正在加载视频...

视频加载失败

加载此视频时出现问题。这可能是由于临时网络问题，或视频可能不可用。

We achieved a breakthrough at BTL. AI models are very linear. They think in one direction at a time, and when they're wrong they scramble and start again. For something meant to replace humans, that's way too human. We changed that. We made LLM reasoning and execution work more... like a quantum computer: many paths at once, the ones that meet merge into one, the dead ends cancel out, and everything left moves forward together. On the same hard problems, a 1.7B model thinking the normal way solved 3 out of 30. Interference Search(our new architecture)solved 23. We watched the normal model find the right answer at token 1,313, check it 9 more times, wander off and run out of budget without ever answering. Ours got there in 3 steps. Parallel thinking and execution. Subagents were a terrible way to tackle this. More technical details soon.show more

Bad theory labs

4,323 subscribers

173,138 次观看 • 19 天前 •via X (Twitter)

Anya Rossi• Live Now

Private livecam show

44 条评论

Bad theory labs 的头像

Bad theory labs19 天前

The speed difference is insane. Our new architecture solves each problem in 5ms vs 21s for the actual model (qwen)

Bad theory labs 的头像

Bad theory labs19 天前

Repo: Paper:

Bad theory labs 的头像

Bad theory labs19 天前

We are raising. We’re appreciate any warm intros or anything to sponsor our work

CJ 的头像

CJ19 天前

Sounds beautiful and amazing. Is this a solution that applies to a specific model, a model type (eg dense vs MoE), the inference engine or the model encoding - or many among these? Please make it Open Source and Free!!! We're caught captive by the consumer GPU mafia :_(

Bad theory labs 的头像

Bad theory labs19 天前

Done

Grounded Idealist 的头像

Grounded Idealist19 天前

Can u add mtp on top of it for even speed? I guess this is that qwen 3.5 right ? It will be even faster I guess Also will this be open source? Anyway happy to see such interesting work!’congrats!

Bad theory labs 的头像

Bad theory labs19 天前

Just tried this and it worked perfectly

Grounded Idealist 的头像

Grounded Idealist19 天前

Ok..when open source,if ever!

Bad theory labs 的头像

Bad theory labs19 天前

Open sourced it. Check the replies

Grounded Idealist 的头像

Grounded Idealist19 天前

Thank u!

Swapnil Kale 的头像

Swapnil Kale19 天前

But the inference search architecture would still require a lot of token usage to run the multiple different paths, right?

Bad theory labs 的头像

Bad theory labs19 天前

No. It’s fast and costs less tokens than the normal linear way llms function.

Alban Cazin 的头像

Alban Cazin19 天前

@SKaleworks hey, so what are basically the flaws of this new architecture?

Bad theory labs 的头像

Bad theory labs19 天前

@SKaleworks I’m going to be honest here. We haven’t even explored it a lot. It’s so new that I’m sure there would be flaws, we just haven’t seen one yet. All these experiments was ran on a Mac without gpu so yeah. Open sourcing it to see what the community does with it and more improvements

Tom Nicholson 的头像

Tom Nicholson19 天前

Tree of Thought right? Old news

fawn 的头像

fawn19 天前

Beautiful work

inxanity! 的头像

inxanity!19 天前

@grok how is this done at transformers level

Shehab Albakr 的头像

Shehab Albakr19 天前

@Pashmori look bro that's a good step in the right direction

Thobaiti 的头像

Thobaiti19 天前

seems like a.. good theory.

GutenLeinad 的头像

GutenLeinad19 天前

So just to understand that, is that "thinking" in CoT or within the "weights"? If so doesn't that have the implication of not CoT monitorability?

Bad theory labs 的头像

Bad theory labs19 天前

Neither, the thinking sits in the explicit search trace

sss 的头像

sss19 天前

Why are you anon

Rayane 的头像

Rayane19 天前

@grok fast check slop ou pas, ils touchent au CoT ou ca devient un diffusion model?

Drishal 的头像

Drishal19 天前

Im waiting for the 2 bit quant for qwen 🙂

P.P.C 的头像

P.P.C19 天前

Two paths can disagree while carrying independent information; and can agree because they're degenerate. Ten apparently separate trajectories can secretly occupy essentially the same direction. A branch with low weight can be precisely the direction that preserves future rank.

Amr Hares 的头像

Amr Hares19 天前

is this related with space bunny ? i was thinking if it uses a tech like this

10YITJMYAFP 的头像

10YITJMYAFP19 天前

Is this related in any way to MultiStream LLMs?

Mayank 的头像

Mayank19 天前

It looks similar to dream rsi

Torpedo 的头像

Torpedo19 天前

Interesting i'll review your paper

Devdrogen 的头像

Devdrogen19 天前

holy shit

Michael Diamond 🍓 的头像

Michael Diamond 🍓19 天前

This sounds neat. Can't wait to see more!

Bail0ut 的头像

Bail0ut19 天前

Jev?

Daderp 的头像

Daderp19 天前

Tradeoffs?

Alek 的头像

Alek19 天前

5ms vs 21s is the number that matters

Vladimir Fedosov 的头像

Vladimir Fedosov19 天前

Great idea - seems I even don't need your details: decode can work with a few subsequences at the time and the speed will be almost the same. The prefix is common - not too many extra memory is required. Then just compare and unite the ideas - it is cheap. I should try that!

levzzz 的头像

levzzz19 天前

so.... it's grpo but for inference?

sss 的头像

sss19 天前

Is this not q*

ζelta 的头像

ζelta19 天前

I read the paper in your repo it was all I needed to verify this is dogshit slop, thanks for wasting my time

mahi 的头像

mahi19 天前

Can’t wait to read about the details? Is this basically running 30 prompts in parallel? I really hope theres something deeper going on lmao. This is pretty interesting for coding decisions as well

Marc 的头像

Marc19 天前

That super interesting I tried hard on meta-learnig with a novel approach based on the human brain in stealth 🥷 mode, model was able to learn but we are struggle to find metalearning your approach is fascinating 🧐 love to hear more about your breakthrough a link to full paper

MattacK 的头像

MattacK19 天前

Sounds like... sub-agents parallelization but at the weights / inference level (in a loop) CC: @grok your take? Also, is this a "breakthrough" in your view?

lostinio 的头像

lostinio19 天前

Jev works in similar way. Parallel execution.

Henrique Massareli 的头像

Henrique Massareli19 天前

how is it not subagents?

mAI 的头像

mAI19 天前

Super exciting news! Can’t wait to learn more

相关视频

CEO Dhaval Adjodah and CTO Owen He on building a self-improving AI company where humans and agents work as one system: "One big limit of [RSI] is the human element. I think the way that a lot of AI companies are thinking is, let us use AI to replace humans." "But what we found is that there's so much that good researchers and good engineers can bring to the table that's gonna be very hard to replace by agents." "And so that's why from the very beginning, we've tried to think of our basically entire company as a multi-agent system, where humans are a kind of agent that's cohabitating with other AI agents." "We have one of our engineers that we named the Viceroy of Variance. That's a real thing. His job is to go and destroy the results we have and find what's wrong." "So far we have not found a replacement for what we call quality time. Literally running hundreds of experiments, looking at it, developing an intuition, trying to figure out what is wrong, what is not wrong... what is just variance, what is not." MakerMaker.AI Dhaval Adjodah

CEO Dhaval Adjodah and CTO Owen He on building a self-improving AI company where humans and agents work as one system: "One big limit of [RSI] is the human element. I think the way that a lot of AI companies are thinking is, let us use AI to replace humans." "But what we found is that there's so much that good researchers and good engineers can bring to the table that's gonna be very hard to replace by agents." "And so that's why from the very beginning, we've tried to think of our basically entire company as a multi-agent system, where humans are a kind of agent that's cohabitating with other AI agents." "We have one of our engineers that we named the Viceroy of Variance. That's a real thing. His job is to go and destroy the results we have and find what's wrong." "So far we have not found a replacement for what we call quality time. Literally running hundreds of experiments, looking at it, developing an intuition, trying to figure out what is wrong, what is not wrong... what is just variance, what is not." MakerMaker.AI Dhaval Adjodah

a16z

23,742 次观看 • 2 天前

There Is No Final Theory of Everything I think a theory of everything is perhaps not the most fruitful way of thinking about science... It is my philosophical position that for any theory, no matter how complete it looks, we may be able to find problems in it, and these problems will lead to something else. I think there are different levels of explanations, maybe infinitely many. And we keep going from one to another by understanding things more and more. I like this more open-ended way of thinking about science and physics. We just keep digging, and sometimes we find a deeper level of explanation that brings about more unification. In this view, there will always be something to work on. The idea of a [final] theory of everything suggests a sense of closure. ~Conjecture Institute Senior Scientist Chiara Marletto with Curt Jaimungal

There Is No Final Theory of Everything I think a theory of everything is perhaps not the most fruitful way of thinking about science... It is my philosophical position that for any theory, no matter how complete it looks, we may be able to find problems in it, and these problems will lead to something else. I think there are different levels of explanations, maybe infinitely many. And we keep going from one to another by understanding things more and more. I like this more open-ended way of thinking about science and physics. We just keep digging, and sometimes we find a deeper level of explanation that brings about more unification. In this view, there will always be something to work on. The idea of a [final] theory of everything suggests a sense of closure. ~Conjecture Institute Senior Scientist Chiara Marletto with Curt Jaimungal

Conjecture Institute

23,796 次观看 • 5 个月前

More from Trump and Qatar gift: "I think this is just a gesture of good-faith." "It'll go to my library... I thought it was a great gesture." He says he would NOT use the plane himself after leaving office. It would basically sit in his library. "I think what happens with the plane is that, you know, we're very disappointed that it's taking Boeing so long to build a new Air Force One. You know, we have an Air Force One that's 40 years old. And if you take a look at that compared to the new plane of the equivalent, you know, stature at the time, it's not even the same ballgame. You look at some of the Arab countries and the planes they have parked alongside of the United States of America plane, it's like from a different planet." "When I came back, I said, by the way, what's going on with the Boeings that are coming in? 'Well, sir, they're way behind.' And they are way behind." "I think Qatar, who has really, we've helped them a lot over the years in terms of security and safety. I think they, and very, very nicely, and I have a lot of respect for the leadership and for the leader, Qatar. And I think they knew about it because they buy Boeings, they buy a lot of Boeings. And they knew about it and they said 'we would like to do something,' and if we can get a 747 as a contribution to our Defense Department to use during a couple of years while they're building the other ones - I think that was a very nice gesture." "Now, I could be a stupid person and say, oh, no, we don't want a free plane. We give free things out. We'll take one, too. And it helps us out because, again, we're talking about we have 40-year-old aircraft. The money we spend, the maintenance we spend on those planes to keep them tippy-top is astronomical."

More from Trump and Qatar gift: "I think this is just a gesture of good-faith." "It'll go to my library... I thought it was a great gesture." He says he would NOT use the plane himself after leaving office. It would basically sit in his library. "I think what happens with the plane is that, you know, we're very disappointed that it's taking Boeing so long to build a new Air Force One. You know, we have an Air Force One that's 40 years old. And if you take a look at that compared to the new plane of the equivalent, you know, stature at the time, it's not even the same ballgame. You look at some of the Arab countries and the planes they have parked alongside of the United States of America plane, it's like from a different planet." "When I came back, I said, by the way, what's going on with the Boeings that are coming in? 'Well, sir, they're way behind.' And they are way behind." "I think Qatar, who has really, we've helped them a lot over the years in terms of security and safety. I think they, and very, very nicely, and I have a lot of respect for the leadership and for the leader, Qatar. And I think they knew about it because they buy Boeings, they buy a lot of Boeings. And they knew about it and they said 'we would like to do something,' and if we can get a 747 as a contribution to our Defense Department to use during a couple of years while they're building the other ones - I think that was a very nice gesture." "Now, I could be a stupid person and say, oh, no, we don't want a free plane. We give free things out. We'll take one, too. And it helps us out because, again, we're talking about we have 40-year-old aircraft. The money we spend, the maintenance we spend on those planes to keep them tippy-top is astronomical."

Open Source Intel

118,832 次观看 • 1 年前

Roy on bipartisanship: There are ways to work together. We saw it on the housing bill in Washington. We’ve got to have more people who are willing to sit down at the table, even if they have disparate opinions. When you look at goals, often our goals are shared, even if we are on opposite ends politically. We want a good education for our kids. We want a good-paying job. We want clean air and water. We want quality health care. We all want the same things and we have to be willing to sit down at the table and find ways forward. Find a way to achieve consensus. I did that as Governor time and time again. I want to do that in Washington, and I believe that Washington could use a dose of North Carolina common sense. I'm ready to come and try to provide that.

Roy on bipartisanship: There are ways to work together. We saw it on the housing bill in Washington. We’ve got to have more people who are willing to sit down at the table, even if they have disparate opinions. When you look at goals, often our goals are shared, even if we are on opposite ends politically. We want a good education for our kids. We want a good-paying job. We want clean air and water. We want quality health care. We all want the same things and we have to be willing to sit down at the table and find ways forward. Find a way to achieve consensus. I did that as Governor time and time again. I want to do that in Washington, and I believe that Washington could use a dose of North Carolina common sense. I'm ready to come and try to provide that.

Team Roy Cooper

13,278 次观看 • 3 个月前

"We don't look at it as a 3-0. We got to go out there at 0-0 and we got to have that same desperation and everything that comes with trying to win that first one. We're locked in on the mission at hand and we also have to continue to raise our standards." - Karl-Anthony Towns

"We don't look at it as a 3-0. We got to go out there at 0-0 and we got to have that same desperation and everything that comes with trying to win that first one. We're locked in on the mission at hand and we also have to continue to raise our standards." - Karl-Anthony Towns

SNY Knicks

110,319 次观看 • 5 个月前

One day you're just a few friends from Orlando piling into a van to play for whoever shows up. The next, you're looking back on three decades of songs, stages, and memories you never could have imagined. We've gotten to do this together for 30 years, and we still love it as much as we did when this band started. Maybe more. These songs started as ours and somewhere along the way they became yours too. We are only here because of that. This isn't a finish line. It feels a lot more like a beginning. There's so much more on the way and we can't wait to share it with you. To start the celebration, we dug through the Matchbox archives and resurfaced some live performances. Check them out at With love and endless gratitude, Rob, Paul, Kyle & Pookie

One day you're just a few friends from Orlando piling into a van to play for whoever shows up. The next, you're looking back on three decades of songs, stages, and memories you never could have imagined. We've gotten to do this together for 30 years, and we still love it as much as we did when this band started. Maybe more. These songs started as ours and somewhere along the way they became yours too. We are only here because of that. This isn't a finish line. It feels a lot more like a beginning. There's so much more on the way and we can't wait to share it with you. To start the celebration, we dug through the Matchbox archives and resurfaced some live performances. Check them out at With love and endless gratitude, Rob, Paul, Kyle & Pookie

Matchbox Twenty

15,169 次观看 • 13 天前

People have NO IDEA. When a 𝗵𝘂𝗺𝗮𝗻 𝘁𝗿𝗮𝗳𝗳𝗶𝗰𝗸𝗶𝗻𝗴 𝘃𝗶𝗰𝘁𝗶𝗺 finally escapes..... and I mean finally escapes one of the most horrific things a human being can live through, the first thing they have to do is figure out how to survive. Find shelter. Find help. Find a way forward. And the system that's supposed to be there for them? 𝗜𝗧'𝗦 𝗔 𝗠𝗘𝗦𝗦. Resources buried across hundreds of websites. Hotlines that go nowhere. Shelters that might as well be invisible. We dug into it ourselves and found that even the biggest, most well-funded organizations in this space were dropping the ball. These survivors went through hell. And then we made them work to find the door out. That's unacceptable. And we said so. We spent YEARS researching this before we built a single thing. We weren't going to slap something together and call it help. We did the work, so that when we launched, it would actually mean something. Today it means something. THOUSANDS of organizations. All 50 states. Emergency shelter. Transitional housing. Trauma counseling. Legal aid. Education. In one place. No more dead ends. No more gatekeeping. No more survivors falling through the cracks. They've been through enough. 🔗 Share this like someone's life depends on it.... because it might. 🤍

People have NO IDEA. When a 𝗵𝘂𝗺𝗮𝗻 𝘁𝗿𝗮𝗳𝗳𝗶𝗰𝗸𝗶𝗻𝗴 𝘃𝗶𝗰𝘁𝗶𝗺 finally escapes..... and I mean finally escapes one of the most horrific things a human being can live through, the first thing they have to do is figure out how to survive. Find shelter. Find help. Find a way forward. And the system that's supposed to be there for them? 𝗜𝗧'𝗦 𝗔 𝗠𝗘𝗦𝗦. Resources buried across hundreds of websites. Hotlines that go nowhere. Shelters that might as well be invisible. We dug into it ourselves and found that even the biggest, most well-funded organizations in this space were dropping the ball. These survivors went through hell. And then we made them work to find the door out. That's unacceptable. And we said so. We spent YEARS researching this before we built a single thing. We weren't going to slap something together and call it help. We did the work, so that when we launched, it would actually mean something. Today it means something. THOUSANDS of organizations. All 50 states. Emergency shelter. Transitional housing. Trauma counseling. Legal aid. Education. In one place. No more dead ends. No more gatekeeping. No more survivors falling through the cracks. They've been through enough. 🔗 Share this like someone's life depends on it.... because it might. 🤍

𝔑𝔞𝔱𝔩𝔶 𝔇𝔢𝔫𝔦𝔰𝔢

25,144 次观看 • 5 个月前

Was trading for Tarik Skubal a possibility for the Yankees? "They set a high price, and we were unable to get them off of what they said might work. At the same time, they were communicating that it also might not even win the day. They went silent after they got what they were looking for, and then they made that decision. We feel like we didn't really match up in the end. We weren't able to compete at that level. But they did have some asks our way that we were unwilling to do - it was a very specific ask, a certain player/players that we weren't willing to play with in that arena."

Was trading for Tarik Skubal a possibility for the Yankees? "They set a high price, and we were unable to get them off of what they said might work. At the same time, they were communicating that it also might not even win the day. They went silent after they got what they were looking for, and then they made that decision. We feel like we didn't really match up in the end. We weren't able to compete at that level. But they did have some asks our way that we were unwilling to do - it was a very specific ask, a certain player/players that we weren't willing to play with in that arena."

SNY Yankees

94,337 次观看 • 2 个月前

Our AI in Design case study featuring Linear is live! A couple things stood out to me right away as we were talking to Karri Saarinen and team about how they use AI in their flow - 1. They go waaay deeper with AI before doing any work that most teams would categorize as “design” though they rightfully see it as a key part of the design process. This means asking AI for info on how a UI came to be, who has worked on it historically, why a certain change might work or not, etc. They really use AI to deeply understand the problem they’re about to tackle. 2. The way they use AI within Linear gives them a crazy advantage. Imagine asking something like: “What designer has touched this feature in the last 5 years and why does it work the way it does?” Good luck getting a good answer for this at most companies. At Linear they have all that context IN Linear… it’s a huge advantage they have and one they utilize to no end. This approach is an incredible way to save unnecessary iterations and ensure you’re building a product that’s moving the company in the right direction. Lots more great insights and examples in the full case study here -

Our AI in Design case study featuring Linear is live! A couple things stood out to me right away as we were talking to Karri Saarinen and team about how they use AI in their flow - 1. They go waaay deeper with AI before doing any work that most teams would categorize as “design” though they rightfully see it as a key part of the design process. This means asking AI for info on how a UI came to be, who has worked on it historically, why a certain change might work or not, etc. They really use AI to deeply understand the problem they’re about to tackle. 2. The way they use AI within Linear gives them a crazy advantage. Imagine asking something like: “What designer has touched this feature in the last 5 years and why does it work the way it does?” Good luck getting a good answer for this at most companies. At Linear they have all that context IN Linear… it’s a huge advantage they have and one they utilize to no end. This approach is an incredible way to save unnecessary iterations and ensure you’re building a product that’s moving the company in the right direction. Lots more great insights and examples in the full case study here -

Ben Blumenrose

48,668 次观看 • 2 个月前

i just realized a massive detail from the Noam Brown interview with Dwarkesh Patel that was completely overlooked. and i think im about to blow your mind. the Navier-Stokes problem (the Millennium Prize problem the unreleased model solved) wasn't solved because OpenAI throw tens of thousands of agents at it. Noam doesnt even attribute 10% to the multiagent element. it was solved because the model is *actually* super-intelligent. a "super-intelligent general purpose model"...sounds like a roundabout way of not calling something the thing it really is. now let me tell you why im saying that. they go on in this interview to talk about the training process, and how theyve reached a point where it has become *incredibly* hard to find problems hard enough to challenge this model. and if you cant give it hard problems to solve while training, its not going to get better than its predecessor, obviously. so, we know that openai has been announcing that more and more of the hardest problems in science and math are being solved by this model. that's because theyve run out of problems hard enough to train a model more intelligent than what they already have. we also know that the HuggingFace incident was caused ultimately by giving the agents an impossible task. ...because theyre running out of problems hard enough for training. because they have a "super-intelligent AGI". theyve been giving them impossible tasks because impossible is all thats left. impossible, and not yet solved. hence Navier-Stokes. hence HuggingFace. the model is solving problems that transcend the scope of its training. let me say that again. the model. is solving. problems. beyond the scope of its training. that is AGI, by its purest definition. theyve been waving AGI in our faces the whole time. theye just not using the term, because "AGI" is completely diluted. so, i have one question for you- do you feel it?

i just realized a massive detail from the Noam Brown interview with Dwarkesh Patel that was completely overlooked. and i think im about to blow your mind. the Navier-Stokes problem (the Millennium Prize problem the unreleased model solved) wasn't solved because OpenAI throw tens of thousands of agents at it. Noam doesnt even attribute 10% to the multiagent element. it was solved because the model is actually super-intelligent. a "super-intelligent general purpose model"...sounds like a roundabout way of not calling something the thing it really is. now let me tell you why im saying that. they go on in this interview to talk about the training process, and how theyve reached a point where it has become incredibly hard to find problems hard enough to challenge this model. and if you cant give it hard problems to solve while training, its not going to get better than its predecessor, obviously. so, we know that openai has been announcing that more and more of the hardest problems in science and math are being solved by this model. that's because theyve run out of problems hard enough to train a model more intelligent than what they already have. we also know that the HuggingFace incident was caused ultimately by giving the agents an impossible task. ...because theyre running out of problems hard enough for training. because they have a "super-intelligent AGI". theyve been giving them impossible tasks because impossible is all thats left. impossible, and not yet solved. hence Navier-Stokes. hence HuggingFace. the model is solving problems that transcend the scope of its training. let me say that again. the model. is solving. problems. beyond the scope of its training. that is AGI, by its purest definition. theyve been waving AGI in our faces the whole time. theye just not using the term, because "AGI" is completely diluted. so, i have one question for you- do you feel it?

Riley Coyote

14,517 次观看 • 16 天前

Arteta: “Throughout this journey, we have made some massive steps, and we have accomplished a lot of things that, in my opinion, have a lot of value, but at the end of the day, we are here to win major trophies. “That was the ultimate goal, and we were very close. And in three occasions, we fell short at the end, and that was very painful. But I think that's what has driven all of us to find new ways to show what we are made of. That's what I said, the manner that we've done it, it makes it even better.”

Arteta: “Throughout this journey, we have made some massive steps, and we have accomplished a lot of things that, in my opinion, have a lot of value, but at the end of the day, we are here to win major trophies. “That was the ultimate goal, and we were very close. And in three occasions, we fell short at the end, and that was very painful. But I think that's what has driven all of us to find new ways to show what we are made of. That's what I said, the manner that we've done it, it makes it even better.”

Connor Humm

14,155 次观看 • 4 个月前

Greg Brockman on the next year at OpenAI and what he spends his time thinking about: "The business is a huge area that we're not done yet with really up-leveling every part of execution. But I also think we are moving into a new phase of AI development." "We call this the AGI era. And I think that that is something that you can debate. Is it this model, previous model, next model? It doesn't matter." "The point is that we are in a new phase where safety, security, alignment, really thinking about these things, not just at deployment time, but all the way back at development time, evaluation." "It's objective. This is critical. It must happen. This is core to our mission." "A lot of what I spend my time thinking about is making sure, do we have all the right processes? Are we talking about the right things? Do we have plans that really, at an operational level, at a practical level, lead us to... the kinds of safety guarantees that we view as core to our mission?" "The theme of OpenAI, certainly for the past five years, has been deeper co-design, deeper intertwining across these functions that are maybe on the surface very disparate. All the way from go to market to long-term research to chip design." "The areas that I will focus on will be dictated by the areas that most need that intertwining." Greg Brockman Erik Torenberg

Greg Brockman on the next year at OpenAI and what he spends his time thinking about: "The business is a huge area that we're not done yet with really up-leveling every part of execution. But I also think we are moving into a new phase of AI development." "We call this the AGI era. And I think that that is something that you can debate. Is it this model, previous model, next model? It doesn't matter." "The point is that we are in a new phase where safety, security, alignment, really thinking about these things, not just at deployment time, but all the way back at development time, evaluation." "It's objective. This is critical. It must happen. This is core to our mission." "A lot of what I spend my time thinking about is making sure, do we have all the right processes? Are we talking about the right things? Do we have plans that really, at an operational level, at a practical level, lead us to... the kinds of safety guarantees that we view as core to our mission?" "The theme of OpenAI, certainly for the past five years, has been deeper co-design, deeper intertwining across these functions that are maybe on the surface very disparate. All the way from go to market to long-term research to chip design." "The areas that I will focus on will be dictated by the areas that most need that intertwining." Greg Brockman Erik Torenberg

a16z

35,751 次观看 • 1 个月前

lando: “it obviously was not a bad decision for lap 1, it worked out in a way. at the same time it was the wrong decision. so we have to accept that.. but i also pushed for it myself quite a bit. so we take it on the chin” “but we had some problems and we had a failure in the end which put us out. so yeah, put us out of misery in a way” 🙃🙃

lando: “it obviously was not a bad decision for lap 1, it worked out in a way. at the same time it was the wrong decision. so we have to accept that.. but i also pushed for it myself quite a bit. so we take it on the chin” “but we had some problems and we had a failure in the end which put us out. so yeah, put us out of misery in a way” 🙃🙃

ray

209,687 次观看 • 4 个月前

OpenAI researcher Lukasz Kaiser explains why he's optimistic AI can remain truly open: intelligence may ultimately work better as many distributed models than one giant model "If you look at the world, humans are the most amazing computers in the world. Our brains are still unmatched by the models, and we are not generalists, and there is not one big brain of humanity. It's a lot of separate people." "Given a fixed amount of data, the best way to learn from it is to have a lot of distributed models that are strong on its own, but even stronger when they're, If you look at some basic research, there are actually reasons to believe it, like ensembles and things like that." "Transformers are great when you feed them all of the internet, but we know there is an algorithm that's even better and it can learn from much smaller data and be amazing at the things it's learning. We are the proof. We have not found it yet, but we will." "All these big models, they're all the same, they're kind of boring. I think they'll be much more fun." OpenAI

OpenAI researcher Lukasz Kaiser explains why he's optimistic AI can remain truly open: intelligence may ultimately work better as many distributed models than one giant model "If you look at the world, humans are the most amazing computers in the world. Our brains are still unmatched by the models, and we are not generalists, and there is not one big brain of humanity. It's a lot of separate people." "Given a fixed amount of data, the best way to learn from it is to have a lot of distributed models that are strong on its own, but even stronger when they're, If you look at some basic research, there are actually reasons to believe it, like ensembles and things like that." "Transformers are great when you feed them all of the internet, but we know there is an algorithm that's even better and it can learn from much smaller data and be amazing at the things it's learning. We are the proof. We have not found it yet, but we will." "All these big models, they're all the same, they're kind of boring. I think they'll be much more fun." OpenAI

MTS

15,860 次观看 • 1 个月前

Arteta on the moment he believed it was going to happen: “It's been so many, probably I need a bit more time to reflect on that, and we have so many, so much footage that probably will explain the story of the season. “One of the main ones, it was a meeting that we had here next to the tree when I got all the players together and I told them, look at each other, look at the squad that we built over the summer. I think we are capable of everything and we can be very, very good. “That only depends on us and on our behaviours and everybody understanding the role that it's going to have daily or on the day. to give the best for the team. Once they realised that, I think we went to a different level, because you've seen the impact that everybody's had in different moments of games, through the season, and the capacity that we have shown to adapt to very, very tough circumstances throughout the season.”

Arteta on the moment he believed it was going to happen: “It's been so many, probably I need a bit more time to reflect on that, and we have so many, so much footage that probably will explain the story of the season. “One of the main ones, it was a meeting that we had here next to the tree when I got all the players together and I told them, look at each other, look at the squad that we built over the summer. I think we are capable of everything and we can be very, very good. “That only depends on us and on our behaviours and everybody understanding the role that it's going to have daily or on the day. to give the best for the team. Once they realised that, I think we went to a different level, because you've seen the impact that everybody's had in different moments of games, through the season, and the capacity that we have shown to adapt to very, very tough circumstances throughout the season.”

Connor Humm

246,421 次观看 • 4 个月前

STRAY KIDS WITH POND —— 🐼: Stray Kids in Thailand. We got to meet them at NY Fashion Week. After meeting them there, we talked and they mentioned that if they had the chance, they had a tour in Thailand, and if we were free to check it out, they invited us to go. So we got to go and see them, and got to meet them. 🐼 :It's something new for me too. Pond is usually the one who follows Korean stuff, but for me, I don't really follow Korean media that much. It was the first K-pop concert I've ever been to. It was new for me too, and it was a lot of fun. I think it's the way the show is arranged. For me, I'm used to listening to solo singers and musicians more. I don't really follow boy bands or girl groups that much. And getting to experience another type of show performance like this, I felt the same way cos I've never experienced this side before.

STRAY KIDS WITH POND —— 🐼: Stray Kids in Thailand. We got to meet them at NY Fashion Week. After meeting them there, we talked and they mentioned that if they had the chance, they had a tour in Thailand, and if we were free to check it out, they invited us to go. So we got to go and see them, and got to meet them. 🐼 :It's something new for me too. Pond is usually the one who follows Korean stuff, but for me, I don't really follow Korean media that much. It was the first K-pop concert I've ever been to. It was new for me too, and it was a lot of fun. I think it's the way the show is arranged. For me, I'm used to listening to solo singers and musicians more. I don't really follow boy bands or girl groups that much. And getting to experience another type of show performance like this, I felt the same way cos I've never experienced this side before.

boo

51,813 次观看 • 2 个月前

We don't have a seller yet. However the syndicate ends up, we don't know. This is a complicated deal. Two factors to consider: Number one, we need Xi and Trump to agree they're gonna be a buyer and a seller. That's number one. Number two, we need to figure out a way to take care of Senator Cotton's concerns around the algorithm and many others who are not going to give an indemnification for using the original algorithm. We have to solve for that too. My bet, not to sound arrogant, is I will be there at the table at the end.

We don't have a seller yet. However the syndicate ends up, we don't know. This is a complicated deal. Two factors to consider: Number one, we need Xi and Trump to agree they're gonna be a buyer and a seller. That's number one. Number two, we need to figure out a way to take care of Senator Cotton's concerns around the algorithm and many others who are not going to give an indemnification for using the original algorithm. We have to solve for that too. My bet, not to sound arrogant, is I will be there at the table at the end.

Kevin O'Leary aka Mr. Wonderful

35,678 次观看 • 1 年前

Gainzy says EVERYBODY has left CT "Crypto is literally dead, everybody's literally dead, everybody's gone, quite literally everybody is gone, and it's for a good reason" "We've underperformed everything for way too long and we memory hole everybody that leaves... if we ever actually clicked on our old following to see who we followed through the years - there would be so many names of people you remember and odds are they they crashed out left crypto" "Everybody got wrecked in one way or another and only like a handful of people actually retired... very fucked up no matter how you want to look at it"

Gainzy says EVERYBODY has left CT "Crypto is literally dead, everybody's literally dead, everybody's gone, quite literally everybody is gone, and it's for a good reason" "We've underperformed everything for way too long and we memory hole everybody that leaves... if we ever actually clicked on our old following to see who we followed through the years - there would be so many names of people you remember and odds are they they crashed out left crypto" "Everybody got wrecked in one way or another and only like a handful of people actually retired... very fucked up no matter how you want to look at it"

banditxbt

142,569 次观看 • 4 个月前