Загрузка видео...
Не удалось загрузить видео
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... show more
14,826 просмотров • 2 месяцев назад •via X (Twitter)
Комментарии: 0
Нет доступных комментариев
Здесь появятся комментарии из оригинального поста
![Gabbit Robotics co-founder Nicole Mattero on why the next AI winners won't fight over your screen, they'll fight to get closer to you: She kicks things off with what she admits is a hot take: "My hot take is that hardware is the only moat going forward because anyone's going to be able to spin up a model and make it really good and really cheap." Her logic is simple. Once a strong model is cheap and easy for anyone to build, the model alone can't set you apart. So what actually protects you? "Where does that leave you? Basically companies that have tactile modes such as hardware are the only ones that actually survive." Hardware, though, is only one piece of her argument. The second piece is data: "The models that win in the long run are the ones where private data is their [moat]." From there, Mattero ties the two together. Hardware gives a company one moat. The ability to gather unique personal data gives it another. Put both in place and the edge starts to multiply: "When you have a company that has hardware as a [moat], but then B is able to collect really unique personal data, that also becomes a [moat]." And that, she says, is the reason so many foundation model companies are reaching beyond the screen and into the physical world: "That's why I think a lot of these foundation models are like, let's get something in the home or on the person. It gives them an edge on both fronts."](https://image.24vids.com/tw-2061832779982328007/amplify_video_thumb/2061832727125782528/img/zf-LabD_ogMlwimO.jpg)
![Catherine Austin Fitts argues that Donald Trump was installed by bankers to construct a digital control grid comprised of programmable money, digital ID, and a social credit system. She maintains that Trump is expendable now that the infrastructure for this system is complete. "So let's look at the control grid. There are three general baskets of what you need to do. The first is programmable money. The second is digital ID... And then the third is you need the hardware and the software infrastructure to do the social credit system." "And so you need all the data centres... But you also need a surveillance infrastructure and enforcement infrastructure to back it." "So how do you persuade conservatives... to embrace building that infrastructure? The way you do it is you say, oh, we have election fraud, or oh, we have an immigration problem. And so we need a digital ID to identify who everybody is so we can't have fraud." "If you look at everything [Trump has] done to build the control grid, or to put the conditions in place to do it, they now have what they need." "Trump is basically expendable now."](https://image.24vids.com/tw-2044699385519386898/amplify_video_thumb/2044699307027185664/img/hyVp_n0t-nVrPCBk.jpg)
![The most interesting part for me is where Andrej Karpathy describes why LLMs aren't able to learn like humans. As you would expect, he comes up with a wonderfully evocative phrase to describe RL: “sucking supervision bits through a straw.” A single end reward gets broadcast across every token in a successful trajectory, upweighting even wrong or irrelevant turns that lead to the right answer. > “Humans don't use reinforcement learning, as I've said before. I think they do something different. Reinforcement learning is a lot worse than the average person thinks. Reinforcement learning is terrible. It just so happens that everything that we had before is much worse.” So what do humans do instead? > “The book I’m reading is a set of prompts for me to do synthetic data generation. It's by manipulating that information that you actually gain that knowledge. We have no equivalent of that with LLMs; they don't really do that.” > “I'd love to see during pretraining some kind of a stage where the model thinks through the material and tries to reconcile it with what it already knows. There's no equivalent of any of this. This is all research.” Why can’t we just add this training to LLMs today? > “There are very subtle, hard to understand reasons why it's not trivial. If I just give synthetic generation of the model thinking about a book, you look at it and you're like, 'This looks great. Why can't I train on it?' You could try, but the model will actually get much worse if you continue trying.” > “Say we have a chapter of a book and I ask an LLM to think about it. It will give you something that looks very reasonable. But if I ask it 10 times, you'll notice that all of them are the same.” > “You're not getting the richness and the diversity and the entropy from these models as you would get from humans. How do you get synthetic data generation to work despite the collapse and while maintaining the entropy? It is a research problem.” How do humans get around model collapse? > “These analogies are surprisingly good. Humans collapse during the course of their lives. Children haven't overfit yet. They will say stuff that will shock you. Because they're not yet collapsed. But we [adults] are collapsed. We end up revisiting the same thoughts, we end up saying more and more of the same stuff, the learning rates go down, the collapse continues to get worse, and then everything deteriorates.” In fact, there’s an interesting paper arguing that dreaming evolved to assist generalization, and resist overfitting to daily learning - look up The Overfitted Brain by Erik Hoel. I asked Karpathy: Isn’t it interesting that humans learn best at a part of their lives (childhood) whose actual details they completely forget, adults still learn really well but have terrible memory about the particulars of the things they read or watch, and LLMs can memorize arbitrary details about text that no human could but are currently pretty bad at generalization? > “[Fallible human memory] is a feature, not a bug, because it forces you to only learn the generalizable components. LLMs are distracted by all the memory that they have of the pre-trained documents. That's why when I talk about the cognitive core, I actually want to remove the memory. I'd love to have them have less memory so that they have to look things up and they only maintain the algorithms for thought, and the idea of an experiment, and all this cognitive glue for acting.”](https://image.24vids.com/tw-1979259041013731752/amplify_video_thumb/1979257571786543104/img/OqUh56yit9gUQWnb.jpg)