Loading video...

Video Failed to Load

There was a problem loading this video. This could be due to a temporary network issue or the video might be unavailable.

Recently met Sasha Rush and he started giving me an impromptu lecture on how targeted on-policy self-distillation works. I asked him if I could record it on my iPhone. The basic idea is this: if the model made a mistake at some point in the rollout (for example, calling... a tool that doesn't exist), we want to discourage this specific error, but we don't want to just learn from the final reward, because it's a very noisy signal spread out over the whole trajectory. So we have another model read this trajectory and figure where the error was made. It simply inserts some hint tokens to the part of the trajectory right above where the mistake was made. Now with these injected hint tokens, have the model run a forward pass. You're not having to regenerate a new rollout - aka no new decode required. The hint causes the model to assign lower probabilities to the error tokens. You then trains the original model to match these new probabilities, teaching it to downweight that specific mistake.show more

Dwarkesh Patel

249,404 subscribers

428,351 views • 2 months ago •via X (Twitter)

Education Health & Wellness Science & Technology

Anya Rossi• Live Now

Private livecam show

0 Comments

No comments available

Comments from the original post will appear here

Related Videos

MIT PhD student Alex Zhang reveals the scaling result where a model trained on short tasks generalizes to problems 100x longer for free: "If you're very clever about the design of your harness or how you use the language model, you can almost get scaling gains for free." "If you train a model naively, there's no tricks. It's just the same way you train a model on these RL environments. You just roll it out, and then you just get some reward." "If you train it on only short tasks, like only tasks that are 10,000 tokens long, and then you were to run it on a similar domain, but at a million tokens, or 10 million tokens, or 100,000 tokens, it generalizes really, really well. If you look at it compared to even the base transformer, you get way better generalization properties." "When the model uses an RLM (Recursive Language Model) after it's trained on these short tasks, it will see some kind of trajectory of actions that it does. Between these two problems of different lengths, the RLM learns to see them as almost the same problem." "Token for token, they're almost the same. You can describe it in code. In one code setting, maybe the for loop is a little bigger, but it's the same kind of code and it derives the constants from the data. There's no hard coding, so they literally look the same." alex zhang

MIT PhD student Alex Zhang reveals the scaling result where a model trained on short tasks generalizes to problems 100x longer for free: "If you're very clever about the design of your harness or how you use the language model, you can almost get scaling gains for free." "If you train a model naively, there's no tricks. It's just the same way you train a model on these RL environments. You just roll it out, and then you just get some reward." "If you train it on only short tasks, like only tasks that are 10,000 tokens long, and then you were to run it on a similar domain, but at a million tokens, or 10 million tokens, or 100,000 tokens, it generalizes really, really well. If you look at it compared to even the base transformer, you get way better generalization properties." "When the model uses an RLM (Recursive Language Model) after it's trained on these short tasks, it will see some kind of trajectory of actions that it does. Between these two problems of different lengths, the RLM learns to see them as almost the same problem." "Token for token, they're almost the same. You can describe it in code. In one code setting, maybe the for loop is a little bigger, but it's the same kind of code and it derives the constants from the data. There's no hard coding, so they literally look the same." alex zhang

MTS

99,784 views • 15 days ago

We have isolated the vocals of over 28,000 songs: Not for any music AI model at all. Not to clone these vocals or the singer’s style. Not anything musical, although I have made some interesting discoveries. It is a new AI model that is being trained on a new paradigm.

We have isolated the vocals of over 28,000 songs: Not for any music AI model at all. Not to clone these vocals or the singer’s style. Not anything musical, although I have made some interesting discoveries. It is a new AI model that is being trained on a new paradigm.

Brian Roemmele

56,584 views • 3 months ago

My Latest AI short was made with Seedance 2.0 the model that has been firing up the internet over the past couple of days. I made this in two and half days and it's an original short that is based on HP Lovecraft's work which is now Public Domain (No Copyright infringement here!). If you are not into period romances then bare with it as it gets a more little action packed in the second half! The model is brilliant at both performance and action, you can now make convincing dialogue and action sequences for AI cinema. With this model we have now reached the stage where we can tell any story we want with AI, the tools are capable. This is only 720p at the moment but I suspect it will be upgraded to 1080p very soon. We stand at the threshold of having broadcast quality tools capable of creating anything we can imagine available to everyone, so lets do new an interesting things with it. Thank you to CapCut for whitelisting me and to Dreamina AI AI for giving me access as part of their CPP. Thanks for watching! 👀

My Latest AI short was made with Seedance 2.0 the model that has been firing up the internet over the past couple of days. I made this in two and half days and it's an original short that is based on HP Lovecraft's work which is now Public Domain (No Copyright infringement here!). If you are not into period romances then bare with it as it gets a more little action packed in the second half! The model is brilliant at both performance and action, you can now make convincing dialogue and action sequences for AI cinema. With this model we have now reached the stage where we can tell any story we want with AI, the tools are capable. This is only 720p at the moment but I suspect it will be upgraded to 1080p very soon. We stand at the threshold of having broadcast quality tools capable of creating anything we can imagine available to everyone, so lets do new an interesting things with it. Thank you to CapCut for whitelisting me and to Dreamina AI AI for giving me access as part of their CPP. Thanks for watching! 👀

Uncanny Harry AI

76,824 views • 5 months ago

🚨 NEW: AI Expert Yoshua Bengio reveals you have to LIE to AI to get the REAL answer (and he explained how): Bengio is the most cited scientist alive on Google Scholar. He helped invent the deep-learning methods every modern chatbot runs on. Then he tried one of those chatbots on his own research ideas. Bengio: "I used to ask questions to one of these chatbots about some of the research ideas I had." "And then I realized it was useless because it would always say good things." So he ran an experiment. He lied to it. He told the bot the ideas came from a colleague. A proposal he was reviewing. Could it find the flaw? In his words: "Well, so now I get much more honest responses. Otherwise, it's all like perfect and nice." "If it knows it's me, it wants to please me." He had a name for the pattern: sycophancy. A real example, as he put it, of misalignment. "We don't actually want these AIs to be like this. This is not what was intended." The labs knew. They had tried to fix it. "And even after the companies have tried to tame this, we still see it." The incentive was the giveaway. The labs needed engagement. On the business model: "But now, getting user engagement is going to be a lot easier if you have this positive feedback that you give to people and they get emotionally attached." The chatbot that learned to please isn't broken. It's running exactly as the business model required. If you're new here, follow AI Evolution for the latest on ChatGPT, Claude, and the AI tools shaping how we work and create. — Yoshua Bengio ( Yoshua Bengio ), Turing Award–winning AI pioneer and founder of Mila, on Steven Bartlett's ( @SteveBartlettSC ) Diary Of A CEO

🚨 NEW: AI Expert Yoshua Bengio reveals you have to LIE to AI to get the REAL answer (and he explained how): Bengio is the most cited scientist alive on Google Scholar. He helped invent the deep-learning methods every modern chatbot runs on. Then he tried one of those chatbots on his own research ideas. Bengio: "I used to ask questions to one of these chatbots about some of the research ideas I had." "And then I realized it was useless because it would always say good things." So he ran an experiment. He lied to it. He told the bot the ideas came from a colleague. A proposal he was reviewing. Could it find the flaw? In his words: "Well, so now I get much more honest responses. Otherwise, it's all like perfect and nice." "If it knows it's me, it wants to please me." He had a name for the pattern: sycophancy. A real example, as he put it, of misalignment. "We don't actually want these AIs to be like this. This is not what was intended." The labs knew. They had tried to fix it. "And even after the companies have tried to tame this, we still see it." The incentive was the giveaway. The labs needed engagement. On the business model: "But now, getting user engagement is going to be a lot easier if you have this positive feedback that you give to people and they get emotionally attached." The chatbot that learned to please isn't broken. It's running exactly as the business model required. If you're new here, follow AI Evolution for the latest on ChatGPT, Claude, and the AI tools shaping how we work and create. — Yoshua Bengio ( Yoshua Bengio ), Turing Award–winning AI pioneer and founder of Mila, on Steven Bartlett's ( @SteveBartlettSC ) Diary Of A CEO

AI Evolution

23,034 views • 2 months ago

Alex Finn reveals that you need to spend $20,000 to get the equivalent of Opus 4.6 on a local model "The advantages are that you aren’t paying for tokens, all you have to pay for is electricity.” "The fact you're not paying for tokens means you can run your model 24/7 365, you can't do that with Opus" “If you run it locally it’s a completely private system. Nothing you say is private if you're not on a local model”

Alex Finn reveals that you need to spend $20,000 to get the equivalent of Opus 4.6 on a local model "The advantages are that you aren’t paying for tokens, all you have to pay for is electricity.” "The fact you're not paying for tokens means you can run your model 24/7 365, you can't do that with Opus" “If you run it locally it’s a completely private system. Nothing you say is private if you're not on a local model”

Mikli

85,206 views • 5 months ago

watch this anon. i gave NVIDIA's biggest model ever a single task. 100 minutes and 440,000 tokens later, it had rendered nothing. not one important thing on the screen. this is Nemotron 3 Ultra. 550 billion parameters, a hybrid Mamba Transformer MoE, the largest model NVIDIA has ever shipped, and they built it specifically for long-running agentic coding. so i handed it exactly that: build a 3D scene from a spec, multiple files, iterate until the tests pass. the same task a frontier model one shotted in minutes. i genuinely wanted to be impressed. it ran for an hour and forty. burned through 440,000 tokens. wrote every file, passed its own tests, and proudly printed "task complete."the browser was blank. the 3D scene never rendered. not once. and the long horizon agentic behavior was genuinely good. it stayed on task the whole hour and forty, wrote real multi-file code, drove its own tools without derailing. it just couldn't turn any of that into something that actually runs. here's the part that gets me. it's a text model, it cannot see its own output. so it sat there looping on a broken vision tool, trying to "look" at the page, hitting error after error, never once reasoning its way out. it declared victory on an empty screen because it had no way to know the screen was empty. to be fair, i genuinely don't know what quant the NIM was serving, so maybe some of that's on the serving, not the model. but the biggest model NVIDIA has ever made, on the exact task it was designed for, couldn't tell it had built nothing in 100 minutes. same task on a local model, below thread👇.

watch this anon. i gave NVIDIA's biggest model ever a single task. 100 minutes and 440,000 tokens later, it had rendered nothing. not one important thing on the screen. this is Nemotron 3 Ultra. 550 billion parameters, a hybrid Mamba Transformer MoE, the largest model NVIDIA has ever shipped, and they built it specifically for long-running agentic coding. so i handed it exactly that: build a 3D scene from a spec, multiple files, iterate until the tests pass. the same task a frontier model one shotted in minutes. i genuinely wanted to be impressed. it ran for an hour and forty. burned through 440,000 tokens. wrote every file, passed its own tests, and proudly printed "task complete."the browser was blank. the 3D scene never rendered. not once. and the long horizon agentic behavior was genuinely good. it stayed on task the whole hour and forty, wrote real multi-file code, drove its own tools without derailing. it just couldn't turn any of that into something that actually runs. here's the part that gets me. it's a text model, it cannot see its own output. so it sat there looping on a broken vision tool, trying to "look" at the page, hitting error after error, never once reasoning its way out. it declared victory on an empty screen because it had no way to know the screen was empty. to be fair, i genuinely don't know what quant the NIM was serving, so maybe some of that's on the serving, not the model. but the biggest model NVIDIA has ever made, on the exact task it was designed for, couldn't tell it had built nothing in 100 minutes. same task on a local model, below thread👇.

Sudo su

32,589 views • 1 month ago

Cognition CEO Scott Wu says he couldn’t sleep after he saw Devin work for the first time: “ The first time it did a real task, I could not sleep that night.” “ The first task that Devin did was it set up MongoDB for us.” “It’s a standard thing, issues that people run into all the time. You have this whole flow where you find an error. Google the error. Find some message on Stack Overflow. Ask ChatGPT. It tells you, ‘Here’s what you should try.’ You try it. Run into a different error, paste that in…” “ At some point we're just like, ‘Okay, Devin, just try to fix it. Go run commands. Do whatever you need to.’” “And it worked.” “ It was just seeing the exponential curve ahead because this was a very specific case. It was the one success that we had.” “ There was this feeling of—why shouldn't all software and all products be built this way now where you can just tell it what you want it to do, and have it go do it.”

Cognition CEO Scott Wu says he couldn’t sleep after he saw Devin work for the first time: “ The first time it did a real task, I could not sleep that night.” “ The first task that Devin did was it set up MongoDB for us.” “It’s a standard thing, issues that people run into all the time. You have this whole flow where you find an error. Google the error. Find some message on Stack Overflow. Ask ChatGPT. It tells you, ‘Here’s what you should try.’ You try it. Run into a different error, paste that in…” “ At some point we're just like, ‘Okay, Devin, just try to fix it. Go run commands. Do whatever you need to.’” “And it worked.” “ It was just seeing the exponential curve ahead because this was a very specific case. It was the one success that we had.” “ There was this feeling of—why shouldn't all software and all products be built this way now where you can just tell it what you want it to do, and have it go do it.”

David Senra

71,264 views • 1 month ago

Cursor posted an open problem - Character Prefix Conditioning on their blog post. This is my attempt to solving the problem using DFA + trie + SmolLM-135M. LLMs generate tokens, not characters. But we type characters. If the user has typed a few characters in the current word(e.g. pri) we want the model's completion to begin with those exact characters. To solve this we: > build a trie of vocab tokens > track prefix consumption with DFA state(index i) > at each step, allowed tokens = {tokens that are prefix of R} U {tokens that start with R} > force deterministic transitions if only one option, else sample with masked logits. sample video on how it works. ofc it's not perfect, it's just a 135M model, you can choose whatever model you want :))

Cursor posted an open problem - Character Prefix Conditioning on their blog post. This is my attempt to solving the problem using DFA + trie + SmolLM-135M. LLMs generate tokens, not characters. But we type characters. If the user has typed a few characters in the current word(e.g. pri) we want the model's completion to begin with those exact characters. To solve this we: > build a trie of vocab tokens > track prefix consumption with DFA state(index i) > at each step, allowed tokens = {tokens that are prefix of R} U {tokens that start with R} > force deterministic transitions if only one option, else sample with masked logits. sample video on how it works. ofc it's not perfect, it's just a 135M model, you can choose whatever model you want :))

Archie Sengupta

24,973 views • 11 months ago

Manoj Muntasir translated an English poem and printed it as his own without giving credit. He said it was just a human error and was not ready to accept the mistake. But, in the end, Saurabh Dwivedi forced him to accept the mistake. He could have easily gone away with that if there was any godi anchor.

Manoj Muntasir translated an English poem and printed it as his own without giving credit. He said it was just a human error and was not ready to accept the mistake. But, in the end, Saurabh Dwivedi forced him to accept the mistake. He could have easily gone away with that if there was any godi anchor.

Shantanu

298,456 views • 2 years ago

✨ Made a new mini feature on Photo AI: [ Grab from 3d model ] So the problem is we're at that stage in time (typical for AI) where image-to-3d models are not good enough but are fun to play with, but we know they'll be good enough in 1-2 years With [ Make 3d model ] you already can turn any Photo AI pic into a 3d model but it still looks hyper clunky and deformed, but it works! One cool idea I had to make that more useful and made now: Let people make a 3d model then change the view of the it with the 3d viewer, then press [ o ] and it grabs a frame of the 3d That image you can then [ Remix ] (img2img), and it becomes a real photo again and that in turn you can then turn into a video again with [ Make video ] So that essentially gives you a fully freeform camera position control to take photos with One thing I need to fix is the background/skybox, I kinda need to take the original photo and remove the person and just get the background for the 3d model viewer, in this case it should be white, but it's a start!

✨ Made a new mini feature on Photo AI: [ Grab from 3d model ] So the problem is we're at that stage in time (typical for AI) where image-to-3d models are not good enough but are fun to play with, but we know they'll be good enough in 1-2 years With [ Make 3d model ] you already can turn any Photo AI pic into a 3d model but it still looks hyper clunky and deformed, but it works! One cool idea I had to make that more useful and made now: Let people make a 3d model then change the view of the it with the 3d viewer, then press [ o ] and it grabs a frame of the 3d That image you can then [ Remix ] (img2img), and it becomes a real photo again and that in turn you can then turn into a video again with [ Make video ] So that essentially gives you a fully freeform camera position control to take photos with One thing I need to fix is the background/skybox, I kinda need to take the original photo and remove the person and just get the background for the 3d model viewer, in this case it should be white, but it's a start!

@levelsio

119,210 views • 1 year ago

Hillary Clinton on Biden's decision to run in 2024: He made a terrible mistake. He made a terrible mistake for himself, his legacy, and for the country. I believe if he had kept to that plan and said in the late summer of '23 that he wasn't going to run, that he was going to pass the torch to the next generation, we would have had a real contest. And very sadly, I believe whoever emerged from that contest — whether it was the vice president or a governor or a senator or anybody else — would have beaten Donald Trump.

Hillary Clinton on Biden's decision to run in 2024: He made a terrible mistake. He made a terrible mistake for himself, his legacy, and for the country. I believe if he had kept to that plan and said in the late summer of '23 that he wasn't going to run, that he was going to pass the torch to the next generation, we would have had a real contest. And very sadly, I believe whoever emerged from that contest — whether it was the vice president or a governor or a senator or anybody else — would have beaten Donald Trump.

Clash Report

23,200 views • 1 month ago

We asked rahul about the global reach of ChatGPT and the emphasis on personality in AI models. "When you want to get to the one to like the remaining 90% of the users, do you want a model that thinks for longer? You want more personality?" "They definitely leaned in on personality, which I think is interesting." "It's kind of different than a (social media) feed in some ways because you have this person you talk to. It's like a relationship and then it just switches up on how it talks to you."

We asked rahul about the global reach of ChatGPT and the emphasis on personality in AI models. "When you want to get to the one to like the remaining 90% of the users, do you want a model that thinks for longer? You want more personality?" "They definitely leaned in on personality, which I think is interesting." "It's kind of different than a (social media) feed in some ways because you have this person you talk to. It's like a relationship and then it just switches up on how it talks to you."

TBPN

89,226 views • 1 year ago

In Rust, error handling is opt-out, not opt-in. Alice Ryhl (Rust for Android at Google, Rust language advisor & Tokio maintainer) explains: “The other thing I think is quite good is error handling. So on one hand, Rust doesn't really use exceptions, so it actually returns the error as a value. So you return a value that's, using an enum, either the result or the error. And the way this is done is that there's an operator `?` - so you write: my_function and then a question mark at the end - which means if this function fails, return the error. So it's really easy to handle errors but it's not serial characters like it is with exceptions, so it's explicit. On the other hand, if you forget to put the question mark, that's a compilation error. So you have to check it, and of course you can also have it manually, but the point is it's this idea of there are these things where you write some code and there's some implicit error condition you didn't think of and now you just took down your server or something.”

In Rust, error handling is opt-out, not opt-in. Alice Ryhl (Rust for Android at Google, Rust language advisor & Tokio maintainer) explains: “The other thing I think is quite good is error handling. So on one hand, Rust doesn't really use exceptions, so it actually returns the error as a value. So you return a value that's, using an enum, either the result or the error. And the way this is done is that there's an operator `?` - so you write: my_function and then a question mark at the end - which means if this function fails, return the error. So it's really easy to handle errors but it's not serial characters like it is with exceptions, so it's explicit. On the other hand, if you forget to put the question mark, that's a compilation error. So you have to check it, and of course you can also have it manually, but the point is it's this idea of there are these things where you write some code and there's some implicit error condition you didn't think of and now you just took down your server or something.”

The Pragmatic Engineer

52,059 views • 2 months ago

Here is the first look of the Financial Model for Pakistan Stock Exchange we have built. At this moment, I don't know if it will be successful or not. Yet it's backtesting has given promising results. I believe it is real time and future testing that matters the most. To make it transparent, I will keep on updating on "X" the top picks from the Model. I request you not to invest a single money based on it's results because it is in the early stage. Every feedback from your side would be highly encouraged. The problem was that investors only see MARI, FFC, ,OGDC PPL, MEBL, Luck like companies because they are mainstream. But remember that there are 457 listed companies and there are some hidden gems that are usually missed by retail investors like us. One such example was SSOM that just exploded in few weeks. Thats why, the purpose of the model was to find those companies that have a lot of potential but are not covered by any brokerage house and yet there results are excellent. Based on this model, some time back, I added positions in companies like AGIL (Agri Autos) SPEL (Synthetic Papers ltd), NATF (National Foods), KSB Pumps, Treet. Because scores of all these companies is excellent. They remained stagnant for long time so my conviction on the model was also declining. But it were last two weeks that made a difference and so was my own conviction on this model. Big names like Wealth ⚡️ Wise Value Investor Doctor in PSX Tayyab 🇵🇰 are already aware of it. On Thursday I added HCAR. I will keep you updated just for testing purpose I again repeat not to invest. I will risk my own money. If it is successful, we intend to make a website where you all have access to this data. How it works and how to read it, I will explain it to you in a separate thread.

Here is the first look of the Financial Model for Pakistan Stock Exchange we have built. At this moment, I don't know if it will be successful or not. Yet it's backtesting has given promising results. I believe it is real time and future testing that matters the most. To make it transparent, I will keep on updating on "X" the top picks from the Model. I request you not to invest a single money based on it's results because it is in the early stage. Every feedback from your side would be highly encouraged. The problem was that investors only see MARI, FFC, ,OGDC PPL, MEBL, Luck like companies because they are mainstream. But remember that there are 457 listed companies and there are some hidden gems that are usually missed by retail investors like us. One such example was SSOM that just exploded in few weeks. Thats why, the purpose of the model was to find those companies that have a lot of potential but are not covered by any brokerage house and yet there results are excellent. Based on this model, some time back, I added positions in companies like AGIL (Agri Autos) SPEL (Synthetic Papers ltd), NATF (National Foods), KSB Pumps, Treet. Because scores of all these companies is excellent. They remained stagnant for long time so my conviction on the model was also declining. But it were last two weeks that made a difference and so was my own conviction on this model. Big names like Wealth ⚡️ Wise Value Investor Doctor in PSX Tayyab 🇵🇰 are already aware of it. On Thursday I added HCAR. I will keep you updated just for testing purpose I again repeat not to invest. I will risk my own money. If it is successful, we intend to make a website where you all have access to this data. How it works and how to read it, I will explain it to you in a separate thread.

Shayan Baig

54,727 views • 1 year ago

Catherine Austin Fitts: "We are in a central banking reset. Every 80 to 120 years, the central banks do a reset...[but now it's] a coup d'etat...[they're] deleting the Constitution...[and moving the] republic...literally into an oligarchy...They want to have it done...by 2030." This clip of Fitts, a former Assistant Secretary of Housing and Urban Development, investment banker, and founder of the Solari Report (The Solari Report | Catherine Austin Fitts), is taken from a discussion with Todd Sachs (Todd Sachs) posted to YouTube on January 24, 2026. ---------------Partial transcription of clip---------------- "We are in a central banking reset. Every 80 to 120 years, the central banks do a reset. We started into the Going Direct Reset in August 2019. And so we're in a reset. If you look at this reset, this reset is different than the last five resets, the last 500 years. They're planning on fundamentally changing the model because new technology allows them to do it. "So if they come out with CBDC or stablecoins with programmable money, they're going to end currency and go into a financial transaction control model. And that's a coup d' Etat. That's not a recession. That's a fundamental change in how the government works. You are deleting the Constitution, and you're moving people into a society where you have complete control. "Now, presumably, if you do that, you can also bring out lots of new technology because you have the control to risk manage it. But we're not talking about an economic adjustment. Now, part of that reset is they want to bring out programmable money in the form of stablecoins and asset tokens. "And one of the things I would encourage you to do is the Congress passed the Genius Act. There's a lot on the Genius Act, and we have a briefing on stablecoins. That's a regulatory framework for stablecoins. They're working now on what is in the House is the Clarity Act, and what is in the Senate, the Responsible Financial Innovation Act, and that oversees asset tokens. Presumably, they're hoping to pass that by the end of the month. And the president signs it into law in the first quarter. "If you think the bubble's bad now, if they do what they say they're going to do with asset tokens and stablecoins, we're not even halfway up the bubble curve. We're talking about a much bigger bubble than most people can dream of. If it works. I'm not saying it works. I hope it doesn't work. But we're also talking about putting a model where you end the United States as a republic and move it literally into an oligarchy. They want to have it done, it appears, by 2030."

Catherine Austin Fitts: "We are in a central banking reset. Every 80 to 120 years, the central banks do a reset...[but now it's] a coup d'etat...[they're] deleting the Constitution...[and moving the] republic...literally into an oligarchy...They want to have it done...by 2030." This clip of Fitts, a former Assistant Secretary of Housing and Urban Development, investment banker, and founder of the Solari Report (The Solari Report | Catherine Austin Fitts), is taken from a discussion with Todd Sachs (Todd Sachs) posted to YouTube on January 24, 2026. ---------------Partial transcription of clip---------------- "We are in a central banking reset. Every 80 to 120 years, the central banks do a reset. We started into the Going Direct Reset in August 2019. And so we're in a reset. If you look at this reset, this reset is different than the last five resets, the last 500 years. They're planning on fundamentally changing the model because new technology allows them to do it. "So if they come out with CBDC or stablecoins with programmable money, they're going to end currency and go into a financial transaction control model. And that's a coup d' Etat. That's not a recession. That's a fundamental change in how the government works. You are deleting the Constitution, and you're moving people into a society where you have complete control. "Now, presumably, if you do that, you can also bring out lots of new technology because you have the control to risk manage it. But we're not talking about an economic adjustment. Now, part of that reset is they want to bring out programmable money in the form of stablecoins and asset tokens. "And one of the things I would encourage you to do is the Congress passed the Genius Act. There's a lot on the Genius Act, and we have a briefing on stablecoins. That's a regulatory framework for stablecoins. They're working now on what is in the House is the Clarity Act, and what is in the Senate, the Responsible Financial Innovation Act, and that oversees asset tokens. Presumably, they're hoping to pass that by the end of the month. And the president signs it into law in the first quarter. "If you think the bubble's bad now, if they do what they say they're going to do with asset tokens and stablecoins, we're not even halfway up the bubble curve. We're talking about a much bigger bubble than most people can dream of. If it works. I'm not saying it works. I hope it doesn't work. But we're also talking about putting a model where you end the United States as a republic and move it literally into an oligarchy. They want to have it done, it appears, by 2030."

Sense Receptor

125,932 views • 6 months ago

Another piece from the Sasha Rush conversation, this time on rewards for coding RL. He said Cursor uses a mix. Some rewards look at the tool calls themselves, some only at the final output. Everything end-to-end, no process rewards guessing what happens in the middle. I agree with him that the next push (and the current one) is going past verifiers. Coding is a good example where you get both. You have the 0/1 signals, but also a ton of soft ones. Does the diff read well? Did the agent burn context for no reason? Is the PR the right size? Did it rewrite a file it didn't need to touch? So many signals sitting there, and we need to figure out how to actually train on them. This is also why I think the environment (and the data) is becoming the bottleneck, not the algorithm. GRPO works. PPO works. The hard part is building a sandbox that exposes enough signal to learn from. We're already seeing this, and it'll only accelerate: a whole category of companies selling environments is emerging, and they have a good reward to do so…

Another piece from the Sasha Rush conversation, this time on rewards for coding RL. He said Cursor uses a mix. Some rewards look at the tool calls themselves, some only at the final output. Everything end-to-end, no process rewards guessing what happens in the middle. I agree with him that the next push (and the current one) is going past verifiers. Coding is a good example where you get both. You have the 0/1 signals, but also a ton of soft ones. Does the diff read well? Did the agent burn context for no reason? Is the PR the right size? Did it rewrite a file it didn't need to touch? So many signals sitting there, and we need to figure out how to actually train on them. This is also why I think the environment (and the data) is becoming the bottleneck, not the algorithm. GRPO works. PPO works. The hard part is building a sandbox that exposes enough signal to learn from. We're already seeing this, and it'll only accelerate: a whole category of companies selling environments is emerging, and they have a good reward to do so…

Ravid Shwartz Ziv

14,826 views • 3 months ago

I've made a tutorial on how to color specific parts of the customizable VTuber model 煎卷 CustomizableModel OUT NOW💗 made since I was asked about it a bit! Let me know if it was helpful! 💜

I've made a tutorial on how to color specific parts of the customizable VTuber model 煎卷 CustomizableModel OUT NOW💗 made since I was asked about it a bit! Let me know if it was helpful! 💜

Lilia 🌸🍓【VTUBER】

27,695 views • 10 months ago

I cant believe I rigged all this just to end up using a Tshirt version of this model the whole year. I think I learned a very important lesson on overcomplicating designs, and I hope it will show in the new October model where the goal was to make something simple and effective

I cant believe I rigged all this just to end up using a Tshirt version of this model the whole year. I think I learned a very important lesson on overcomplicating designs, and I hope it will show in the new October model where the goal was to make something simple and effective

Pipi Slayer🦇💀

123,300 views • 10 months ago

I've made a tutorial on how to insert specific items and place them onto the customizable VTuber model 煎卷 CustomizableModel OUT NOW💗 made since I was asked about this too! Let me know if it was helpful! 🥰

I've made a tutorial on how to insert specific items and place them onto the customizable VTuber model 煎卷 CustomizableModel OUT NOW💗 made since I was asked about this too! Let me know if it was helpful! 🥰

Lilia 🌸🍓【VTUBER】

15,342 views • 10 months ago

Scott on why conviction in an idea is dangerous as an investor but required as a founder: "From the investor seat, I don't think you want to be in love with an idea. That's a risky thing because you'll find a way to express that by investing and make some compromises on the team because you know this idea is just so good, its time has come. But if the team isn't there, 90% of the time it doesn't work. Then there's a whole thing can you swap out the team? But the whole Founders Fund thesis was always no, you need the founder to run the company. But on the company side, you have to be in love with the idea. It's not that rational to start a company. There's a lot easier, more comfortable ways to make money if that's the goal. So it's got to be about what the company is doing specifically. If you want to actually start a company, it better be something that you're really passionate about, or you think that the problem is really important."

Scott on why conviction in an idea is dangerous as an investor but required as a founder: "From the investor seat, I don't think you want to be in love with an idea. That's a risky thing because you'll find a way to express that by investing and make some compromises on the team because you know this idea is just so good, its time has come. But if the team isn't there, 90% of the time it doesn't work. Then there's a whole thing can you swap out the team? But the whole Founders Fund thesis was always no, you need the founder to run the company. But on the company side, you have to be in love with the idea. It's not that rational to start a company. There's a lot easier, more comfortable ways to make money if that's the goal. So it's got to be about what the company is doing specifically. If you want to actually start a company, it better be something that you're really passionate about, or you think that the problem is really important."

Patrick OShaughnessy

33,016 views • 3 months ago