OpenAI’s AI math proofs face outside review; Meta’s 25 billion dollar sale draws five times the demand

Compact Conversations for 2026-10-27: 6 AI stories, ai news worth knowing in just 5 minutes.

[Audio embed placeholder]

The Lead: OpenAI’s AI models produce proofs for three major open questions, now under outside review

OpenAI says its internal AI models have cracked three major open questions in math, computer science, and physics. One is part of a Millennium Prize Problem, carrying a $1 million prize. Outside experts, including a Nobel laureate, are now checking whether the proofs hold up.

Why it matters: If the computer science proof holds, it challenges the Exponential Time Hypothesis, a foundational assumption in computer science. Hundreds of research papers built on that assumption could lose their foundation, making this a potential paradigm shift.

Source: The Decoder

Number to Know: Meta’s $25 billion share sale oversubscribed 5x as it raises $98 billion for AI

Meta’s upsized $25 billion share sale attracted demand for about five times the available shares. The company seeks to raise $98 billion for AI spending, with the 15 largest investors taking roughly 70% of the offering.

Why it matters: The oversubscription signals strong investor confidence in Meta’s AI strategy and the broader AI buildout. It also highlights the massive capital requirements for AI infrastructure, which could influence enterprise AI investment decisions.

Source: Techmeme

The Feed

OpenAI launches GPT-6.1 Terra, a lower-cost mid-tier reasoning model

OpenAI has launched GPT-6.1 Terra, a new reasoning model that delivers near-frontier performance on visual reasoning, professional task execution, and computer-use benchmarks at a 50% lower price than its predecessor. It’s aimed at organizations seeking cost-effective automation for large-scale operational workloads.

Why it matters: This model targets the enterprise sweet spot: balancing intelligence, speed, and cost. With pricing at $1 per million input tokens, it could make AI automation more accessible for routine tasks, but analysts caution that cheaper tokens don’t eliminate the cost of human review and integration.

Source: InfoWorld

Google’s Genie 4 world model generates interactive simulations with audio and real-time steering

Google announced Genie 4, a world model powered by Gemini 3.8 Flash that generates interactive AI simulations. It offers more realistic worlds, real-time natural language updates, audio generation, and experiences up to two minutes long. It’s launching as a research preview to trusted testers.

Why it matters: World models could transform how we prototype games, pre-visualize films, and train robots. While still in preview, Genie 4’s capabilities signal Google’s push into interactive AI, with potential applications across creative and industrial fields.

Source: The Verge

OpenAI plans to buy Google’s TPUs, diversifying away from Nvidia

OpenAI, historically a major Nvidia customer, plans to buy Google’s TPUs. A job listing shows it’s seeking an engineer to lead optimization of its models on TPUs, indicating a strategic shift toward diversifying its compute suppliers.

Why it matters: This move could reshape the AI chip market, breaking Nvidia’s near-monopoly. For enterprises, it signals more competition and potentially better pricing and availability for AI compute.

Source: The Information

Anthropic’s S-1 shows 85% gross margin, challenging AI profitability skeptics

Anthropic’s S-1 filing reportedly shows an 85% gross margin in the first half of the year from running its models. This is a direct rebuttal to skeptics who argued pay-as-you-go model sales wouldn’t be profitable.

Why it matters: This data point is crucial for the AI economics debate. If Anthropic can sustain such margins, it validates the business model of selling AI models as a service, which could encourage more investment and innovation in the space.

Source: The Information

One Thing to Try

Write a short rules doc (about 15 lines) covering your style, conventions, and common mistakes. Paste it at the start of every new chat or project, or drop it in your project folder if your tool reads it automatically. This stops the AI from repeating mistakes and saves you from re-explaining your preferences each session.

Sources

Transcript

Host A: Welcome to Compact Conversations, the show that compresses the day’s AI news into 5 minutes.

Host A: Today’s lead is OpenAI saying its internal AI models have produced proofs for three major open questions in math, computer science, and physics. [conversational] One of them is part of a Millennium Prize Problem — that’s one of the famous unsolved math challenges that each carries a one million dollar prize. The company has asked outside experts, including a Nobel laureate, to check whether the proofs actually hold up.

Host B: [curious] And one of those checks matters a lot more than a curiosity. The computer science result is a challenge to the Exponential Time Hypothesis, a long-standing assumption about which problems are fundamentally hard for computers to solve. According to The Decoder, if that proof holds on its own, hundreds of research papers lose their foundation, because they built on that assumption. Until the outside reviewers sign off, OpenAI’s claim stays exactly that, a claim.

Host B: One number to know today: twenty-five billion dollars. That’s the size of Meta’s upsized share sale, which Bloomberg reports ended oversubscribed by about five times — investors wanted far more shares than were available. [with emphasis] Meta is seeking to raise ninety-eight billion dollars for its AI spending, and the fifteen largest investors took roughly seventy percent of the offering. More than half went to long-only and existing investors.

Host A: In the feed, OpenAI has launched GPT-6.1 Terra, a mid-tier reasoning model that replaces GPT-5.6 Terra at a fifty percent lower price. InfoWorld reports OpenAI says it gets near-frontier results on visual reasoning and computer use. On one professional-tasks benchmark, it scored sixty percent at about a dollar per task, compared to GPT-6 Astra at seven dollars twenty-four. [quick aside] An analyst cited in the piece recommends enterprises test it on their own workloads. Another analyst cautions that cheaper tokens don’t erase the cost of human review. Pricing starts at a dollar per million input tokens.

Host B: Next up, Google’s Genie 4. The Verge reports Google announced this new world model — that’s a system that generates interactive, simulated environments rather than just text or video. It runs on Gemini 3.8 Flash. It adds audio, lets you steer the world with a natural language prompt in real time, and generates experiences about two minutes long, up from one. [lighter] It’s in a research preview for trusted testers first, coming soon to AI Ultra subscribers. Google warns details can drift as simulations run longer.

Host A: [curious] In chip news, The Information reports OpenAI plans to buy Google’s TPUs — a shift for a company that’s been a longtime Nvidia customer. A job listing shows it’s hiring an engineer to lead the optimization of its models on TPUs.

Host B: And finally, The Information reports Anthropic’s S-1 filing shows an eighty-five percent gross margin in the first half of the year from running its models — a reality check for skeptics who said pay-as-you-go model sales wouldn’t be profitable.

Host A: Your One Thing to Try comes from a popular post on the Prompt Engineering subreddit. [conversational] The problem: you explain how you like things done, the model nails it, and then three chats later it forgets everything and does it the old way again. The fix that costs nothing is a rules file. Write one short doc — the poster’s is about fifteen lines — covering your style, your conventions, and the mistakes the model keeps repeating. [with a small lift] Then paste it at the start of every new chat or project, or drop it at the top of your project folder if your tool reads it automatically. The payoff is you stop re-explaining yourself in every session. The catch is you have to actually keep it current — when the model does something new you like, add it to the file instead of hoping it remembers.

Host A: That’s Compact Conversations for Tuesday. More AI news tomorrow. Until then, happy prompting.