Okay, wait, hold on a second. I just need to make sure I heard you read the pre -show. You said Sam Altman's kids were supposed to run the Mars colony. Yes, that is exactly what the leaked email said. It's part of the whole lawsuit discovery. Altman proposed this $80 billion project for a Mars colony. And yeah, the text implies that... the governance structure that involves his children. Your biological children or his AGI children. You know, knowing that whole circle of tech philosophy,
genuinely hard to tell. Maybe they see them as the same thing eventually. But here's the real kicker. Elon Musk's reply to this whole pitch was just, it was brutal. I can only imagine. He just replied with a single statistic. He called the probability of success zero percent. Zero. Zero. Okay. Welcome back to the Deem Time. I'm really glad you're here with us. We usually spend our time trying to filter out all the noise, you know, find the signal in these huge data
dumps. But today. I mean, the noise is incredibly loud. It is loud. But if you listen really closely, there's a very distinct rhythm underneath it all. Right. And we aren't just talking about the usual chatbot headlines today. We're looking at a fundamental shift in competence. We're talking about AI putting on a lab coat at Stanford. Yeah. To cure diseases. We're talking about AI literally breaking out of the screen. To what? Fix your
dishwasher. And maybe most significantly, we're talking about a massive geopolitical shift in hardware. Something coming out of China that just changes the entire board. It feels like the theme for this dive is doing work. We're moving past the era of, you know, generating interesting text and into the era of executing complex tasks from the biology lab to the factory floor and then all the way to the global supply chain. There's so much to unpack here. Let's
start with the scientists. Let's do it. So we have this really fascinating study from Anthropic. It's titled The AI Frontier, Scientific Research and Physical Evolution. And I think for a lot of people listening, this is going to require a real mental shift about what an LLM actually is. Ugh, a huge shift. Because usually when we talk about large language models, you know, Claude, GPT, whatever, we think of them as writing assistants. Like, Claude, summarize this PDF. Or, hey, write
me a polite email to my boss. That kind of thing. Exactly. It's a retrieval tool, a formatting tool. It's basically a sophisticated autocomplete. Right. But what Anthropic did was profile three different labs, and these are serious, high -level research labs, where they're using Claude to do work that... That normally takes human scientists months, and the AI is doing it in minutes. And it's not just doing the grunt work. It's actually doing the thinking. I mean, take Stanford's Biomne
project as the prime example. Biomne is fascinating. They call this a research agent. Okay, before we go further, we should probably clarify that term. When we say agent here, we're not just talking about a chatbot with a personality, are we? No, definitely not. A research agent is a system that doesn't just chat. It uses tools to execute tasks. Think of it this way. A chatbot is a brain in a jar. It can talk to you. An agent is a brain that's been given digital hands that
can open files, run code, search databases. It can execute a whole workflow on its own. So it has agency. It can actually interact with its environment. Exactly. And Biomni is working across 25 different scientific fields. It's building this memory bank of skills. It learns expert workflows so it doesn't have to start from scratch every single time. See, that's the key differentiator. It's not just retrieving data. It's figuring
out. how to do the science. Yeah. But the example that really stood out to me, the one that really made me pause, was from MIT. Oh, the Matsuram Project. Such a great name, by the way. Scientists love a good pun. They really do. So they were using it for CRISPR screens. For anyone listening who might not be a biologist, CRISPR is basically gene editing. And these screens, they're usually a brute force process, right? Oh, incredibly
brute force. You basically have to disable thousands of genes one by one just to see what breaks. It's tedious. It's expensive. And it's so prone to human error because you're just staring at these massive data sets for weeks. And this is where the story gets a little wild. Because Claude didn't just speed up the spreadsheet work. No. This is the part that just stopped me in my tracks.
I mean, whoa. Yeah. Get this. Claude actually analyzed the data and it spotted an RNA modification pathway that the human scientists had completely missed. Wait, hold on. Let's pause right there. It found a biological reality that the human experts overlooked. Yes. When they benchmarked it, you know, against other models and against the human analysis, Claude identified a pathway that wasn't obvious. It saw a pattern in the noise that the humans just didn't see. That feels
like a pivotal moment. It's one thing to say AI can calculate faster than me. We're used to that. It's a totally different thing to say the AI noticed something about biology that I didn't. It completely changes the value proposition. It moves from assistant to collaborator. And there's a third example, right? It touched on the economics of all this. The Lindbergh lab at Stanford. Right. This is all about cost. Gene screens are incredibly expensive. We're talking
something like $20 ,000 per screen. 20 grand. Yeah. So you can't just test everything. You have to be picky. You have to place your bets very, very carefully. Usually, scientists use a mix of spreadsheets, some intuition, and existing research to guess which genes to target. It's an educated guess, but at $20 ,000 a pop, it's still a big gamble. So how did the agent change that gamble? Well, they had Claude navigate a
molecule relationship map. So instead of just looking at lists, the AI analyzed this whole network of connections between molecules, and it suggested the targets based on those relationships. So it's replacing intuition with network analysis. Precisely. And the study emphasizes this point over and over again. None of these breakthroughs came from just chatting with the bot. Right.
That goes back to your agent definition. Yeah, the breakthrough is connecting Claude to tools, giving it a custom workflow, adding guardrails so it doesn't just hallucinate and make things up. It's like the difference between asking a carpenter to describe a table and giving a carpenter a hammer and wood and saying, build it. That's a perfect analogy. So this brings up the big question for me here. If AI can cut the time from months to minutes and find things that we
miss, does this make science cheaper? Or does it just mean we start asking much, much harder questions? It removes the grunt work so brains can focus on the breakthrough. So we just level up the difficulty because we have a smarter teammate. We have to. All the easy problems are already solved. All right. Let's move from the sterile lab environment to something a little messier. The kitchen, the garage, the factory floor. The real world. Exactly. We're talking about physical
AI. Yeah. Now, I have to admit, when I first saw that term in the source material, my marketing alarm bells went off. Physical AI just sounds like a buzzword somebody made up to sell more chips. It definitely has that ring to it. I get it. But if you look at the actual capital expenditure, the money is telling a different story. The source material is adamant about this. 2026 won't be on screens. That is a very bold claim considering how much of our lives are on screens. It is.
But look at the investment stack. We're looking at a $123 billion ecosystem. Yeah. NVIDIA, Tesla, the entire robotics supply chain, they are all racing towards this. So when they say physical AI, what are they actually building? It's really about embodiment. For the last few years, AI has been a brain in a jar, right? It's a server farm in Virginia. It outputs text or images. Physical AI is about giving that brain a body. We're putting that brain inside the bodies we
already have. You know, our cars, our appliances. Right. The source material had the slightly hilarious example of a future where your fridge is negotiating with your dishwasher. I saw that. Yeah. And honestly, it sounds like a bad cartoon. Why do I need my appliances to gossip with each other? I know, it sounds ridiculous. But think about the internet of things. We've been hearing about IoT for a decade. Your toaster will talk to the internet.
And for 10 years, all that really meant was that your toaster needed a firmware update and had a security vulnerability. It was just dumb connectivity. Exactly. Physical AI is about competence. It's not just connected. It has agency. It's about the fridge, analyzing what's inside, realizing you're out of milk, checking the dishwasher cycle to coordinate energy use to the cheapest time of day, and maybe ordering groceries based on
your diet plan. So it's the fulfillment of the smart home promise, but like actually smart this time. And it's about the supply chain. If you're a consumer or a creator, you are now part of this loop. This isn't just about cool robots doing backflips. It's about the entire physical world getting a new operating system. That is both exciting and, frankly, a little terrifying. I'm just picturing my fridge judging my late -night snack choices. Oh, it absolutely will.
I'm sorry, Dave. I can't open the door. You've exceeded your calorie limit for the day. See, that's where I pull the plug. But a serious question, then. Yeah. Are we ready for our appliances to have agency? Or do we just want better automation? We want them to do the dishes, not argue about the soap. Right. We want the labor, not the debate. Exactly. Okay. Let's pivot from the domestic to the geopolitical. Because while the U .S.
is focusing on robots, there's this massive story coming out of China about the chips that run them. Hardware wars. We have a story here about Zippo AI launching a new model. It's called GLM Image. And on the surface, okay, it's an image generator, like Midjourney or DALI. Right. It's a 16 billion parameter model. Decent size, pretty capable. But the specs of the model aren't the real headline here. The headline is the silicon
it was trained on. Exactly. Usually when you read a paper like this, you scroll down to the hardware section and it always says, you know, trained on NVIDIA H100 clusters. Industry standard. But this one, no NVIDIA, no AMD, no US tech at all. This was trained entirely on Huawei hardware. That is a huge, huge signal. We've been talking about the chip bans for years. The U .S. government placed these really strict restrictions to try and limit China's access to high -end AI training
chips. And the prevailing theory in Washington was that this would, you know, cripple their ability to train foundation models. It was supposed to be a chokehold. But Zipu AI just proved that that chokehold might be slipping. They build a full domestic stack. This is what China has been chasing for years. Independence. Let's look under the hood for a second, because the way they did this is technically pretty interesting. It's not just a copy paste of a U .S. model.
No, the architecture is quite clever, actually. They use a two stage system. OK. Bring that down for us. So stage one uses what's called an autoregressive transformer. All right, jargon alert. Let's define autoregressive transformer for everyone listening. Yeah, the simplest definition is it's a model that predicts the next piece of data in a sequence. Kind of like how chat GPT predicts the next word in a sentence. Exactly like that. But here it's predicting the layout of an image. It's kind
of acting like an architect. So it creates the blueprint first. Right. It predicts what they call semantic VQ tokens. Basically, it figures out the meaning and the structure of the image first. That's a nine billion parameter model doing the planning. And then stage two. Stage two is a diffusion transformer. This is the painter. It takes that blueprint and it renders the final pixels. So logic first, then aesthetics. Yes.
Because of this split approach, it's really good at text -heavy images, which has been a weak point for some other models. Now, the source material does mention that in terms of pure artistic vibes, it might not beat the absolute top -tier Western models yet. You know, models like Nano Banana or Seadream. Maybe not yet, no. But it proves the concept. It proves you can train a massive functional commercial -grade model without
ever touching American silicon. So if the hardware bands were meant to stop them, did they just force them to accelerate their own development? Necessity is the mother of invention. We force them to build their own. And now that stack exists. And once it exists, you can't un -invent it. The divergence is real now. We basically have two separate hardware ecosystems developing in parallel. Which complicates, well, everything. All right, let's take a quick breath. That was
a lot of heavy lifting. Science agents, robotic supply chains, and semiconductor geopolitics. The trifecta of modern anxiety. Let's move to our rapid fire segment. Culture, chaos, and what we're calling cursed phrases. Oh, I love this list. First up, Leo Tolstoy. Poor Tolstoy. Can't even rest in peace. People on X, formerly Twitter, are accusing Leo Tolstoy of using AI. Because of the em dashes. Apparently using em dashes, that long dash you use for a break in a sentence,
is now considered a sign of AI writing. Never mind that War and Peace was published in the 1860s. Right. Or that, you know, he died in 1910. The paranoia is real. People are seeing ghosts in the machine everywhere. It's becoming a witch hunt. If you use good grammar or complex punctuation, you must be a bot. Which leads us right to the next item, the cursed phrases list. Ah, yes. There's a new list of words circulating that supposedly out you as an AI. Words that the models
are statistically more likely to use. I actually checked this list before we started recording. And did you pass? I have to be honest. Yeah. I failed miserably. I used Delve and Tapestry all the time. You're a bot, confirmed. I know. I felt so called out. But it's a real problem for writers now. I actually still wrestle with this. I find myself pruning my own vocabulary, deleting words I love just because I'm afraid someone will think I use chat GPT to write it.
It's this weird feedback loop of anxiety. It's a crisis of authenticity. We're stripping down our language just to prove we're human, which is so ironic. Item three. We touch on this in the intro. Elon versus Sam. The lawsuit that keeps on giving. We mentioned the Mars colony email, but there was another detail that really stuck out to me. Elon Musk. in that correspondence called OpenAI 0 % likely to succeed before he left. That is such a definitive statement, not
unlikely, zero. It really just highlights the split in vision, doesn't it? You have Altman with this expansive, almost hallucinogenic optimism, and then Musk with this brutal probability -based skepticism. And yet, strangely, they're both pushing the frontier harder than anyone else. True. Item four. Energy. The elephant in the room. The Trump administration is reportedly asking tech giants for $15 billion. For power plants. Power plants they might never use. They
want an upfront commitment. And the tech companies aren't thrilled. They say they weren't even consulted on the price tag. But it just shows the scale of the infrastructure problem. We're not talking about coding anymore. We're talking about AI consuming electricity on a nation state level. The constraints aren't software anymore. It's physics. It's electrons. And finally, the bear case. Michael Burry. The big short guy himself.
He's quoting Warren Buffett, and he's predicting the AI stock boom, specifically NVIDIA, is going to end badly. He thinks it's a bubble. He thinks the hype has just completely outpaced the reality. So is Burry right, or is he just early again? He's betting against the momentum of the entire world. That's a bold move. Very bold. OK, we're back. Let's try to pull all of these threads together before we go. We covered a lot of ground today. We did. But I think there's a pretty clear
through line here. I agree. It starts with competence. That's Claude. It's not just chatting. It's doing real science at MIT and Stanford. It's discovering biology that we missed. Then you have embodiment, the whole physical layer. That $123 billion stack, the fridge, the car, the factory. AI is growing legs and entering the supply chain. And then divergence. China building a non -U .S. stack with Zipu and Huawei. The hardware world is splitting in two, and it's proving that bands might just
accelerate independence. And finally, hype versus reality. Mars colonies run by kids versus the reality of crumbling power grids and these $15 billion demands for electricity. It really feels like the play phase is over. Yeah, it's not a toy anymore. Now the hard work and the hardware begins. So here's my question for you, listening to this. We talked about research agents and physical AI. Are you using these tools to chat or are you using them to build workflows? That's
the shift you really need to make. Because if you're just chatting, you might be missing the whole point. Oh, and definitely check the show notes for that cursed phrases list. Save yourself. Don't be a tapestry person. Please don't. I want to leave you with one final thought. We spent a lot of time on the code and the chips today. But think about that $15 billion demand for power plants. It's a massive number. The ultimate constraint on AI isn't code anymore. It isn't even really
chips, as China just proved. It's electricity. The smartest model in the world is completely useless if you can't plug it in. Keep diving. See you next time.