Is this in response to now Mythos/Fable and OpenAI starting to compete more directly with Palantir for defense business, and it is timed with the day Fable is re-released?
The Enterprise versions of Anthropic/OpenAI's products say they don't train on input data so either this is a misinformed rant or it's an allegation that the AI labs are perpetrating an enormous fraud on their enterprise customers.
It is like questioning the purpose of shovels, because surely if shovels generate so much wealth for the miners, wouldn‘t shovel companies ask for a percentage of the gold (say 30%) instead of selling shovels at a unit price?
But saying just because openai didn't adopt a particular business model therefore tokens must be worthless seems like a stretch. Maybe it was regulatory. Maybe they just don't want to be a conglomerate of that sort. Could be any number of reasons.
Everything he says about AI companies charging a subscription and then possibly extracting value from customer data is true for every single cloud hosted SaaS since the dawn of that model. If you put your data on someone else’s computer, you have only their word they won’t misuse or leak it. Often, if you read the ToS, you don’t even have that.
Google Drive and Gmail anyone? They’ve had the Crown Jewels of a ton of businesses since forever. Even governments. Can you imagine what they have? Just Gmail alone. It’s staggering.
Dropbox? Microsoft OneDrive? Mind blowing.
Palantir? Pretty rich for a guy who runs a company named after an insecure communication channel used to steal data and run side channel attacks to talk about the risks of losing data sovereignty.
Obviously nobody would ever… say… insider trade with that data. That would be unethical and illegal. Has that ever happened? Who knows. Palantir, Google, and every other company I listed are definitely using that data for AI training. I'd be shocked if they were not, ToS or enterprise agreements be damned. The AI race is so hot and hype-loaded that they'd run the risk of getting sued if they thought the data would give them an edge. They'd probably just settle the lawsuit and call it COGS.
Not your storage, not your data (unless it's encrypted with keys only you possess).
As for why they don’t take equity positions when the products save time or boost productivity… lol wut?
Yeah. My Mac boosts my productivity. It requires less maintenance than any other computer I can buy. So Apple gets equity in my company, right?
Steam engine companies got equity in mining ventures when they sold them automated water pumps. Railroads got equity in food companies for shipping fast and reducing spoilage. The inventors of refrigeration too. Can you imagine how much agro company equity they must have gotten!
AFAIK a bunch of customers run Palantir on prem specifically so that their data doesn't leak and they can use a firewall instead of a TOS. Some of the data is so sensitive that Palantir doesn't want to touch it. I don't think it's accurate to lump Palantir in with all the other SaaS.
I mean, if companies really are furious, why tf are they paying for the service? don0t pay, don't use the service, host yourself some open weight model, or train one yourself?
Are these companies being pressured by some higher entity to subscribe to these services?
I agree with the "IP" stealing part, because yes that's what they did, but the point on companies being livid with Anthropic/OpenAI makes no sense.
I wouldn't be surprised if a bunch of organizations made a temporary exception to try Opus and then got "addicted" in the sense that it's too good to give up.
Yeah, but that has literally been their business model from the start yes?
Even now, after all the development, if you can somehow get Claude Fable to run without bricking due to the safety 'features', whatever it imitates is still a poor replacement compared to the slew of available open source code repositories that it stole its training from.
Problem is humans are lazy and its really easy to just yell at a chatbot until you get some slop that mostly does it.
Wait till you connect Slack with LLM's - all your chat and internal data and meta data that includes confidential information will be given away to companies who most likely already have access to all your engineering codebase.
If you value your data, then don't upload confidential documents to LLM services. Don't connect confidential systems to LLMs via connectors. Don't enter confidential data in LLMs. If you do any of these things, then fully expect the documents to be mined for further training of the LLM. That's unless you're running the Gov partition of the LLM which contractually doesn't do it.
There was a submission earlier but it didn't get any upvotes https://news.ycombinator.com/item?id=48755074. I wouldn't mind if this thread is merged into that one... as long as the discussion stays in the front page. Many times conversations got suppressed due to "dupe" merged into older submissions!
I really don't like Karp but I think he is spot on this point.
I hear a thread of envy throughout that is basically "why isn't my magic money machine as hyped as their magic money machine"
Is this in response to now Mythos/Fable and OpenAI starting to compete more directly with Palantir for defense business, and it is timed with the day Fable is re-released?
The Enterprise versions of Anthropic/OpenAI's products say they don't train on input data so either this is a misinformed rant or it's an allegation that the AI labs are perpetrating an enormous fraud on their enterprise customers.
They have ignored any copyright during initial creation of these models, why would they start respecting copyright now?
If it's profitable, they would. Companies aren't moral, with very rare exceptions. Legal risks and costs are just part of the overall equation.
But that's not in defense of what Karp said, or that slop tweet summarising it. That's... wild. I regret reading it.
Seems like a misread to me.
Tokens don't have inherent value. You can't hoard them and stuff them under a pillow. They're valuable when applied to a problem.
It's like asking a shovel making company why they need gold miners.
It is like questioning the purpose of shovels, because surely if shovels generate so much wealth for the miners, wouldn‘t shovel companies ask for a percentage of the gold (say 30%) instead of selling shovels at a unit price?
It's not an either or thing. Value chains have an equilibrium like anything else - naturally balances out.
You don't see every step being abandoned because everyone converged on the most profitable link in the real world.
>wouldn‘t shovel companies ask for a percentage of the gold (say 30%)
They could go that route. Nvidia is doing that right now
https://www.cnbc.com/2026/07/02/nvidia-plans-to-offer-start-...
But saying just because openai didn't adopt a particular business model therefore tokens must be worthless seems like a stretch. Maybe it was regulatory. Maybe they just don't want to be a conglomerate of that sort. Could be any number of reasons.
If you invented god that can make you billions, why wouldn't you extract that value yourself instead of leasing it in form of an API access.
No one claims their showels find gold for you.
Oh boy. This edgelord. Where to start.
Everything he says about AI companies charging a subscription and then possibly extracting value from customer data is true for every single cloud hosted SaaS since the dawn of that model. If you put your data on someone else’s computer, you have only their word they won’t misuse or leak it. Often, if you read the ToS, you don’t even have that.
Google Drive and Gmail anyone? They’ve had the Crown Jewels of a ton of businesses since forever. Even governments. Can you imagine what they have? Just Gmail alone. It’s staggering.
Dropbox? Microsoft OneDrive? Mind blowing.
Palantir? Pretty rich for a guy who runs a company named after an insecure communication channel used to steal data and run side channel attacks to talk about the risks of losing data sovereignty.
Obviously nobody would ever… say… insider trade with that data. That would be unethical and illegal. Has that ever happened? Who knows. Palantir, Google, and every other company I listed are definitely using that data for AI training. I'd be shocked if they were not, ToS or enterprise agreements be damned. The AI race is so hot and hype-loaded that they'd run the risk of getting sued if they thought the data would give them an edge. They'd probably just settle the lawsuit and call it COGS.
Not your storage, not your data (unless it's encrypted with keys only you possess).
As for why they don’t take equity positions when the products save time or boost productivity… lol wut?
Yeah. My Mac boosts my productivity. It requires less maintenance than any other computer I can buy. So Apple gets equity in my company, right?
Steam engine companies got equity in mining ventures when they sold them automated water pumps. Railroads got equity in food companies for shipping fast and reducing spoilage. The inventors of refrigeration too. Can you imagine how much agro company equity they must have gotten!
Just wat.
AFAIK a bunch of customers run Palantir on prem specifically so that their data doesn't leak and they can use a firewall instead of a TOS. Some of the data is so sensitive that Palantir doesn't want to touch it. I don't think it's accurate to lump Palantir in with all the other SaaS.
I mean, if companies really are furious, why tf are they paying for the service? don0t pay, don't use the service, host yourself some open weight model, or train one yourself? Are these companies being pressured by some higher entity to subscribe to these services?
I agree with the "IP" stealing part, because yes that's what they did, but the point on companies being livid with Anthropic/OpenAI makes no sense.
I wouldn't be surprised if a bunch of organizations made a temporary exception to try Opus and then got "addicted" in the sense that it's too good to give up.
Yeah, but that has literally been their business model from the start yes?
Even now, after all the development, if you can somehow get Claude Fable to run without bricking due to the safety 'features', whatever it imitates is still a poor replacement compared to the slew of available open source code repositories that it stole its training from.
Problem is humans are lazy and its really easy to just yell at a chatbot until you get some slop that mostly does it.
Wait till you connect Slack with LLM's - all your chat and internal data and meta data that includes confidential information will be given away to companies who most likely already have access to all your engineering codebase.
If you value your data, then don't upload confidential documents to LLM services. Don't connect confidential systems to LLMs via connectors. Don't enter confidential data in LLMs. If you do any of these things, then fully expect the documents to be mined for further training of the LLM. That's unless you're running the Gov partition of the LLM which contractually doesn't do it.
Your worst enemy has just made a great point.
The tweet itself is AI-generated slop, flagging. If there's a discussion to be had, a primary source seems like a better link.
Agreed. But please don't flag for the link as we'll lose the discussion. I don't care much about the tweet. It was the one I saw circulating.
Mods can change the link to the full interview https://www.youtube.com/watch?v=0A3sGymV6kY
There was a submission earlier but it didn't get any upvotes https://news.ycombinator.com/item?id=48755074. I wouldn't mind if this thread is merged into that one... as long as the discussion stays in the front page. Many times conversations got suppressed due to "dupe" merged into older submissions!
Edit: aaaand... it's gone. Congrats.
what is this slop of a tweet
[dead]
[dead]