MultiTranscript

Everyone can be everything now. PMs can ship code, designers can write PRDs, engineers can product, and there's this confusion and frustration of what is my job anymore. >> Anytime a new technology comes along, you go through a storming phase before you go through the forming phase of things. We are in the middle of that right now. [music] I don't think that means we should put AI back into the box and say let's not use it. >> If we all become builders, will we still need separate functions? >> I still see a craft excellence that's really important that [music] I don't think is going away anytime soon. I still find great engineering to be scarce, great data science to be scarce, great creativity to be scarce. >> If you look at the early culture deck of Netflix, high agency, autonomy, paying top of market, this is what I hear constantly now from how the top AI labs operate. >> Netflix's culture has always been excellence as an operating system. It's a resistance [music] to do the thing that a lot of bigger companies would do and to feel comfortable in that discomfort very often. >> What are the ingredients to make this happen? >> Talent density is the non-negotiable, being very comfortable with risk-taking in cases where things are not going well, not assume that process is going to fix it. >> What have you added to the career ladders within this AI world? >> more systems thinkers, people who can look across all the business domains and abstract that [music] to here's the buil

For thousands of years, humans have built tools to make hard things easier. A hammer does not make you worse at building, but is AI different? Because AI helps your brain, not your hands. And that raises an uncomfortable question. What happens to the skills we outsource? Is AI brain rot real? What should we do? Well, I see a big difference in the media headlines and actual real-world studies. So, let's see an actual experiment with a group of 52 junior software engineers split into two groups. One that used AI for coding and one that didn't. I have two questions here and the results on both kind of surprised me. One, did AI make them faster? Well, the AI group finished their tasks a good 2 minutes, or about 8% faster. So, they had time to watch a 2-minute Papers episode, which is never 2 minutes. But then, wait a second, the result is not statistically significant. I'll downgrade this a bit. So, did AI make people code faster? Maybe. Now, second question, did the AI make them dumber? Now, hold on to your papers, fellow scholars, because the answer is kind of yes. After having them take a quiz, we get 50% for the AI people and 67% for the no AI people. Woof, that is almost a two-grade difference. Huge difference. And this is statistically significant. So, this one would be much harder to dismiss as noise. I got to say, this is a fairly large effect for such a smaller study. So, where is the largest gap? The AI group really underperforms when exactly? Well, in debugging. Yes, y

What would you see if your whole world was made out of numbers? Yes, like in the Matrix movie. A bat sees with sound. A reindeer shifts its vision into ultraviolet for the winter. But, an AI system does not have eyes. All it sees is a flood of numbers. So, how come it can understand your questions? How can it pass the bar exam? I mean, what is going on here? You see, today's AI systems are able to turn a bunch of input tokens into geometry. Yes, it quietly builds little curved shapes inside. They denote areas, concepts, and decision boundaries. And if we do this at a really large scale, these systems have properties that almost feel like magic. Getting a gold medal performance at the International Mathematical Olympiad, no problem. Now, check this out. Let's ask a simple question. Little AI, if I add the word aluminum here, does it fit the page or does it go over? Well, this simple task is surprisingly difficult for it. Why? Because the model gets no character counts and no page width. Yet, it still answers correctly. Wait, what? It does not have eyes, so how come it can judge that? Even the input tokens it gets do not come with character count information. So, before it can count characters, it has to learn to count by itself. Then, figure out how wide the page is and then subtract. Okay, okay. Now, I hear you asking, "Carolyn, who cares? Why does this problem matter?" Well, it matters because of two things. One, whenever we give an AI a task it has seen or done before, it u

Bad news for the tech community. >> Burnout at work is a major problem and AI could add to this already overwhelmed employee problem. >> The honeymoon period with AI is over. When people are asked, what what are you afraid of? Losing my job to AI is actually second to last. What we saw rise up to the top is the expectation to do more for the same pay. We did this survey on how people are feeling in tech right now. Burnout is [music] increasing significantly. Optimism is declining. It's never been crazier. When are we going to get to the optimistic part of this [laughter] uh of this episode? Half of the people in tech are feeling incredible. And the other half told us, "My brain is rotting. My work feels worse." Elon had this really interesting way of describing it. We're definitely living in a simulation if we're alive right now when we're about to start building data centers in space and AI is going to be as smart as human. What does this tell us we should be doing if we can do anything? We're in the second inning of a massive shift. No one knows how it will end, but all you can do is Today, my guest is Noam Seagull. Noam and I go a long ways back. We worked together at Airbnb for many years where he was my research partner. Since then, he went on to be a research leader at Intercom, Twitter, Wealthfront, Meta, Zapier, and Figma. Gnome is so amazing, and over the past couple years, I've been lucky to partner with him on a bunch of research projects. The most ambitious of whi

I was horrified as many were when Fable suddenly um got banned. >> We have obtained the mandate to build um the UK's first sort of sovereign LLM. >> How can you do in millions what they are doing with billions? >> There isn't a huge amount of room for error or a huge amount of wiggle room. It was not something that was on my bingo card in January. Numbers like 10 trill being knocked around. We we we're compressing most of the internet at that point. It's >> really funny. after Claude deletes your production database, um it it'll say, "Oh, you're right to point that out." >> Agentic harnesses are getting less important over time. A model can probably do with bash only basically any task these days. Also, thank you, Donald Trump. >> Yeah, I mean for the first time I feel like a secondass citizen because they are going faster than I am and I really hate that >> me that boils my blood more than anyone else. Um and I we are going to do everything we can to pull this off. We have no choice but to make it happen. Alistister, it's it's great to meet you, mate. We are here in London where it's customary to say hello, Giza. >> Hello, Geyser. >> So, where in London are we? >> We are in Hawkton right now. So, we're in Shortitch. Um, we're about half a mile away from where Cosign started in my apartment, which was in Hawkton Square, just over there. Um, so we haven't come very far, but we have expanded a fair bit since then. >> And what is Cosine? >> Cosign is a frontier lab based here in

Imagine you could hold an entire planet in your hands, a world with mountains that never existed, coastlines stretching past the horizon in every direction forever. A solo scientist invented something incredible here. One, it is a truly infinite terrain generator. This panel shows a region that is roughly the size of a country like Congo, millions of square miles of land. And if you would zoom in to one of these regions over and over, it just keeps going and going. Amazing. Now, many of the examples are shown within Minecraft, but it can generate fully continuous worlds depending on the game engine being used. But it gets better. Two, this generator also learns. More on why that is amazing in a moment. Now, wait, wait, wait. We already have terrain generator programs, a ton of them. So, is this really new? Well, there are existing methods that build terrain out of noise. These can generate infinitely in every direction, but they aren't that organic. They are a bit too uniform, too repetitive. It just generates new stuff with no plan for the whole planet, no large-scale coherence. So, these methods don't really learn. This is the price of speed. Or, there are AI-based methods. These can learn. Yes, you can feed them the statistical distribution of real terrain from Earth and have it generate something similar without copying. Ooh, that is amazing. But unfortunately, they are quite inefficient. Why? Well, because every newly generated area depends on every other area in the wor

What if an AI could see into its own future? Not perfectly, no, just a little. DeepSeek did something absolutely amazing in their new research work. You see, when you ask an AI how to rewrite a difficult email, it answers you one word or one token at a time. That can get pretty slow. So, why not put out multiple words, like a whole sentence, at the same time? That would be much faster. And that would be a game-changer because it would boost many slow AI systems into usability, maybe one right on your phone in your pocket. Well, yes, except that it doesn't work. A big, smart AI is like a senior editor, brilliant but expensive. If you ask it to write five words one by one, it still has to think five times, slow, expensive. So, what do you do instead? The solution is unexpectedly brilliant. What is that? Well, you hire a junior writer. The junior writer writes the next few words quickly. Of course, these words are not always accurate. So, you ask the senior editor to verify them. It looks and says, "Yes, yes, yes, no, no." And everything after the first no is thrown away. If the junior writer guessed well, we get our five words for cheap quickly. Hooray! We call this speculative decoding. But, not so fast. There is a problem. The writer is not that experienced. It messes up. For instance, it forgets things, it starts a phrase and finishes another. And I don't want this little writer to get fired. So, here is DeepSeek's incredible new work called DeepSpark with three new tricks t

Today, we are going to create crazy virtual worlds where we simulate squishy things. Yes, things that deform. This new method is incredibly impressive. It can do absolutely wonderful things. Elastic barbarian ships and rods, really squishy letters, super detailed cloth wrinkles, and then whatever the heck this is. Now, that's really tough, but I'll try to summarize 30 years of research for you. Here it goes. We have fast simulations that are wrong, or we have accurate ones that are painfully slow. Why? Well, the input is a bunch of shapes and forces that act upon them, and the output is the new positions of every tiny point that describes these objects. This is a really tough problem because these things are squishy. So, every tiny point affects every other point, and there are millions of them. So, one small mistake can spread everywhere and break the whole simulation. Okay, so what can this new technique do? Well, imagine a tree that you really don't like, and you hope to never see it again. So, what does a computer graphics researcher do? Well, of course, throw jack-o'-lanterns at it until there are enough so we never see it again. Can we do that? Let's see. Oh, yes. Kind of crazy. So, what is going on here? Dear fellow scholars, this is Two Minute Papers with Dr. Karoly Zsolnai-Feher. So, previous techniques had a problem with scenes like this. You see, they split the problem into small pieces and solve them separately. But, the problem is that those pieces ignore each ot

But we have the benefit of millions of years of evolution, right? And so it's almost a bit unfair that we're expecting AI algorithms to be able to do all of that. >> Exactly. >> I think it boils down to one of the big open questions in the field like is language critical to intelligence. >> I will say let's say at least when I'm playing the games myself and I think that goes for all of us. I'm using language. >> So my name is Jon Kotar. Uh, I have a background in physics and mathematics and I've been working in the industry for about 15 years with the last 5 years specializing more towards AI. It was immediate to him what the goal was. He recognized the pattern and I showed him the game. Within 3 seconds, he finished the first level. Often the agents start thinking that reducing the energy bar to the minimum is the goal or that stepping 10 times in a region is the goal, which for a human is kind of clear that that it's not the actual goal. Uh, my name is Stephano. Um, I study computer science, machine learning. I did some research in reinforcement learning at EPFL. >> I guess the million-dollar question though is, do you think it's possible in principle to do really well on RKI3 and be no closer to AGI? >> Uh, yes, I do think it is possible. >> I don't know. I guess the main idea is a bunch of uh bright people in the room and do good research together. The transformers can't plan but they they can do a very good job of pretending essentially that is in a sense indistinguishab

The US government essentially banned the use of fable anthropics frontier level AI system. And if this kind of capability is locked away from us, even from some of its own creators, we have to ask if any other AI model reaches that capability, will that get the banhammer too? So far, the answer seems to be yes. And even if it comes back for us, it might, but with an identity and a nationality verification system. So, is this the last time we laid our hands on a Frontier AI system? Well, I think I have an answer. I hope. You see, there are free and openweight AI models out there that we can download and run forever. Yes, something you can actually own. I know it sounds weird. Now, these open systems are typically behind what these trillion dollar companies can offer. It has been like this for a while now. But then a system called GLM 5.2 appeared. The headlines say this is a fable level system. Some benchmarks say it matches some Frontier models. As always, it depends. During my internal testing, I was very surprised myself. In most of my usage, it leaves all other open systems in the dust. It is insanely good. A huge jump forward. Now, for me, let's be measured here. It did not match the Frontier systems, but it came so close, way better than their 5.1 system in general knowledge, coding, math, fixing things in the terminal, you name it. And this is just a minor version number jump from 5.1. This in less than 3 months. That is insane. How on earth did they do that? Dear fello

Yeah. So, why not try and build a chip that's just inherently random? >> Meet Thomas Alley. I caught up with him in Zurich, and he's one of these rare galaxy brain people who's comfortable in probabilistic machine learning, formal verification, and chip design. However, there's a small problem. When the token god hands you something that looks like it works, how do you know it's actually right? So my background is in theoretical computer science. I used to do algorithms for highdimensional data locality sensitive hashing. Then I moved to normal computing to develop thermal computing uh also to speed up bashian intelligence. Sometimes I think about it as the lovable for chip design. So we take it all the way from your intent uh through the design through optimizing your design to formalizing and verifying your design uh all the way to tape out. Now I didn't fully appreciate this before. These days a chip doesn't necessarily start in a factory. It can start as code. So engineers design the whole circuit in a language called verilog almost written like software. And only much later does any of it become physical silicon. But first that code has to be simulated and formally verified. It has to be proven correct. Because once a chip is fabricated, if there are any bugs, you're in big trouble. So, a few months ago, Thomas blogged about building a Verarilog simulator using a swarm of AI agents collaborating with each other, and it generated over half a million lines of code in 43 da

I don't really love the bitter lesson as people try and apply it. In fact, AlphaFold 2 is the opposite of that. >> Protein folding is one of these holy grail type problems in biology. >> We predict nature level science with the press of a button in a very narrow category of nature level science of the structure of a specific protein. >> John Jumper led the team behind AlphaFold, the system that predicted 200 million protein structures. In 2024, he won the Nobel Prize for chemistry. And now, Jumper is leaving DeepMind. But what did AlphaFold solve? What remains unsolved? And could AlphaFold be the template for AI for science? >> We are not trying to tell you everything. We are not a model of the entire cell. You try it, you measure. Nine times out of 10, you find out you're wrong, right? If you're wrong nine times out of 10, you're a very successful machine learner. You're incredibly productive. >> So, for half a century, structural biology had a massive bottleneck. DNA was easy to read, but protein structures were not. A protein structure begins as a chain of amino acids. And then, often with help from the cell, they settle into a three-dimensional shape. And that shape determines what it binds, what chemistry it catalyzes, where it sits in the cell, and whether it even works at all. But from a machine learning perspective, if you only have the sequence, can you predict the fold? Can you predict the structure? >> We've discovered more about the world than any other civilizati

Scientists at Deep Seek have invented something amazing and exactly at the right time when we need it most. You see, we are entering the age of AI. But, I am really surprised. I just found out that the way these AI systems run on our computers is incredibly inefficient. So, if you want your AI assistant to answer quicker, you need more compute power, clear as day. But, you may find that as you add more compute, it does not get faster. But, how can that be? You know, it's kind of shocking given that companies are paying billions and billions of dollars for more compute to run these AI systems. How is this possible? Imagine reading a book and now imagine that every time you turn the page, you forget about the characters. That's not a great way to read books, right? Here is what happens in practice. Assume we have a huge brain the size of a mountain and we want to talk about a book. If the book is one page, we just memorize that one page and just talk about it, quick and easy. Now, imagine that the book grows. It is now huge and since we forget about everything the moment we turn the page, ouch. If we want to talk about it, we have to reread it all the time. So, our brain is huge and hungry, but there is a problem. Information is coming in through a straw. So then, we spend most of our time not thinking, but reading slowly. And that is exactly what the graphics cards of today are doing when you run an agentic AI system on hard problems. All those billions of dollars sitting at 4

The number of AI agents on the internet is increasing at such an insane rate. I don't think I've seen anything like this. This is crazy. And this is an area that is quite new, and the technology is still pretty rough. Improving rapidly, but pretty rough. And the promise of agents is incredible. It would book the cheapest plane ticket for you, or run 24 hours a day to manage your schedule, submit insurance claims, continuously scan a codebase for vulnerabilities and patch it. Well, this is the good, but at the same time, you get so many news headlines about spam, security issues, and system breakdowns. And it gets even tougher when you have not one agent, but multiple agents. Imagine two agents organizing a holiday for you. The flight agent hallucinates a cheaper airport 400 miles away from your real destination. Then, the hotel agent says, "Let's book something super cheap nearby." Well, super cheap is often non-refundable. And now, congratulations. You now have a non-refundable room you will never see. And so many of these problems come from the fact that agent coordination is super difficult. Now, check out what this paper says we should do. Here is a math problem. First agent writes a plan. The next one critiques it, and the third one solves the problem. And at this point, I said, "Okay. I see nothing interesting here. This is what everyone does with agents." Yes, but here's the key. Most agents communicate a bit like we do, in words. Wait a second. Why should we do that?

AI systems today are really powerful and can do a lot. No question about that. But, how do they really work? We have so many questions. Do they think like humans? How do they beat the best human chess player? How do they beat the world champion video game players? And how is it possible that an AI chooses to not play the game, but just collapse and can trick the brain of another AI to malfunction? Why does Claude think about blackmailing people? I mean, who what is going on here? If you look at the activations inside an AI system like Claude, you see a bunch of gibberish, millions of numbers. Researchers tried to make sense of it for years and years now, but the results were very thin and situational. We now see that it understands that if you look at an image and you have floppy ears, a black snout, and so on, then it might be a dog, a good boy. But, we asked a bunch of questions and still no answers to those. But, now Anthropic has excellent new research with new insights on this. This is when Anthropic is at its best, in my opinion. I love seeing it. Here's the idea. Take this bunch of numbers that the AI thinks about and ask another AI to translate it into text. Translate from machine to human. And it did something. Okay, but these systems often make stuff up. So, how do we know if this is a good translation? We don't. So, what do we do here? Try it separately with a bunch of different models and see if they translated the same way. Is that a good idea? Mm, not quite. Ima