd consider them more like bets. The frontier AI companies are betting "the house" on them, and if they pay off they could, hypothetically, make them financially competitive.
> The frontier AI companies are betting "the house" on them, and if they pay off they could, hypothetically, make them financially competitive.
They are not betting the house, they are betting the American economy on it. When this crashes it will take everyone down.
I've been keeping my cash/stock ratio a bit higher than usual recently because I'm waiting for the bubble to burst. My "stock" side is mostly a company I work for that is riding the wave.
The moat is enterprise contracts and artificial friction moving between harnesses. Moat is a strong word, more of a puddle.
If we use the same logic as the AI boosters and just assume we're in exponential growth, that means in a year open models will be handling 523T tokens.
The prose is, of course, LLM-generated. https://www.pangram.com/history/29a71663-e6b2-4db6-87bd-b943...
I'd be curious to learn more about how/why executives come to sign their names to this kind of writing. Maybe it feels like a natural iteration of the pre-existing experience of signing one's name to an assistant's manuscript, or a press release from staff in Comms? I don't know execs who do this kind of thing, so would be grateful for any observations others might be able to share (anonymously or otherwise).
Aside: I think it would be very good for the culture if HN automatically ran Pangram on any text-centric articles. Even perhaps just those which hit the front page? I have my own browser plugin, of course, but judging by comments here, at least some seem not to notice?
Agreed re: automatically running Pangram, maybe once the articles hit some vote threshold (or just, frontpage) in case cost is a concern.
I think aggregators running AI checks automatically seems quite good, rather than individual readers all doing it on their own sporadically. We recently implemented auto Pangram checks on Manifund, and I think LessWrong has done this for a while now.
I agree that it would help the site a lot but Pangram is notoriously inaccurate.
This presentation is painful to read. It's an LLM's idea of a CTO presentation. I'm overwhelmed by charts, only slightly connected to the text around them. But no matter, it looks like a CTO slide deck. HIGH IMPACT.
Much better would be if the CTO of Mozilla had actually articulated their own analysis.
As someone that's generally for the proliferation of open models, I want to take this seriously, but it's really difficult when it was clearly written by AI.
I guess they fired whoever used to write copy for these things.
Edit: to be clear, I'm not trying to just dunk on them, I think it's actively hurting their own point to do this, and counter-productive when people can easily clock it - it makes some percent of the audience immediately tune out.
> Parity reached. The contest is one layer up.
I want to vomit reading that.
Agreed. Maybe we've become jaded to the writing style of AI, but it reads as disingenuous.
> Mozilla exists because one company tried to own the front door to the web, and an open community rose up to make sure it never could.
I'd say that the front door to the web is pretty much owned by Google and Apple at this point given Firefox current marketshare. And maybe that's enough, maybe a future where a low percentage of open models keep the rest of the system honest but that doesn't seem the argument of this article
It's also just a very cherry-picked framing
Mozilla exists because one company owned the front door to the web, and another company abused their market position to push their free (as in beer) browser. Mozilla is the phoenix rising from the ashes of that first company.
Then another company came along, abused their market position to push their free browser, demolished Firefox's market share but keep handing them cash to avoid the appearance of a monopoly
Mozilla exists because Google gives them billions to keep Google as default search engine.
And to keep around a concurrent browser with full adblock abilities so they can cripple their own without too much of an outcry.
Do you think Firefox development is unsustainable without this substantial stipend?
At their current trajectory, absolutely. But if they focused their development efforts, cut down on careless spending especially to third parties, and looked for alternative sources of funding (it is presently not possible for users to donate to Firefox development, for example), I feel like they could survive without Google for at least a little while. Frankly, I think these things should be started on now, while they still have Google's backing.
It sure is nice to see that Mozilla is still doing all that they can to keep on top of current trends, except developing a decent privacy-focused web browser for developers and power users.
Firefox's AI Chatbot feature only proposes one open model provider (Mistral) and zero local option. They're not walking the talk.
They get paid half a billion a year to do exactly what they have been doing and keeping FF at <5% of market share on desktop and 0% on mobile.
I’ve been using Firefox on mobile (and desktop) for years. I still don’t understand all the hate thrown at them. Whatever the downsides/ shortcomings people see, they are irrelevant to me, because I hate Google more & refuse to give them my data. They are good browsers & work just a well as Chrome 99.9% of the time for me
But they supply all the source for those projects to flourish and have an ecosystem of their own: Librefox, Iceweasel, Reynard, etc, etc.
That's a very low bar, though. Chromium is open-source too and has a bunch of privacy-focused forks such as Ungoogled Chromium, Brave and Chromite.
This used to be my attitude, but I do feel like Firefox has made good ground in 2026.
I've also become sympathetic to their AI strategy. They don't seem delusional in their approach. They're not building models or selling slop, they're building OSS compatibility layers.
I don't want to see a world of vertically locked AI, so if Mozilla truly can dust off the old playbook for OSS AI, we'll all be better off for it.
Yes, Firefox itself is a “general purpose” browser, and that’s probably for the best in terms of wide market appeal. Other developers have taken the engine and made power-user-focused browsers with it. Lately I’ve fallen in love with Zen (though after 2 months of use its pinned tabs features still confuse me a bit).
https://zen-browser.app/
https://stateofopensource.ai/state-of-open-source-ai-2026.pd...
the pdf is easier to read
I am sad that there doesn’t seem to be any community whatsoever around _truly_ open models that are released with source data and training methodology, such that they could actually be reproduced given the resources. We’ve allowed the term “open” to be diluted to a shocking extent.
The design and layout made it harder to read than it needed to be.
Regardless, the inference costs dropping almost 50× is really amazing to see. And now Kimi K3 release has shown how open models are getting closer to the frontier level already. Open source AI is moving a lot faster than Anthropic and OpenAI would have expected lol.
> The cloud era already ran this experiment: proprietary APIs plus data gravity made exit punitive. The repatriation wave is the receipt.
Some sentences smell a lot like AI.
Virtually every article about AI these days is written by AI. It's obvious why, but boy is it nauseating to read
Almost all about open weight, but the title says Open Source.
When europes mismanagement of AI suddenly turns out to be the best thing that has ever happened. At best only few billions lost to training models that become obsolete next month.
I see a gap in the ecosystem: too few mature open source harnesses.
I’d like a community led, BYOK, modular project where I can define, orchestrate, monitor and maintain agents.
Of course this is a new area and projects like this take time. But still, IMHO, a gap exists.
Unless someone wants to recommend their favorite FOSS tool; please do.
I use opencode as my harness which is pretty ok. I daily drive it with no issues. It's not going to do any orchestration for you but it works.
I'm slowly moving to Pi but it's a very DIY system. Think neovim for agents. I think it has the potential to be where I end up for a while but it's a long road.
I’m not ready to celebrate the victory of open models just yet considering all the good ones are built by private, VC funded companies. How long will they continue to be charitable? And what’s their actual business model once the money stops and investors start demanding returns?
Gemini had a pop-up on my phone today, asking me if I wanted to "bring Gemini up to speed" by importing conversations from my other AI apps. This tells me that Google is threatened, or data hungry as always, or both. Open source AI, Anthropic, and OpenAI, knocking on doors.
I don't love the appeal to romanticism portrayed in this article.
Maybe its the wildfire smoke in my eyes, but that font choice feels aggressive.
It's their own fonts: Mozilla Headline and Mozilla Text.
No idea why they'd be using the display font for the abstract though, that kind of defeats the whole purpose. It's supposed to be quirky and bold, but used far more sparsingly.
Forget aggressive, I just find the main text font harder to read (vs. ... you know ... a normal font).
It's AI slop
It appears open models were used to create this slop.
That opening is so hard to understand what they are trying to say, from the font and how it's written. It took me several times rereading to even grasp.
Plus the article is filled with cryptic things like:
What?! Is it a meta answer to "the state of open source AI" question?From the title of a chart:
> The venture-funded open-source ecosystem: total disclosed funding, USD M
> Bars grow as you scroll.
The bars, in fact, don't grow as you scroll. And I don't even see why they should.
On my device, bars grow as I scroll. I want your feature, being able to just scroll the static page without elements jumping around.
> The bars, in fact, don't grow as you scroll. And I don't even see why they should.
On my device, they grow as I scroll to them.
I think it’s supposed to mean “open source is easily shipped, but open source is hard to deploy”? Or perhaps “deploys hard” is a figure of speach, as in “we are deploying this open source and we are deploying it /hard/“? I don’t know, it’s not good.
This is truly some proper slop. The "PRODUCTION RATE BY COMPANY SIZE" graph has bars that start offset from the text underneath them, which LOOKS like a mistake that happened due to word wrap, but if you visibly compare the 54% to the 55% bars they seem to have compensated for this?! I can't tell if his was on purpose or accident and it's impossible to take the data seriously!
This is on mobile in portrait. In landscape the text doesn't wrap or offset anything.
So good at style, so weak on substance
Is the CTO a bot?
Can a vending machine operator do better than Mozilla management? - that's the golden question.
> Open weights are no longer a compromise. They are where the work happens.
> They require owning the layers above it — the harness, the memory, the permission model — while those layers are still open.
> Open isn't a vendor choice. It's a sovereignty choice.
Quick fix for the font, which many people are (rightly) complaining about.
The issue is that all of the text is a quote, and that renders enormous. That’s probably fine for a tiny quote amongst more text, but here it is jarring.querySelector/querySelectorAll() are great for plucking out deeply nested elements but if all you need is to find all elements of a certain class and the API gives you a tool to do exactly that, why not do that instead of reaching for the general-purpose Swiss army knife? Sure, the execution speed difference may be only measurable in microseconds, but it takes about the same amount of time to type so why not use the specific tool?
This new trend of content appearing while scrolling down is so terrible accessibility-wise, I do not understand how Mozilla of all institutions would do it.
Not every trend needs to be followed. Have some backbone. You receive donations to have that.
___
Apart from the website being - frankly - bullshit, the content is also - frankly - bullshit.
It's just on the frontpage because the title says "open source AI".
> This new trend of content appearing while scrolling down is so terrible accessibility-wise, I do not understand how Mozilla of all institutions would do it.
Could you explain what is wrong with the accessibility of this page? All the content is included in the html payload, so it is accessible to screen readers and text-based browsers; and as for the "reveal" effect, it seems to respect user's choice of "prefers reduced motion" and is disabled when that is user's preference.
> it seems to respect user's choice of "prefers reduced motion".
Cool, that I didn't check, because it is impossible to enable that setting, as it breaks _huge_ amounts of websites.
I'm not aware of a way to enable it selectively, but one could also just display the content at all times. It's a static page. It's static content. None of this makes any sense.
___
The idea behind that style of gradual reveal is probably some kind of storytelling, but the only story it tells is that mozilla is wasting donations on people with incorrect opinions that could be used on.. idk not building torment nexii?
It’s literally a paradox. It cannot be changed. Open Source AI will ultimately win. Think it through.
Seems quite odd to use OpenRouter as “proof” that open weights models won. If you’re using OpenRouter, you’re already looking to bypass frontier models. To suggest there’s no longer a tradeoff simply isn’t true. But this isn’t the first time I thought Mozilla was a less than trustworthy source of information.
I think "If you’re using OpenRouter, you’re already looking to bypass frontier models." is false. Our company uses both Claude subscriptions and OpenRouter ... and a lot of what we use OpenRouter for is more Claude.
We do a little exploration with other models through it, but it's not at all accurate to say we use it because we are "already looking to bypass frontier models".
... or at least, no more than any other company that doesn't want to overpay for their tooling, but is basically happy (ATM) with the current state of Claude.
The UI is really hard on the eyes. Personally, I think the font size is way too big, and the animation timing feels off. If this is a benchmark page and not a product page, I feel like the information should be scannable at a glance. The UX is bad.
I'm unsure what it is about AI developers seemingly not having eyeballs. The Hermes Agent website is absolutely eye-searing and the application itself resembles some sort of weird "RETVRN" greek-styled travel agent website.
https://hermes-agent.nousresearch.com/
Really wish websites weren't allowed to force smooth scroll on. Hijacking basic browser functionality is so hostile.
Oh that's easy: they outsource design to the LLM, which doesn't have eyeballs.
I agree 1000%, Mr. Jake.
I use hermes only ever saw their repo. Atrocious. I was sure you were exaggerating.
Feels like a mobile website that was never optimized for desktop usage.
There isn't any open-source AI. There is Open AI (not to be confused with the closed company called OpenAI, which was unable to trademark its name). There's no open source AI both because the open source community doesn't have the resources to train a useful AI and because AI doesn't have source code.
I'm not sure what kind of point you're trying to make. There are projects to train competent modern LLMs in which the entire pipeline (data, training process, final weights) is all completely transparent, shared, and reproducible by anyone with the compute to try it out.
Or is your definition of "open source" mean that a small indie dev should be able to reproduce the entire pipeline? Because that would disqualify more than just LLMs, but also hardware platforms like Arduino where you need to pay for manufacturing to get the underlying stuff built... is Arduino "open source?"
Is training code and dataset not source?
Are they open?
I’m just challenging your point that ai doesn’t have source.
This is just wrong on multiple levels, the open source model ecosystem is very much a thing.
I don't know of any useful model that would match the usual definition of "open source". That is, where everything needed to build the thing is provided, such as the training data and code. Not that it would be tremendously useful anyways considering how expensive training a model is.
I much prefer the "open weights" term. It is not open source in the sense that you only get the finished product, not the actual source, but it is still open in the sense that it is not only accessible as a service.
For an analogy, take Quake for instance. When it was launched, its game server was available as an executable, so you could run it your machine, but that didn't make it open source. Only much later it was released as true open source software.
Olmo 3? K2 V2? There are definitely LLMs with very compelling capabilities where the dataset, training process, and final weights are all open. There are also initiatives in the EU and various national government levels (e.g. Switzerland) to develop LLMs in an open and transparent way, not just "open weights."
There are fully open models, it's just not as common because that's basically giving away the sauce, which is non-viable for many.
eg: https://allenai.org
mozilla only began babbling about open source ai when they got their teeth extracted for free via the entire community giving them the curbiest of curbstomps.
i still use firefox but hot damn did they utterly fail to read the room initially. the only other company i can think of in recent memory -- besides sony cutting out discs -- is logitech when their ceo began gibbering about a subscription mouse or microsoft and its copilot button(s).
It's frustrating because I truly believe they are well-meaning, honest actors. Every single "controversy" surrounding Mozilla can be directly attributed to miscommunication rather than malice. They do often struggle to read the room.
That said, the takeaways from this report are exciting, and I do feel that Mozilla now has the right lens in their assessment of OSS AI and their own approach to ensuring interoperability and setting open, modern standards.
Just like how the web was won?
I think Mozilla is chasing a past formula, but the projection isn't linear enough to remain consistent, and the critical parts of the outcome, utter centralization of the market dominance of the three C's, are left out of the equation.
We might get the consolation prize, a few nerds having competitive alternatives to applaud, but we will be left with the hidden costs: stagnation by bloated market leaders, consumers and businesses pouring trillions of dollars into the commercial offerings while open development wonders where money comes from, and the leakage of these imbalances into political and social spheres.
If we follow a Mozilla template and get to the peak of Mozilla's success at the web, look at what that really is. Facebook, Amazon, Google etc are orthogonal to that equation.
"Open won"... to be fair cause "google paid it".
I think the fact that Mozilla survival model ultimately depends on Google's money means Google is keeping a corpse propped up just to have a defense argument that a browser competitor still exists, so they don't get hit with monopoly regulations.
Particularly in this day in age when the FF market share is down to low single digits
Title: The state of open source AI.
First sentence: In New Zealand's far north, a Māori broadcaster...
...oh boy, that's all you need to read to know what kind of media diet the writer is on.
Normally I'd agree, but the rest of the sentence does actually discuss open models and their use cases.
Mozilla presents themselves as an "open community". But they are a commercial corporation bankrolled by Google, and with an oppressive organizational culture. I would interpret this post as being not just about an opportunity for their self-aggrandizement, but also Mozilla trying to whitewash its mass surveillance of users, likely used also to train some model or another, or perhaps just to feed their patronn (Alphabet)'s ad machine.
They tell us about how the farmers and native people and whateve are all happy with their chatbots and models. The major effects are a massive and ever-increasing energy use - in a time where we must cut back and economize to avoid further global warming; a massive diversion of investment capital - especially in the US; fantastic stock valuations for a few tech giants (gee, I wonder whether any of them is related to Mozilla somehow); and other effects one could survey, all more significant by far than the examples they bring.
Haven't been following the articles and snippets we get from these labs about training their models for a while. But I'm guessing the latest chinese models are way less based on distilling? If not, then your speed of progress is still limited by the two labs (which we are collectively, in various forms subsidizing).
This is really insane to me.
There's nothing practical about open-source models yet that makes them even remotely comparable to closed frontier models.
All the hype around GLM, Qwen, now Kimi.... Are people really this naive that they believe these reports or, more worringly, are people NOT using these models and seeing the HUGE gap that still exists?
Take a task, any medium-sized task, decently scoped that you'd trust to give to Sonnet to finish without a hitch. Now give it to ANY open-source frontier model and watch them struggle and go in circles while failing tool calls and randomly assuming things.
Open-source is and has been amazing but its so hard to deploy reliably and at scale and there's still big problems in the underlying models with instruction following and tool calling that makes it basically unusable for production workloads at a decent price point...
> Take a task, any medium-sized task, decently scoped that you'd trust to give to Sonnet to finish without a hitch. Now give it to ANY open-source frontier model and watch them struggle and go in circles while failing tool calls and randomly assuming things.
Claude used to be much worse than it is now, just as bad the open weights models are. And the open weights were worse. The labs will also try to keep the lead, but at some point people start seeing real value from open models. Maybe you say they're not ready yet for medium tasks, but everyone sees the writing on the wall.
I hope you're right and I want you to be right, but, even seeing the current hype around local models, etc... and open-source models, I think the industry is currently under a big confusion where they see the benchmarks of things like Kimi, GLM, Qwen, they play with it via opencode, and they think like: "Wow this is pretty good, I want to deploy this". But they don't understand how the KV cache grows over time and can take almost as much memory as needed for a 30B param model, they dont understand that a quantized model WILL NOT be the same as a full precision one, and they surely don't see the engineering work needed to serve inference to even tens of customers at a decent quality and latency level.
The biggest moat of these giant labs and models is increasingly shifting towards deployment capabilities and (debatably) having better (proprietary) harnesses.
The models themselves can be impressive on benchmarks, but unless they can be served reliably to customers either at scale, hosted somewhere, or even on edge with predictable latency and memory usage, then frontier will always be leading.
It sounds like you're focusing on the problems of running local models, or running models yourself, but I don't think that many people seriously expect near term improvement on that, it's definitely more just hopeful thinking there. That's not what I meant to address, and I also am in more of a "wait and see" mode.
But at this point we do expect that open weights _hosted_ options become feasible for the tasks they're using the frontier models for. And because of the lack of "legal monopoly" (intellectual property of whatever kind), they're way cheaper, not mention more flexible.
The launch of the tinker platform from Thinking Machines is an example of the "more flexibility" part that people want (and they chose to make their model open weights, maybe because this is the angle they want to push).
At this point I think it's realistic enough that the ball is in OpenAI / Anthropic's court to figure out how to respond to this threat to their business model.
That said, I think it's concerning that there are apparently only a couple of providers of hosted open weights inference, due to the complexities of doing so (per Dax from OpenCode's tweets).
> but everyone sees the writing on the wall
i'm... not sure? This assumes ~stagnation in task-possibility. We've had ~exponential progress for like 3+ years now; I'd have never dreamed the tooling I hammer daily would exist in my lifetime just.. 3? years ago. And it's improving daily.
Maybe Open will win, maybe Closed will keep pushing the envelope. The world here is raw enough i don't think anyone can make any significant claim other than 'holy shit this is useful and moving Fast'.
If your doing things the closed models won't let you do; its the whole ball game.
Have you really given GLM 5.2 an honest go ?