These videos are worth a watch. There are tons of impressive moments, but they had me at the very first one where a woman says: "I'm going to tell you a story," and then pauses for a long, luxurious sip from a cup of coffee, and the model ... does nothing, just waits. Take my money.
Speaking of taking my money, what's the economic model for a company like this? They've published a fair amount about their architecture - enough that I imagine frontier labs could implement. Patents? Trade secrets? It's hard for me to understand how you'd be able to beat that training compute and knowhow at Anthropic/GOOG/oAI/Meta without some sort of legal protection.
I can't wait to see what these model architectures do with like 30-40% lower latency and more model intelligence. Very appealing. For reference, these look to be roughly 1/10 the size of Opus 4.7 / GPT 5.x series -- 275B, 12B active. So there's lots of room to add intelligence, and lots of hope that we could see lower latency.
> They've published a fair amount about their architecture - enough that I imagine frontier labs could implement.
i think the real ones know this is the tip of the iceberg? hparam tuning, data recipes, data collection, custom kernels, rl/eval infra, all immensely deep topics that would condense multiple decades of phd lifetimes to produce SOTA performance (in both senses of the word) like this.
i would also calibrate what you are impressed by. simply waiting is a posttrain thing - the fact that gemini and oai have not prioritized it is not something you should overindex on as hard. what they showed with full duplex is technically far far harder to achieve
I agree that full duplex is the amazing bit. For instance, the three engineers shouting trivia questions while a timer is running — that’s extremely novel as far as I can tell.
I’d like to believe from the demos that this ability to wait kind of falls out of the model as an emergent property — perhaps coming out of a small RL loop - rather than a specific behavior trained, a-la a VAD component in a stack. Either way, I would guess that VAD absolutely cannot do this right now — interruptions are highly annoying on all voice interaction experiences, and if it were a simple matter of better post training, SOMEONE would have done this, e.g. elevenlabs.
But, I disagree on your idea that this is too expensive/too hard to replicate. For me, yes. But, there’s an existence proof — a small team at a new company just did this without a real roadmap, certainly for less than $1b dollars and probably in less than two years. They are almost certainly less skilled at your list of needs to replicate than teams at the frontier labs, who have been given a roadmap.. So I don’t think it’s as difficult as you propose, from an organizational skills perspective.
In China it's become well known that promising new companies will get an offer from either Alibaba or Tencent. In the US, it's probably simmilar. Everything that's out in the open can get acquired or simply copied. Maybe that is what Thinking Machines is hoping as well?
Publish a Demo -> acquihire for anthropic/oAI/GOOG/META stock and cash is an understandable economic model. In this case, I feel like they built more than would be needed though — and I hope they deploy something useful, I’d love to play with it.
> Companies can’t afford to just give things away right?
Let's say a cutting-edge young researcher is making a name for themselves in their field and earning $300k/yr at a company where they're encouraged to publish and speak. You're trying to headhunt them for a company where they'll be forbidden from sharing their work which will likely stall their career and reputation outside of that company. How much do you think you'd have to offer? $600k? $1M? $1.5M?
When faced with the choice to paying significant salaries, hiring lower-tier researchers, or just letting their people publish, many companies conclude that giving away some of their work is the best option. (And that doesn't even include the benefits of boosting the company's profile which makes it easier to attract other cutting-edge researchers.)
Yes they can. Your research papers are not the whole story. It’s like google could open source their entire monorepo and very little would change. No one else could operate it.
The noteworthy things to me are that the architecture is a transformer that takes in text, image, and audio input and produces text and audio output, all trained together, and it works in near real-time through interleaving inputs and outputs rather than pure generation of the output from a given prompt.
> Time-Aligned Micro-Turns. The interaction model works with micro-turns continuously interleaving the processing of 200ms worth of input and generation of 200ms worth of output. Rather than consuming a complete user-turn and generating a complete response, both input and output tokens are treated as streams. Working with 200ms chunks of these streams enables near real-time concurrency of multiple input and output modalities.
That's probably the main thing that distinguishes it from the multimodal models from other frontier labs as far as I can tell.
What's really interesting for me about multimodal architectures from the ground up is that we might start to see applications where different modalities are "facets" of the same thing. Like a coding agent that sees "code" + "IDE" + "memory mapping" + feedback from different plugins as different modalities. And it gets to output in them as well - text where it needs to, actions (not <action>call_something(params)</action> like we have today) and so on. Being able to "sit still" until one of the modalities triggers is really interesting.
We can do these things today, but they're "bolted on" as afterthoughts. Yet they work remarkably well. I wonder how well they'd work if trained int his combined regime, from the ground up.
I hate to say it but while this does seem very impressive and a step forward in how we interact with AI, the use-cases they present and the UX both seem unrealistic and/or unhelpful.
With the exception of the real-time translation (which seems like it should be a separate product all by itself), none of the use-cases they presented had much utility. I don't want anything to count the number animals in my stories or time a trivia quiz for me. The auto-slouch-detector, while the demo was pretty funny, just seems so dystopian and weird. AI interrupting you to scold you about taking elderly parents mountain biking instead of waiting for you to finish to scold you? No thanks.
The UX is also an issue - the model interrupting the user (even when apparently required by these strange use-cases) is jarring and makes one lose their flow. You can even see this in the demo videos that they put out - the employees/actors had to really concentrate to continue speaking as if they weren't being interrupted by a brash robotic machine. A human, when participating in this (rare) "invited interruption" has the ability to speak "under" the main speaker and I feel it's generally timed with a lot of nuance.
Even in the auto-translation demo, they ducked the human's audio but the AI steamrolled him and it would have been impossible to actually do that demo without either an incredible amount of control over one's speaking, or (more likely) muting the output. A human translator has a way of "pointing" the "output" to the intended speaker.
The very best part of this tech was presented in the first video where it shows the AI not needlessly interrupting the user. This seems to me more of an important bug fixed that the current models still (somehow) have.
Maybe a good use-case for this would be counting "um's" and the like while practising public speaking.
Agree that this is interesting/impressive, and the demos are nice.
But I completely cracked up at the unexpected physical comedy of the woman in the "slouching" demo, haha omg that was comedy gold, no notes...
I do appreciate less of that flavor of demo that we get from OpenAI/Anthropic, and more of this "human"-feeling vibe. Dare I go as far as calling this an example of "human-centered design" even (https://en.wikipedia.org/wiki/Human-centered_design)?
Yes! This is a big thing ive noticed in all AI demos. If the best use case you can think of to show off yor tech is to book a holiday, that I could easily do myself, does your service really add much value? Or is it simply because the real uses will be nuanced and specialsed, and not suited for a quick general audience demo? I'm not sure.
In theory I would expect it to do everything the current frontier models are capable of but with the added benefit of real time interactivity for better collaboration. The biggest benefit may be the real time video input so it can take in that input in parallel with producing outputs steered by the input rather than taking in a video or all images at once and then producing a single output for all of that.
Presumably it will be possible to adjust that behavior with settings, the system prompt, etc. Not that most users will make such adjustments, though.
I'm currently teaching a class on AI-related issues at a university in Tokyo. Many of the students were surprised when I showed them that they can change the response behavior of chatbots to make them more or less verbose, sycophantic, etc. It shifted the direction of our discussions on the possible impacts of AI on the people who use it.
hard agree, there's already "voice ai" companies that use the normal models and have this "interaction" engine on top of them to produce better results than I've seen in these demos. idk why people are impressed
What I will say is that this is probably the first model after gemini live to do some of these things. It feels similar to gemini live, which I don't think is what they were going for exactly, but IMO it is still impressive as I don't think anyone else has matched full duplex video/audio/tool calling.
Next gemini releases coming next week though, we will see how that matches up!
This deserves to be at the top of HN, shame it seems like it's not going to make it. Some of the demos are hilarious. Clearly having the model appropriately choose when to speak is a major thing that has been missing from voice models to date. It seems like the latency is still a touch too high to be truly human-like though.
These videos are worth a watch. There are tons of impressive moments, but they had me at the very first one where a woman says: "I'm going to tell you a story," and then pauses for a long, luxurious sip from a cup of coffee, and the model ... does nothing, just waits. Take my money.
Speaking of taking my money, what's the economic model for a company like this? They've published a fair amount about their architecture - enough that I imagine frontier labs could implement. Patents? Trade secrets? It's hard for me to understand how you'd be able to beat that training compute and knowhow at Anthropic/GOOG/oAI/Meta without some sort of legal protection.
I can't wait to see what these model architectures do with like 30-40% lower latency and more model intelligence. Very appealing. For reference, these look to be roughly 1/10 the size of Opus 4.7 / GPT 5.x series -- 275B, 12B active. So there's lots of room to add intelligence, and lots of hope that we could see lower latency.
> They've published a fair amount about their architecture - enough that I imagine frontier labs could implement.
i think the real ones know this is the tip of the iceberg? hparam tuning, data recipes, data collection, custom kernels, rl/eval infra, all immensely deep topics that would condense multiple decades of phd lifetimes to produce SOTA performance (in both senses of the word) like this.
i would also calibrate what you are impressed by. simply waiting is a posttrain thing - the fact that gemini and oai have not prioritized it is not something you should overindex on as hard. what they showed with full duplex is technically far far harder to achieve
I agree that full duplex is the amazing bit. For instance, the three engineers shouting trivia questions while a timer is running — that’s extremely novel as far as I can tell.
I’d like to believe from the demos that this ability to wait kind of falls out of the model as an emergent property — perhaps coming out of a small RL loop - rather than a specific behavior trained, a-la a VAD component in a stack. Either way, I would guess that VAD absolutely cannot do this right now — interruptions are highly annoying on all voice interaction experiences, and if it were a simple matter of better post training, SOMEONE would have done this, e.g. elevenlabs.
But, I disagree on your idea that this is too expensive/too hard to replicate. For me, yes. But, there’s an existence proof — a small team at a new company just did this without a real roadmap, certainly for less than $1b dollars and probably in less than two years. They are almost certainly less skilled at your list of needs to replicate than teams at the frontier labs, who have been given a roadmap.. So I don’t think it’s as difficult as you propose, from an organizational skills perspective.
hasn't the economic model always been enterprise llms?
tinker - for fine tuning a custom enterprise model,
interaction models - for working as a digital paired employee (as opposed to a company having to reinvent their entire process around ai agents)
In China it's become well known that promising new companies will get an offer from either Alibaba or Tencent. In the US, it's probably simmilar. Everything that's out in the open can get acquired or simply copied. Maybe that is what Thinking Machines is hoping as well?
Publish a Demo -> acquihire for anthropic/oAI/GOOG/META stock and cash is an understandable economic model. In this case, I feel like they built more than would be needed though — and I hope they deploy something useful, I’d love to play with it.
Purely out of curiousity, I see you are using an em dash. Did you use voice transcription or something? It looks hand-typed though. I'm confused.
I just typed two single hyphens from my iOS device. One: - two: —
Edit: when I edit this comment they have been merged in the form so I speculate this is an iOS keyboard feature.
they hire leading researchers, and leading researchers won't work for you unless they're able to publish
That was true 10 years ago. It’s most definitely not true now. The arms race is very real.
> leading researchers won't work for you unless they're able to publish
oh, honey.
Do we want the whole humanity to get richer, or few individuals (company owners)?
Which seems bizarre. Companies can’t afford to just give things away right?
> Companies can’t afford to just give things away right?
Let's say a cutting-edge young researcher is making a name for themselves in their field and earning $300k/yr at a company where they're encouraged to publish and speak. You're trying to headhunt them for a company where they'll be forbidden from sharing their work which will likely stall their career and reputation outside of that company. How much do you think you'd have to offer? $600k? $1M? $1.5M?
When faced with the choice to paying significant salaries, hiring lower-tier researchers, or just letting their people publish, many companies conclude that giving away some of their work is the best option. (And that doesn't even include the benefits of boosting the company's profile which makes it easier to attract other cutting-edge researchers.)
Yes they can. Your research papers are not the whole story. It’s like google could open source their entire monorepo and very little would change. No one else could operate it.
The noteworthy things to me are that the architecture is a transformer that takes in text, image, and audio input and produces text and audio output, all trained together, and it works in near real-time through interleaving inputs and outputs rather than pure generation of the output from a given prompt.
> Time-Aligned Micro-Turns. The interaction model works with micro-turns continuously interleaving the processing of 200ms worth of input and generation of 200ms worth of output. Rather than consuming a complete user-turn and generating a complete response, both input and output tokens are treated as streams. Working with 200ms chunks of these streams enables near real-time concurrency of multiple input and output modalities.
That's probably the main thing that distinguishes it from the multimodal models from other frontier labs as far as I can tell.
What's really interesting for me about multimodal architectures from the ground up is that we might start to see applications where different modalities are "facets" of the same thing. Like a coding agent that sees "code" + "IDE" + "memory mapping" + feedback from different plugins as different modalities. And it gets to output in them as well - text where it needs to, actions (not <action>call_something(params)</action> like we have today) and so on. Being able to "sit still" until one of the modalities triggers is really interesting.
We can do these things today, but they're "bolted on" as afterthoughts. Yet they work remarkably well. I wonder how well they'd work if trained int his combined regime, from the ground up.
> interleaving the processing of 200ms worth of input and generation of 200ms worth of output.
How does this work? Don't LLMs/transformers need whole context to output next chunk of tokens?
I hate to say it but while this does seem very impressive and a step forward in how we interact with AI, the use-cases they present and the UX both seem unrealistic and/or unhelpful.
With the exception of the real-time translation (which seems like it should be a separate product all by itself), none of the use-cases they presented had much utility. I don't want anything to count the number animals in my stories or time a trivia quiz for me. The auto-slouch-detector, while the demo was pretty funny, just seems so dystopian and weird. AI interrupting you to scold you about taking elderly parents mountain biking instead of waiting for you to finish to scold you? No thanks.
The UX is also an issue - the model interrupting the user (even when apparently required by these strange use-cases) is jarring and makes one lose their flow. You can even see this in the demo videos that they put out - the employees/actors had to really concentrate to continue speaking as if they weren't being interrupted by a brash robotic machine. A human, when participating in this (rare) "invited interruption" has the ability to speak "under" the main speaker and I feel it's generally timed with a lot of nuance.
Even in the auto-translation demo, they ducked the human's audio but the AI steamrolled him and it would have been impossible to actually do that demo without either an incredible amount of control over one's speaking, or (more likely) muting the output. A human translator has a way of "pointing" the "output" to the intended speaker.
The very best part of this tech was presented in the first video where it shows the AI not needlessly interrupting the user. This seems to me more of an important bug fixed that the current models still (somehow) have.
Maybe a good use-case for this would be counting "um's" and the like while practising public speaking.
Aside from how impressive the model is, the demos here are very well done! Quirky and short, unlike what we're used to from Anthropic and OpenAI.
Agree that this is interesting/impressive, and the demos are nice.
But I completely cracked up at the unexpected physical comedy of the woman in the "slouching" demo, haha omg that was comedy gold, no notes...
I do appreciate less of that flavor of demo that we get from OpenAI/Anthropic, and more of this "human"-feeling vibe. Dare I go as far as calling this an example of "human-centered design" even (https://en.wikipedia.org/wiki/Human-centered_design)?
Very cool! The demos felt fairly contrived - e.g., count things while I talk. I wonder what more useful or commercial applications look like.
Yes! This is a big thing ive noticed in all AI demos. If the best use case you can think of to show off yor tech is to book a holiday, that I could easily do myself, does your service really add much value? Or is it simply because the real uses will be nuanced and specialsed, and not suited for a quick general audience demo? I'm not sure.
In theory I would expect it to do everything the current frontier models are capable of but with the added benefit of real time interactivity for better collaboration. The biggest benefit may be the real time video input so it can take in that input in parallel with producing outputs steered by the input rather than taking in a video or all images at once and then producing a single output for all of that.
Their corporate sound system is sick!
the intentions may be good but it looks like a boost to surveillance tech in the wrong hands, time to react
This does feel like where things should be going for more natural human-AI interaction patterns. Nice write up and demos.
This looks similar to things people are already building locally with Gemma4 and TTS; just a bit fancier.
Local models will catch up soon.
Very cool demo, I wonder what would be the billion dollar applications of a thing like this.
Very cool tech. I think people are underrating how this will be used.
That's neat and definitely the next step. But to be honest, I don't want an AI talk to me like that.
Same here.
Presumably it will be possible to adjust that behavior with settings, the system prompt, etc. Not that most users will make such adjustments, though.
I'm currently teaching a class on AI-related issues at a university in Tokyo. Many of the students were surprised when I showed them that they can change the response behavior of chatbots to make them more or less verbose, sycophantic, etc. It shifted the direction of our discussions on the possible impacts of AI on the people who use it.
Simultaneous speech is best.
Really really cool. If they can serve this efficiently it would disrupt a lot of things.
incredibly impressive demos. I wonder how the training data for these models look like?
is it separate batches of special "skills" that are added post training? how can they guarantee the models won't eventually lose a skill?
am i the only person not impressed by this ? it just feels akward still with pauses and doesnt openai offer voice cadence already
hard agree, there's already "voice ai" companies that use the normal models and have this "interaction" engine on top of them to produce better results than I've seen in these demos. idk why people are impressed
Same here. I dont see anything there that nobody else can catch up on eventually. I must be missing something here. It's all cute, but mmm
What I will say is that this is probably the first model after gemini live to do some of these things. It feels similar to gemini live, which I don't think is what they were going for exactly, but IMO it is still impressive as I don't think anyone else has matched full duplex video/audio/tool calling.
Next gemini releases coming next week though, we will see how that matches up!
A bunch of companies made light bulbs after Edison, that doesn't mean that light bulbs weren't an interesting invention.
This deserves to be at the top of HN, shame it seems like it's not going to make it. Some of the demos are hilarious. Clearly having the model appropriately choose when to speak is a major thing that has been missing from voice models to date. It seems like the latency is still a touch too high to be truly human-like though.