EP 653 : Meta, Google DeepMind & xAI Power $1.8B AI Cell Quest & Grok Bot Model Routing

0:00 Welcome to episode 653 of AI Brief. We're Amin and Leah and today we are tracking a pretty massive phase shift. >> Oh yeah, massive is definitely the right word for it. >> Right. So just to give you all a quick preview, we're talking about AI basically pivoting from, you know, just solving math to literally simulating human biology, >> which is I mean that's straight out of science fiction. And we're also going to look at Elon Musk's Grockbot making a really surprising partnership with its fiercest rivals.

0:32 >> Yeah, that one definitely caught me off guard. Plus, we've got a massive shakeup happening in AI interface features and subscription value. >> So much happening today. >> Exactly. The whole mission of today's deep dive is really about watching AI break out of standard text boxes. Whether it's mapping the building blocks of human life, creating these interactive visual charts, or just acting as an autonomous agent, routing tasks across the internet, >> it's just operating completely independently of human micromanagement now, which is wild.

1:00 >> Totally. So, jumping right into today's new updates, let's talk about the biological frontier because this this Biohub moonshot is huge. >> Yeah. Mark Zuckerberg and Priscilla Chan's Biohub, they just expanded their virtual biology initiative from what was a $500 million effort, right, to a staggering $1.8 billion effort. >> $1.8 billion. That is just wow. And the goal here, right, is to build what they are calling a universal virtual cell. >> Exactly. It it's basically AI software that can accurately predict how a cell is going to react to drugs or environmental changes before any physical lab tests even happen.

1:39 >> Right. With the ultimate mission to just cure or prevent all disease, >> which you know, no big deal. >> Yeah. Casual Tuesday. But the heavy hitters involved in this are what's really crazy to me. You've got the US government jumping in. >> Yeah. The NIH is providing data sets from, I think, over $500 million in past funding, >> right? And then the energy department is committing another 500 million over the next 5 years. >> Plus you have Meta, Google DeepMind, and Isomorphic Labs chipping in 300 million.

2:07 >> So it's this massive alliance. But but there is a catch here, right? >> There is. Yeah. So the data from this partnership is actually going to remain private for a full year before they release it to the public. >> A year in the AI and pharma spaces. I mean, that's an eternity. >> It's a massive first mover advantage for sure. And the scale of this is just hard to wrap your head around. Alex Rives, the science head over at Biohub, he noted that today's data sets have maybe hundreds of millions of cells, >> right?

2:34 >> But to get accurate models, he says they need data on trillions of cells. >> Trillions. Okay. So we let me ask you this because it makes me think about predicting the weather, right? >> Oh yeah, good analogy. >> We use AI to simulate global weather patterns, which is huge. But doing this at a microscopic level inside the human body requires like an unimaginable leap in data. Why does biology require trillions of data points compared to say the text data we use for standard LLMs?

3:05 >> Well, think about it like this. Language models are basically playing a highly advanced game of autocomplete, right? >> Sure. >> They map probability across static human created rules like grammar. But a living cell isn't static. >> No, it's constantly changing. >> Exactly. It's like trying to map the traffic patterns of a city where the roads literally change shape depending on what cars are driving on them. >> Oh wow. >> You've got proteins folding in 3D, genes expressing differently based on chemical gradients.

3:32 >> Just absolute chaos. >> Yeah, biological chaos. So you need trillions of data points just to establish a baseline of how that system reacts when you introduce a foreign chemical. >> Which explains why the energy department is involved, right? Because they have those insane supercomputers usually used for like climate modeling or nuclear decay. >> Exactly. You need that level of computational physics to even process this. >> Man, that is just incredible. But okay, speaking of unexpected collaborations, let's pivot from biology to the AI agent wars because Elon Musk just made a move that I I genuinely did not see coming.

4:08 >> The Grockbot announcement. >> Yes. So Musk announced that Grockbot is now going to route tasks to rival models, specifically Anthropics Claude using the Opus 5.5 model, right? And Midjourney to get the user the best possible outcome. >> It's a huge shift. So basically for lower lift queries, it's going to use this lightning fast upcoming Gro 4.8. Yeah. But for the heavy lifting, XAI is effectively becoming an Anthropic customer, which is wild because Anthropic already rents compute on XAI's Colossus 1 supercomputer.

4:41 >> Right. So, the money is just flowing both ways now. >> Exactly. And it's clearly a strategic move to compete with rival agents, right? Like OpenAI's DOTS, which is powered by GPT6 Astra and Meta's Muse. But wait, okay, I have to point out the irony here because Musk has spent literally billions of dollars to make Grock a frontier model. >> Yeah, that Colossus cluster was not cheap. >> Exactly. So, is opening the door to Claude an admission of defeat or is this just like a brilliant aggregator play to stop people from jumping ship to open AI's dots?

5:12 >> I mean, it's definitely a bit of both. It's a defeat on the foundational model front for sure, but from a business survival standpoint, it's brilliant >> because users don't want to juggle three different subscriptions, >> right? Grock operates as this 247 assistant with its own computer. It becomes the single interface. You keep the user locked in the XAI ecosystem. Even if Enthropic is doing the actual cognitive heavy lifting, >> that makes a lot of sense. And you know, connecting that reliance on Enthropic directly into everything else in AI today, there's a lot happening with these specific labs.

5:46 >> Oh, totally. The pricing wars are heating up. Anthropic just launched Claude Haiku 5.5 and it is insanely cheap. >> How cheap are we talking? >> Just 10 cents per million input tokens. >> Wow. Okay. That matches OpenAI's GPT6 Luna price, right? >> Exactly. But it's actually beating Luna on multiple benchmarks. So, Anthropic gets a huge budget win there. Meanwhile, OpenAI is completely revolutionizing their UI. >> Yeah, ChatGpt is getting GPT6 for all users now, not just the paid ones, >> right? And they're pairing it with this new intelligent UI, which I watched some demos of this. It's crazy.

6:24 >> It's so cool. Interactive charts, graphics, maps. >> Yeah. Clickable buttons right inside the chat. You aren't just getting text back. you're getting like mini software applications generated on the fly. >> But that brings up a really interesting value debate because semi analysis just published research showing that cloud subscriptions offer about five times more value per dollar than chat GPT's everyday model plans. >> Five times. >> Five. >> That is a massive gap. >> Yeah. And you have to remember the context here. Open AAI just have the limits on its $200 tier last week.

6:57 >> Wait. Okay. That is such a paradox because how can budget models like Haiku 5.5 be getting so incredibly cheap while top tier AI is becoming so resourceheavy that OpenAI literally has to slash user limits on a $200 subscription. >> It's the compute overhead, >> right? >> Generating stateful web apps and interactive charts in real time requires a massive amount of server space compared to just streaming text. So the bottom of the market is pennies, but the top end is skyrocketing in cost, >> which naturally pushes people toward open source, right?

7:29 >> Exactly. Like Noose Research, they just raised $90 million at a $ 1.5 billion valuation. Wow. >> And they revealed that their open-source Hermes agent is already handling $2.5% of global token usage. >> That's a serious chunk of the global pie for an open source model. >> It really is. And as all this open- source and agentic stuff scales, detection is becoming a massive priority. >> Oh, for sure. Google just released their synth ID detector today.

7:57 >> Yeah, it finds invisible watermarks in media. And it's not just for Google stuff. It works for OpenAI, Nvidia, Cacao, and they said Apple is coming soon, too. >> That's going to be crucial. >> Oh, and Google also dropped Playground, right? >> Yes. >> A tool for creating custom video games without coding. Just totally natural language. I am definitely going to be losing some hours to that this weekend. >> Oh, same here. >> But okay, looking at today's trending AI tools, because models are gaining these intelligent visual UIs and acting as agents, how can you actually put this stuff to work today?

8:30 >> Well, velocity is a big one trending right now. >> Velocity. Yeah. >> It lets you set up a multiplayer AI office where human teams and AI agents actually collaborate, >> right? And within that, you can deploy Anna, which is their AI negotiation agent. >> Yes. Anna can literally handle purchasing your software for you. >> Okay, but push back on this. If you use Anna to negotiate software prices and the software vendor uses their own AI to negotiate back against Anna, don't we just end up in an infinite loop of AI hegling over pennies?

9:00 >> You'd think so, right? But it actually behaves more like highfrequency trading. They just exchange boundary constraints, run the math, and find the equilibrium in milliseconds. No endless emails. Okay, that is super efficient. Another tool I'm seeing everywhere today is Parakeet AI. >> Oh, I love the concept of this one, >> right? It acts as a live sparring partner to help you prep for job interviews. Simulates high pressure situations, real-time feedback. >> That's huge for career prep. And on the marketing side, there's Gen Page 3.0, >> which instantly builds and personalizes landing pages. Right.

9:34 >> Exactly. It uses dynamic UI generation to optimize conversion for every single visitor instantly. >> Wild. And if you're actually building these agents, 11 Labs just dropped architect. >> Yeah, it's an assistant specifically designed to help users build and fix their AI support agents, like an automated structural engineer for your workflows. >> Super necessary right now. And obviously, Grockbot and the synth ID detector are trending heavily today, too. Tying back into the news we covered earlier, >> everyone is testing that Grock routing architecture right now. Seriously. So, wrapping up today's deep dive, the scale of what we're looking at is just it's hard to process. We're going from micro tools like Anna negotiating our software budgets all the way up to macro efforts trying to simulate trillions of cells in the human body.

10:19 >> It's a fundamental shift in ambition. AI is moving out of the chat window and trying to manipulate physical reality. Which leaves you with a pretty crazy thought to ponder because if Biohub actually succeeds in creating a universal virtual cell over the next few years. >> Yeah. >> How long until AI can simulate an entire virtual human organ or even like a complete digital twin of a patient so doctors can test surgeries on a digital you before they ever happen?

10:48 >> It honestly feels like we are right on the edge of that. >> It really does. Well, if you found today's deep dive valuable, please make sure to follow, rate, and share the show. >> Yes, definitely share it with anyone trying to keep up with all this craziness.