Two researchers just closed a math problem that had been open since 1995, with AI agents doing much of the legwork. And they ran a control: the same token budget, several leading models, nobody steering. Those runs never got there. So the humans keep their jobs. This week, anyway. It's Friday, October 23rd, and this is AI Daily Briefing. OpenAI lines up a bridge round at a trillion and a quarter, a research lab raises a billion dollars to squeeze more out of every GPU, Microsoft Research explains how that proof came together, and a small coding model, built largely by an automated research system, runs on a laptop. Tap follow so the next episode finds you.
Kate Rooney and Ashley Capoot at CNBC, who broke the story Thursday:
The San Francisco-based AI company is raising up to $30 billion at a $1.25 trillion valuation, according to people familiar with the deal, who asked not to be named because the details were confidential. That valuation does not include new capital raised.
So that's pre-money. Investors from the UAE are leading alongside Coatue, and the round is being pitched as an opportunistic bridge to an IPO. OpenAI told investors it wants signatures by the end of next week, and the final size is still being worked out. On Wednesday we passed along reports that OpenAI wanted about thirty billion at above one point four trillion. AI Affairs now says the target came down from one point four to one point two five. If it closes, it's still a record for a private company. But it's a trim, and it lands while Anthropic is reportedly pitching public investors on one and a half to two trillion. The line I'd underline is the side deal. One of CNBC's sources says OpenAI is also negotiating for more computing capacity from the UAE, so the money and the compute come from the same place. OpenAI told investors in February it's targeting roughly six hundred billion dollars in total compute spend by 2030. Thirty billion doesn't go far against that. Some timeline. OpenAI filed confidentially with the SEC in June. CFO Sarah Friar told staff in August it "will be a public company in 2027." A month later Sam Altman called this "an ill-advised moment to go public." And Mike Wheatley at SiliconANGLE notes OpenAI raised a hundred twenty-two billion in March at an eight hundred fifty-two billion post-money valuation. Seven months later, it wants a bridge. That's the burn rate of a frontier lab. And one more lab raising this week. Per AIToolsRecap, Reuters reports DeepSeek is raising about a hundred billion yuan, roughly fifteen billion dollars, at a pre-money near five hundred billion yuan, with CATL, Tencent, Meituan, Geely and GigaDevice named. We first covered that raise on October 6th, when it was still a plan. What's new is who's writing the checks: battery makers, carmakers and chip firms.
Jayujyoti Mullick at The American Bazaar, following Bloomberg's reporting via Quartz:
Co-founder Asher Spector told Bloomberg that the company believes demand for AI will outpace the available supply of computing resources and data. He argued that companies capable of extracting greater performance from existing infrastructure could gain a significant economic advantage.
The company is Flapping Airplanes. One billion dollars at a six billion valuation, led by Index Ventures and Lux Capital, with Sequoia, GV and Kleiner Perkins in and Nvidia among the backers. It was founded in late 2025, and per Quartz the valuation has quadrupled since January, when it raised a hundred eighty million. It's one of the so-called neolabs: research first, no product yet. The work is models that learn from less data and less compute, plus making the chips themselves more efficient. Co-founder Aidan Smith says they've been building their own version of the software that operates Nvidia's GPUs. That's the part that gets my attention. Writing your own version of the software that drives Nvidia hardware is brutally hard, and Nvidia is on the cap table. Spector's pitch is that even a ten percent gain on every chip is worth a fortune. Sure. But there's no benchmark, no paper and no product yet. A billion dollars for a pre-revenue thesis is exactly where my eyebrow goes up. Set it next to the first story, though. OpenAI is raising to buy more compute. This round is a bet that the cheaper path is using the compute that already exists better. Both bets can pay off. It's harder to believe both are priced right.
Xinzhi Zhang, a senior researcher at Microsoft Research, writing on the company's blog Wednesday about who did what:
The recursive method, its repeated application, and the proof that the accumulated errors remain bounded are original contributions of the researchers. AI tools helped explore candidate approaches, carry out detailed derivations, prepare the manuscript, and perform internal proof checks.
The problem is the Kannan-Lovász-Simonovits conjecture, proposed in 1995. Rough version: for a broad class of high-dimensional distributions, the best flat cut is nearly as good as the best cut of any shape, within a constant, no matter how many dimensions. It governs how fast random-walk sampling algorithms can explore a distribution. Zhang and co-author Zhao Song ran agents in parallel, pruned the dead ends by hand, and fed the most promising result into the next round. Roughly a hundred research trajectories went by before the key criterion emerged. The complete proof is in the October 4th version of their paper. And here's why I buy this one. On the day the full proof went up, a separate group of experts posted their own proof, built on Microsoft's first version and calling it "the crucial step in the proof." A third proof appeared that same day. Then came a Lean 4 formalization. Compare October 14th, when we covered Scientific American's report that OpenAI's Navier-Stokes proof wasn't what it first seemed. This time the checking came from outside the lab, and from a proof assistant. Back to that control from the top. Unguided runs on the same token budget fell short. Agents fanning out with an expert cutting branches got there. For anyone building research agents, that's the design lesson: the pruning step is the product.
TPS, in its open-weight release log for Wednesday:
Recursive released Flint, a free, open-weight mixture-of-experts coding model with 40 billion total and 2.5 billion active parameters, under Apache 2.0. The company claims its Eureka Machine automated research system contributed to nearly every stage of development, and that Flint runs at about 131 tokens per second on an M4 Pro MacBook Pro.
Two and a half billion active means it decodes like a small model. But all forty billion still have to fit in memory, as the QuelLLM catalog points out. It's already in the Ollama library under the flint tag. Doddi Priyambodo at Bicara IT, summarizing Recursive's launch post, says it was built in under four months and leads mixture-of-experts models under fifty billion parameters on SWE-Bench Pro v2 and Terminal-Bench 2.1. Those are Recursive's numbers. We haven't seen a model card or an independent eval. And "built largely by AI" is the headline everybody will run. Microsoft just showed what an honest version of that claim looks like: a breakdown of who did what, plus a control. Recursive says its system touched nearly every stage. I'd like the same breakdown. Per University 365's roundup, Wired reports the model performs strongly, and that's why the debate over self-improving research systems is heating up. For me it's simpler. A private, fast coding model on my own laptop is something I'd actually run. Whether it holds up at step seven of a ten-step refactor is what I'm testing this weekend.
Half of today's stories came back to compute. If that's the layer you care about, try our sister show, The Data Center Daily. It's a daily briefing on AI compute: hyperscaler capex, the power grid, semiconductor supply, and energy markets reshaped by intelligence at scale. Search for The Data Center Daily wherever you're listening.
So that's OpenAI's bridge round, Flapping Airplanes, the KLS proof and Flint, with links to all the reporting in the show notes. AI Daily Briefing is made by Lantern Podcasts, and our next episode arrives Monday, October 26th. Have a good weekend.