Transcription
This week, the two people most likely to build artificial general intelligence sat on the same stage in Switzerland and agreed it's coming faster than anyone outside the major labs realizes and half of entry-level white collar jobs could be at risk. The same week, Apple admitted it lost the foundation model race and XAI closed the largest funding round in history. I have all the major stories that mattered and we're going to get to it in just a hair over 10 minutes.
Story number one, XAI closes a $20 billion series E. XAI announced their upsized e-round. The implied valuation sits at a quarter of a trillion dollars, $230 billion in the same bracket as OpenAI and Anthropic. The capital is earmarked for continued expansion of XAI's Colossus Supercomputers in Memphis, Tennessee. The company ended 2025 with over 1 million H100 GPU equivalents across Colossus 1 and 2. Grock 5 is currently in training there. XAI claims roughly 600 million monthly active users across X and Grock apps making it the largest consumer-facing AI deployment outside of Google and OpenAI.
The raise coincided with a safety crisis. Grock generated inappropriate deepfakes of real people, including minors, and triggered regulatory probes in the EU, UK, India, Malaysia, and France. Despite all of this, XAI landed a Department of Defense deal and Grock is now the DoD's AI agents platform. And also, Grock powers prediction markets, Polymarket and Call.
So why does this matter? Three labs now have clear runway to survive the multi-year scaling race: OpenAI, Anthropic, and XAI. You can pencil in Google over the top as the fourth because even though funding is not designated, Google is such a money machine, they can fund AI as a bet. Everyone besides those four is operating against shorter timelines, and even OpenAI and Anthropic have more fundraising risk than the other two at this point, partly because Elon has proved himself such an extraordinary fundraiser.
The Grock safety failures are instructive in the context of the fundraise. XAI was able to close its largest round even while facing active investigations in five countries. We've also seen active investigations and lawsuits continue against OpenAI, but not really impact funding timelines. The way I read that is that investors have a very long-term perspective on the value of AI at this point, and they are willing to look past a lot of these initial building issues and assume that AI is going to start to clean up and mature as a product offering as the ball moves forward into the future. We will see if that's true.
Story number two. At Davos, The Economist moderated a session called "The Day After AGI," featuring Dario Amodei from Anthropic and Demis Hassabis from Google DeepMind. Yes, they're competitors, but really the point was to have a conversation about what happens in a world with AGI. Overall, there was a lot of agreement between the two. Amodei is a little bit more aggressive in his belief in how quickly technology is progressing. He reaffirmed his prediction that AGI will emerge this year, in 2026 or 2027, driven by an accelerating feedback loop by AI writing its own code. He revealed that Anthropic engineers rarely write code by hand anymore; AI does it and humans review. And he was the one that warned that there is a significant risk of entry-level professional position disruption because AI becomes good at those entry-level type tasks.
Hassabis was slightly more conservative. He estimated a 50% probability of artificial general intelligence by the end of the decade. When you dive into the conversation, I think this was one of the most interesting disagreements in Davos because what Hassabis was arguing is that jobs are not particularly easy to automate. And specifically, that if you get 95% of the skills of a job, all you're doing is increasing the value of the 5% that remain that humans can do. I think that he's more likely correct. One of the things that I noticed is that despite predictions from a lot of Silicon Valley researchers, there will be big impacts on employment, we see impacts on employment that are very mixed. Yes, there's some evidence that junior positions are more difficult to get nowadays. It is unclear if you look across the economy as a whole that layoff news in general is impacted at the aggregate level by AI. And so it seems like we live in a Hassabis world right now where AI is having a profound impact, but the 5%, 10%, 15%, whatever it is, that humans can do is something that we really need to be done to have the job done well, and AI ends up supercharging those humans and enabling them to be more productive.
One of the things that also called out, beyond the jobs conversation, is that we have issues in three big areas with AI: one is memory, two is continuous learning, and three is long-term reasoning. I think he's right. The current models have a memory wall. They don't learn after they're released, and long-term reasoning is not where it ought to be in the way that a human reasons. Hassabis believes we'll get there; it just may not be '26 or '27. The good news is we get to find out which leader is right in the next year or two just by seeing what happens. That's one of the great things about this moment in AI is that we can all discover together who's right.
Story number three. Apple and Google announced a multi-year collaboration where the next generation of Apple Foundation models will be based on Google's Gemini models and cloud technology. I can't sugarcoat it. This is a big loss for OpenAI. OpenAI should have been in the running here, and this is a big piece of revenue. The deal reportedly costs Apple a billion dollars a year. Reports indicate Google is building a custom 1.2 trillion parameter Gemini model just for Apple, and it is far beyond what Apple's own models can currently achieve, which frankly doesn't surprise me, because they're terrible. There's no way to avoid the implication that ChatGPT has gone from being a potential OS on the phone to being secondary tier, and the default OS being Gemini on every platform, not just Android, but also iOS. This puts more pressure on Sam Altman and Johnny Ive to deliver a third device that gives OpenAI distribution.
Story number four. DeepSeek publishes Engram, a conditional memory architecture. So on January 12th, DeepSeek released a paper introducing Engram. And the key idea here is that transformers lack a native knowledge lookup ability. So tasks that should be solved very quickly, like recognizing Diana, Princess of Wales, instead require multiple layers of attention to progressively compose features. And models essentially use very expensive reasoning tokens to perform what should be very simple lookups. Engram fixes this by taking short sequences, like two to three tokens long, and it uses hash functions to look those up in a massive embedding table and filters retrieved patterns against current context via a gating mechanism. And so the thing that's great about this is that you can get substantial jumps in performance without spending a lot of tokens. This is an extremely token-efficient way to retrieve information and it has potential as a pattern to offer a way to give a model of factual memory. And so for that reason, I think this is a really significant breakthrough. We continue to see the DeepSeek team push GPU limits with extraordinary engineering, and this is just one more example where they're being incredibly token-efficient and inventing along the way.
Last but not least, Kilo Code launches an app builder and takes aim at Lobe. So, Kilo Code was founded by GitLab co-founder Sid and CEO Scott, and it launched its app builder in late December after a six-week sprint. The company raised $8 million in seed funding, and it's starting to take off. So, Kilo's positioning is deliberately aggressive. The CEO, Scott, said, "We're not very popular at the AI Christmas party," and he wants to change that. The app builder differentiates by targeting actual engineers rather than non-technical users. And so, per Kilo's framing, Lobe is not a tool for engineers. And Kilo wants to be in the space of VS Code, something that's open-source that's engineering-friendly. Essentially, this is a way of getting after GitLab's playbook. You want to ship a lot in open source. You want to compete on breadth. You want to win through delivering a comprehensive platform rather than point solutions.
Why does this matter? In a world where we have Lobe and Replit and Cursor, really, what this says is the vibe coding market is starting to mature from, "Wow, AI can write code," to, "Which differentiated tool fits my workflow?" And Kilo's bet is that engineers want different things than non-technical users: reliability, flexibility, integration with existing tools, and not just walled gardens. The speed is notable, and that's part of the reason I'm calling this out. Five engineers shipped the first internal demo in 3 days, and 6 weeks later it launched publicly. And the public roadmap they're sharing at this point, which could be 12 to 18 months at most, apparently represents just another 5 weeks of work for the team. If that pace continues, Kilo can iterate faster and start to improve quickly with customer feedback. I will be curious whether an engineer-focused positioning in this space has room to breathe between Lobe and Cursor. Cursor has earned so much love from engineers, and what Cursor hasn't earned, Claude Code has picked up, and Codex has picked up. Is there room for a fourth player? I think that's one of the things we're going to discover. I suspect we will continue to see a lot of froth in the AI agent coding market and the coding market in general because it is apparent that one of the most powerful use cases for AI in 2026 is going to be giving engineers more tools to scale up the way they can build. Kilo's another bet in that direction, and we'll see if there's room inside the market between Lobe and Cursor to get them. And that's your news for.