OpenAI says an unreleased model produced 372 new math results, mostly from a single prompt to a single agent, though outside experts want human verification. SpaceX is reportedly lining up about $40 billion in debt to buy Nvidia chips, and Lambda is reportedly raising up to $4 billion ahead of a 2027 IPO. Also: Meta, Walmart, Stripe and Sierra propose a standard for AI shopping agents, Mistral unveils a roughly one-trillion-parameter open-weight model, and Google ships Nano Banana 2.1.
OpenAI's unreleased model claims 372 new math results
Watch from 2:24
- What happened
- OpenAI posted 722 manuscripts grouped into 372 families to GitHub Tuesday evening, including a claimed solution to the four-dimensional Kakeya conjecture. The model was pointed at about 4,000 problems, and the average result used about 3 hours of ChatGPT Pro-level thinking. Many proofs are checked in Lean, a proof-checking language, which Scientific American says makes them all but certain to be correct. OpenAI has not released the model, the exact prompts, or the compute per problem. Fields medalist Terence Tao has criticized the pace, and MIT's Andrew Sutherland says single-agent one-shot claims should be treated as unverified until humans test them.
- The claim
- OpenAI says each result resolves or makes real progress on an open question in math or theoretical computer science, that nearly every result came from one prompt to one agent, and that the model is coming soon.
- Austin's take
- Austin calls it a phenomenal advancement, possibly timed to counter OpenAI's recent security troubles. He sees little direct everyday impact, but highlights three things: AI producing novel results at this volume and difficulty, the small compute per result compared with last month's Navier-Stokes claim (about 10,000 agents over 88 hours), and how fast progress moved in roughly a month.
- Why it matters
- If it holds up, AI is now producing research results faster than mathematicians can check them, and doing it with far less compute than before.
- What to do
- Treat the headline number as a claim until independent mathematicians weigh in; the Lean-checked proofs are the strongest part. Watch for the model's release and whether OpenAI shares prompts and compute details.
- Not confirmed yet
- The model, prompts and compute per problem are unreleased. Experts say the single-agent claims are unverified until humans check them.
SpaceX reportedly lining up $40B in debt for Nvidia chips
Watch from 7:50
- What happened
- The Financial Times reported Tuesday that SpaceX is arranging roughly $40 billion of debt, about $10 billion in bank loans and $30 billion in investment-grade bonds, organized by Apollo with PIMCO among firms in talks, to buy Nvidia chips for AI data centers. It may not close until 2027. Reuters reports SpaceX plans to use Nvidia chips exclusively in future data centers; SpaceX shares fell 1% and Nvidia rose half a percent in extended trading. Morgan Stanley estimates AI infrastructure needs about $1.5 trillion of outside financing by 2028.
- The claim
- Elon Musk has said xAI's Colossus 2 data center could double its Nvidia chips by December.
- Austin's take
- Austin is enthusiastic, reading big lenders' backing as a signal there is something real here. He frames it as a bet on the infrastructure layer of the AI stack, but flags the Cisco precedent: it overbuilt during the internet buildout and took about 20 to 25 years to recover its stock price. He thinks demand looks real for now.
- Why it matters
- AI is increasingly built on borrowed money, like railroads and power plants. Debt gets paid back whether or not the AI pays off.
- What to do
- If you hold Nvidia, SpaceX or AI infrastructure exposure, note how much of the buildout now rests on debt, and keep an eye on whether the deal actually closes.
- Not confirmed yet
- The deal is reported, not confirmed, and may not close until 2027.
Meta, Walmart, Stripe and Sierra propose rules for AI shopping agents
Watch from 14:24
- What happened
- Sierra published the Personal Agent Protocol on Tuesday with Meta, Walmart and Stripe, an open standard for how AI agents shop and act on your behalf. It uses OAuth, the tech behind 'log in with Google,' so you can grant an agent specific read or write access to an account; the business decides what agents can do and can see what they did. Shopify, Genesys, Instinct and Rocket also signed on, and a v0.1 spec is due later this month. Google, Amazon, Apple and OpenAI are not named partners.
- The claim
- Backers say it lets a business verify that a bot is acting for a real person. Sierra co-founder Bret Taylor, who also chairs OpenAI, told CNBC it is 'kind of chaos until such a standard exists.'
- Austin's take
- Austin agrees with Taylor and sees real time savings, like having an agent shortlist and buy a sofa. He doubts everyone will settle on one protocol, noting Amazon actively blocks shopping agents because agent searches undercut its sponsored ad revenue.
- Why it matters
- Right now a store can't tell whether it's dealing with you or a bot acting for you. Whoever writes this standard shapes how AI shopping works.
- What to do
- Nothing to install yet. If you run an online store, watch for the v0.1 spec later this month; as a shopper, expect to grant agents scoped permissions rather than full account access.
- Not confirmed yet
- It exists on paper only; the spec isn't out and major platforms haven't joined.
Mistral unveils a roughly one-trillion-parameter open-weight model
Watch from 20:22
- What happened
- France's Mistral unveiled Large 4, nicknamed Lechon, with a little over one trillion parameters and 49 billion active per word, a mixture-of-experts design. It was trained from scratch on 4,000 Nvidia Grace Blackwell GPUs over two months in Mistral's own European data centers and covers more than 160 languages. It's available through Mistral's API now; weights are due October 27 under Mistral's own license, not a standard open-source one. VentureBeat notes its coding score doesn't establish an outright lead.
- The claim
- Mistral calls it the most powerful open-weight AI outside China; co-founder Guillaume Lample says it is at the frontier of open-weight models. Mistral's own benchmarks put it at 62% on a software-engineering test.
- Austin's take
- Austin welcomes more competition and open models from Europe, and ties the compute-heavy training to why companies like SpaceX are borrowing for chips. He's curious what the custom license actually means.
- Why it matters
- Two big non-Chinese open models in two days means the open-weight race is no longer China's alone.
- What to do
- Developers can try it via the API now; read the license terms before planning to self-host once weights land October 27.
- Not confirmed yet
- Benchmark figures are Mistral's own, and license implications are unclear.
Lambda reportedly raising up to $4B before a 2027 IPO
Watch from 24:37
- What happened
- The Wall Street Journal reported Tuesday that AI cloud company Lambda is raising up to $4 billion at a $14.5 billion pre-money valuation, led by Blackstone and Coatue, as its last private round before a targeted 2027 IPO. Its order backlog jumped from $15 billion in June to $50 billion in September, anchored by a $35 billion compute deal with Anthropic. It was valued at $5.9 billion last November. Lambda rents Nvidia GPUs to AI labs, and Nvidia is an investor.
- Austin's take
- Austin doesn't think the industry is overbuilt yet, contrasting real backlogs with revenue-less dot-coms. He notes a contact's view that OpenAI's constraint is money, not compute capacity, and predicts everyday life will change dramatically by the end of 2027.
- Why it matters
- Companies renting chips to AI labs are growing as fast as the labs themselves. But a backlog is promised future spending, and one customer, Anthropic, accounts for most of it.
- What to do
- If you're watching the 2027 IPO, weigh customer concentration risk alongside the headline backlog.
- Not confirmed yet
- The raise is reported, not confirmed.
Google releases Nano Banana 2.1 image model
Watch from 29:26
- What happened
- Google released Nano Banana 2.1, built on Gemini 3.6 Flash, rolling out in the Gemini app, Google AI Studio and the Gemini API. Developers pay about 3 cents for a standard 1K image and about 7.6 cents for 4K. Reports disagree on price: The Decoder says about half the cost of Nano Banana 2, others say unchanged. AlphaSignal reports it beats the Pro model on editing.
- The claim
- Google says it improves on previous versions across the board; its claims are the only benchmarks so far.
- Austin's take
- Austin jokes that the world needs more AI slop, then says the real excitement is artists losing tedious manual work and becoming far more prolific, while purely human analog art rises in value.
- Why it matters
- Image editing is one of the AI features regular people actually use, and this lands directly in the Gemini app.
- What to do
- Try it in the Gemini app for edits; developers should confirm current API pricing before budgeting.
- Not confirmed yet
- Pricing reports conflict, and performance claims come only from Google so far.
