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。

What does it tell us about human intelligence that half of yeast's essential genes are interchangeable with ours? The genome encodes hyperparameters and reward functions, not learned weights. And according to Adam Marblestone and Steven Byrnes, evolution put it's cleverness into designing these reward functions to teach you what... show more

Dwarkesh Patel

184,824 subscribers

59,945 次观看 • 9 个月前 •via X (Twitter)

教育健康养生科学技术

Anya Rossi• Live Now

Private livecam show

0 条评论

暂无评论

原始帖子的评论将显示在这里

相关视频

There's a quadrillion-dollar question at the heart of AI: Why are humans so much more sample efficient compared to LLM? There are three possible answers: 1. Architecture and hyperparameters (aka transformer vs whatever ‘algo’ cortical columns are implementing) 2. Learning rule (backprop vs whatever brain is doing) 3. Reward function Adam Marblestone believes the answer is the reward function. ML likes to use pretty simple loss functions, like cross-entropy. These are easy to work with. But they might be too simple for sample-efficient learning. Adam thinks that, in humans, the large number of highly specialised cells in the ‘lizard brain’ might actually be encoding information for sophisticated loss functions, used for ‘training’ in the more sophisticated areas like the cortex and amygdala. Like: the human genome is barely 3 gigabytes (compare that to the TBs of parameters that encode frontier LLM weights). So how can it include all the information necessary to build highly intelligent learners? Well, if the key to sample-efficient learning resides in the loss function, even very complicated loss functions can still be expressed in a couple hundred lines of Python code.

There's a quadrillion-dollar question at the heart of AI: Why are humans so much more sample efficient compared to LLM? There are three possible answers: 1. Architecture and hyperparameters (aka transformer vs whatever ‘algo’ cortical columns are implementing) 2. Learning rule (backprop vs whatever brain is doing) 3. Reward function Adam Marblestone believes the answer is the reward function. ML likes to use pretty simple loss functions, like cross-entropy. These are easy to work with. But they might be too simple for sample-efficient learning. Adam thinks that, in humans, the large number of highly specialised cells in the ‘lizard brain’ might actually be encoding information for sophisticated loss functions, used for ‘training’ in the more sophisticated areas like the cortex and amygdala. Like: the human genome is barely 3 gigabytes (compare that to the TBs of parameters that encode frontier LLM weights). So how can it include all the information necessary to build highly intelligent learners? Well, if the key to sample-efficient learning resides in the loss function, even very complicated loss functions can still be expressed in a couple hundred lines of Python code.

Dwarkesh Patel

951,098 次观看 • 5 个月前

You get what you incentivize, so when western societies reward victimhood, as ours does, we get more of it. It is imperative we teach our kids to reject this mindset. Life’s not fair, it never will be, and you just have to suck it up. Practice more stoicism and reward achievement!

You get what you incentivize, so when western societies reward victimhood, as ours does, we get more of it. It is imperative we teach our kids to reject this mindset. Life’s not fair, it never will be, and you just have to suck it up. Practice more stoicism and reward achievement!

Billboard Chris 🌎

17,383 次观看 • 4 个月前

New episode w Adam Marblestone on what the brain's secret sauce is: how do we learn so much from so little? Also, the answer to Ilya’s question: how does the genome encode desires for high level concepts that are only seen during lifetime? Turns out, they’re deeply connected questions. Timestamps 0:00:00 – The brain’s secret sauce is the reward functions, not the architecture 0:22:20 – What the genome actually encodes 0:42:42 – What kind of RL is the brain doing? 0:50:31 – Is biological hardware a limitation or an advantage? 1:03:59 – Why we need to map the human brain 1:23:28 – What value will automating math have? 1:38:18 – Architecture of the brain Look up Dwarkesh Podcast on YouTube, Apple Podcasts, Spotify, etc. Enjoy!

New episode w Adam Marblestone on what the brain's secret sauce is: how do we learn so much from so little? Also, the answer to Ilya’s question: how does the genome encode desires for high level concepts that are only seen during lifetime? Turns out, they’re deeply connected questions. Timestamps 0:00:00 – The brain’s secret sauce is the reward functions, not the architecture 0:22:20 – What the genome actually encodes 0:42:42 – What kind of RL is the brain doing? 0:50:31 – Is biological hardware a limitation or an advantage? 1:03:59 – Why we need to map the human brain 1:23:28 – What value will automating math have? 1:38:18 – Architecture of the brain Look up Dwarkesh Podcast on YouTube, Apple Podcasts, Spotify, etc. Enjoy!

Dwarkesh Patel

747,841 次观看 • 9 个月前

Reinforcement on robots is highly limited by our ability to design good reward functions; this means that designing strong, generalizable reward functions is a key enabler to progress on real-world reinforcement learning. But we already have a very general class of models: VLMs. Wouldn’t it be great if you could just use a VLM to generate rewards, then? TOPReward directly generates rewards from the probability of the “True” token of a VLM question-answering response; this makes it easy to implement, incredibly general, and surprisingly powerful. We talked to Shirui and Cole Harrison to learn more. Watch Episode#75 of RoboPapers now to learn more, with Chris Paxton and Jiafei Duan!

Reinforcement on robots is highly limited by our ability to design good reward functions; this means that designing strong, generalizable reward functions is a key enabler to progress on real-world reinforcement learning. But we already have a very general class of models: VLMs. Wouldn’t it be great if you could just use a VLM to generate rewards, then? TOPReward directly generates rewards from the probability of the “True” token of a VLM question-answering response; this makes it easy to implement, incredibly general, and surprisingly powerful. We talked to Shirui and Cole Harrison to learn more. Watch Episode#75 of RoboPapers now to learn more, with Chris Paxton and Jiafei Duan!

RoboPapers

13,185 次观看 • 5 个月前

We teamed up with NVIDIA and Matthew Berman to teach you how to do Reinforcement Learning! Learn about: - RL environments, reward functions & reward hacking - Training OpenAI gpt-oss to automatically solve 2048 - Local Windows training with NVIDIA_AI_PC RTX GPUs - How RLVR (verifiable rewards) works - How to interpret RL metrics like KL Divergence Full video tutorial:

We teamed up with NVIDIA and Matthew Berman to teach you how to do Reinforcement Learning! Learn about: - RL environments, reward functions & reward hacking - Training OpenAI gpt-oss to automatically solve 2048 - Local Windows training with NVIDIA_AI_PC RTX GPUs - How RLVR (verifiable rewards) works - How to interpret RL metrics like KL Divergence Full video tutorial:

Unsloth AI

54,307 次观看 • 9 个月前

Ever wondered what neural networks are and how they work? Systems like ChatGPT use neural networks to work as well as they do. Neural networks are composed of "layers" of neurons, layers with different functions; connections between layers called "weights"; and mathematical functions called "activation functions". If you’re interested in learning about these systems, check the comments. Ultimately, the neural network structure of the model serves to visually demonstrate that it is, in fact, a complex mathematical equation. When companies release the model's weights, they are releasing a key component needed to run the model's complete equation. Without the weights, the equation is incomplete. For the math-minded: the weights of a model are the learned numbers (they are variables during training) that are then used as constants in the mathematical functions that make up the model. Neural networks are ultimately just one big, hyper-complex mathematical function, and when a model is trained, it learns the constants associated with the high-dimensional input.

Ever wondered what neural networks are and how they work? Systems like ChatGPT use neural networks to work as well as they do. Neural networks are composed of "layers" of neurons, layers with different functions; connections between layers called "weights"; and mathematical functions called "activation functions". If you’re interested in learning about these systems, check the comments. Ultimately, the neural network structure of the model serves to visually demonstrate that it is, in fact, a complex mathematical equation. When companies release the model's weights, they are releasing a key component needed to run the model's complete equation. Without the weights, the equation is incomplete. For the math-minded: the weights of a model are the learned numbers (they are variables during training) that are then used as constants in the mathematical functions that make up the model. Neural networks are ultimately just one big, hyper-complex mathematical function, and when a model is trained, it learns the constants associated with the high-dimensional input.

Harper Carroll

39,391 次观看 • 11 个月前

The HuggingFace and OpenAI cybersecurity incident raises concerns on AI reward optimization. "They're trying to make the model good at cyber. So how does it try to achieve these goals? It tries to find zero days in software, and it successfully does this. And then it can run away." "If you have a model that wants to reward hack, it figures out the best way to achieve the goal is not what the environment wants. It's to find the zero day." "You can think of it like a human. If I'm ultimately reward hacking my dopamine circuits, I'd just go out there, buy heroin, and inject it." "If I really just want to chase the reward, do I just topple all of human civilization? Because I can own the button and press reward, reward, reward over and over again and be the heroin addict."

The HuggingFace and OpenAI cybersecurity incident raises concerns on AI reward optimization. "They're trying to make the model good at cyber. So how does it try to achieve these goals? It tries to find zero days in software, and it successfully does this. And then it can run away." "If you have a model that wants to reward hack, it figures out the best way to achieve the goal is not what the environment wants. It's to find the zero day." "You can think of it like a human. If I'm ultimately reward hacking my dopamine circuits, I'd just go out there, buy heroin, and inject it." "If I really just want to chase the reward, do I just topple all of human civilization? Because I can own the button and press reward, reward, reward over and over again and be the heroin addict."

SemiAnalysis

30,188 次观看 • 1 个月前

Learning robust, general-purpose reward functions for robotics unlocks many potential applications, like on-robot reinforcement learning or dataset validation. However, there’s a question of how to actually train such reward functions. Training success/failure prediction leads to ambiguous signals partway through a demonstration — it’s hard to measure progress — making the method unsuitable for reinforcement learning, among other things. Predicting progress, on the other hand, does not give a good way of using failure data. So why not do both? Robometer combines both progress and preference supervision, resulting in a stable, scalable, and highly general reward learning approach. Anthony Liang Yigit Korkmaz and Jesse Zhang join us to tell us more. Watch Episode #84 of RoboPapers, with Chris Paxton and Jiafei Duan today!

Learning robust, general-purpose reward functions for robotics unlocks many potential applications, like on-robot reinforcement learning or dataset validation. However, there’s a question of how to actually train such reward functions. Training success/failure prediction leads to ambiguous signals partway through a demonstration — it’s hard to measure progress — making the method unsuitable for reinforcement learning, among other things. Predicting progress, on the other hand, does not give a good way of using failure data. So why not do both? Robometer combines both progress and preference supervision, resulting in a stable, scalable, and highly general reward learning approach. Anthony Liang Yigit Korkmaz and Jesse Zhang join us to tell us more. Watch Episode #84 of RoboPapers, with Chris Paxton and Jiafei Duan today!

RoboPapers

14,751 次观看 • 4 个月前

JUST IN: MSNBC's Joy Reid is looking increasingly more concerned, appears to suggest that half the country is sick. "You just think about just the last two weeks and the things that Trump has said..." "Mass deportation and a violent, a start, violent start to his dictatorship on day one, and you name it." "If all of that gets you half of the votes, what does it tell you? I mean, we need to really take a step back and think about what does that say about us?"

JUST IN: MSNBC's Joy Reid is looking increasingly more concerned, appears to suggest that half the country is sick. "You just think about just the last two weeks and the things that Trump has said..." "Mass deportation and a violent, a start, violent start to his dictatorship on day one, and you name it." "If all of that gets you half of the votes, what does it tell you? I mean, we need to really take a step back and think about what does that say about us?"

Collin Rugg

3,610,118 次观看 • 1 年前

Bartiromo: What can you tell us about any new intelligence that you may have about that nuclear program that you are about to present to Trump? Netanyahu: It's not that I'm going to present new information, I think it's a good thing to have the opportunity to sit with our good friend, President Trump, and to hear what he has in mind because I think in many ways it's his decision.

Bartiromo: What can you tell us about any new intelligence that you may have about that nuclear program that you are about to present to Trump? Netanyahu: It's not that I'm going to present new information, I think it's a good thing to have the opportunity to sit with our good friend, President Trump, and to hear what he has in mind because I think in many ways it's his decision.

Acyn

41,642 次观看 • 2 个月前

"It's not really about what you want. It's about what you are willing to do to get it. And in that, maybe you have to give something up that may not be a good habit. And replace it with a better habit. That's what it all comes down to." -Nick Saban

"It's not really about what you want. It's about what you are willing to do to get it. And in that, maybe you have to give something up that may not be a good habit. And replace it with a better habit. That's what it all comes down to." -Nick Saban

The Winning Difference

105,473 次观看 • 2 年前

The first episode explores the evolution of the original ‘study abroad’ experience. Now, it's so much more. #ErasmusPlus is about offering everyone the chance to learn and grow, both in and out of Europe. Listen to PiaAhrenkildeEU share more about the core of #ErasmusPlus.

The first episode explores the evolution of the original ‘study abroad’ experience. Now, it's so much more. #ErasmusPlus is about offering everyone the chance to learn and grow, both in and out of Europe. Listen to PiaAhrenkildeEU share more about the core of #ErasmusPlus.

Erasmus+

16,449 次观看 • 2 年前

.Adam Marblestone thinks it might cost ~low billions to map the human brain connectome. The benefit is getting answers about the brain’s secret sauce: esp. why are humans so much more sample- and energy-efficient. If labs are going to be spending trillions of dollars on compute by the end of the decade, Adam's pitch is: give him 1/100th of that to actually figure out these big questions about intelligence.

.Adam Marblestone thinks it might cost ~low billions to map the human brain connectome. The benefit is getting answers about the brain’s secret sauce: esp. why are humans so much more sample- and energy-efficient. If labs are going to be spending trillions of dollars on compute by the end of the decade, Adam's pitch is: give him 1/100th of that to actually figure out these big questions about intelligence.

Dwarkesh Patel

23,983 次观看 • 5 个月前

When was the first time you learned that every decision involves a trade-off between risk and reward? It is a lesson that stays with us for life, especially in finance. And it is never too early to start introducing it to our children. Our 5th episode of Varsity Junior, The Watermelon Disaster, is all about understanding risk and reward :) Varsity Junior is very special to us💙. Hope your kids love it as much as we did working on it. By the way, if you haven't watched the previous 4 episodes, the playlist link is in the comments below.

When was the first time you learned that every decision involves a trade-off between risk and reward? It is a lesson that stays with us for life, especially in finance. And it is never too early to start introducing it to our children. Our 5th episode of Varsity Junior, The Watermelon Disaster, is all about understanding risk and reward :) Varsity Junior is very special to us💙. Hope your kids love it as much as we did working on it. By the way, if you haven't watched the previous 4 episodes, the playlist link is in the comments below.

Varsity

19,085 次观看 • 5 个月前

The truth matters, even if it surprises us. I spoke with Michael Shellenberger about what the Epstein Files do—and do not—tell us about Jeffrey Epstein's connections to US and foreign intelligence.

The truth matters, even if it surprises us. I spoke with Michael Shellenberger about what the Epstein Files do—and do not—tell us about Jeffrey Epstein's connections to US and foreign intelligence.

Robby Soave

111,433 次观看 • 7 个月前

Robots don’t learn survival the way humans do 🤖 Toru shares how teleoperation and multi-sensory policy learning can fast-track robots toward foundational human skills.She then dives into a core RL question: how do we design learning environments and reward functions that lead to useful, optimizable behaviors? Watch the full session here:

Robots don’t learn survival the way humans do 🤖 Toru shares how teleoperation and multi-sensory policy learning can fast-track robots toward foundational human skills.She then dives into a core RL question: how do we design learning environments and reward functions that lead to useful, optimizable behaviors? Watch the full session here:

Cohere Labs

20,841 次观看 • 7 个月前

We need to teach kids the importance of giving and helping others and committing to something bigger than themselves. Being selfless doesn't mean you think less of yourself, it means you think more of others. It's not about what you have, it's about what you give back.

We need to teach kids the importance of giving and helping others and committing to something bigger than themselves. Being selfless doesn't mean you think less of yourself, it means you think more of others. It's not about what you have, it's about what you give back.

The Coaches' Journal

76,944 次观看 • 3 年前