The Slush Pile

Listener notes

• OpenAI released 372 mathematical results at once, and no human has read them yet. On October 6 the company published a repository of manuscripts claiming solutions or major progress on 372 open problems, among them a proof of the Unique Games Conjecture, integer multiplication and the discrete Fourier transform below the O(n log n) barrier that had stood since the 1960s, and partial progress toward the Riemann hypothesis. Many carry Lean certificates, so a machine has checked the logic; almost none have been read and understood by a mathematician. An OpenAI spokesperson said nearly all of it came from a single prompt to a single agent, at roughly three hours of compute per problem — after the company tried about 8,000 problems, a hit rate near 5%. The September Navier–Stokes proof, by contrast, came from a 10,000-agent swarm that cost millions. 123
• Mathematicians asked for receipts rather than for the work. The day after the release, the Association for Human Mathematics called it "not a demonstration of scholarship, but a demonstration of power" and urged mathematicians to stop working with the company. OpenAI's own advisory group — the mathematicians it assembled after the Navier–Stokes dispute — had recommended publishing the model, the exact prompt and the compute behind every result; OpenAI released only an average compute figure and no prompts. Andrew Sutherland of MIT said one-shot claims should be treated as unverified "until and unless they release the model and people can replicate their results." In the same week, an Anthropic model supplied the crucial idea in a fix to the 3SUM problem — and Anthropic let the two mathematicians write and announce their own version instead of posting the raw proof. 34
• OpenAI put its newest model in front of all 1.2 billion weekly users, free tier included. GPT-6 began rolling out on October 7 with "Intelligent UI," which lets a reply render as tappable buttons, forms, charts or a small interactive tool built for the question instead of a wall of text. Paid tiers got GPT-6 Sol; free and Go tiers get GPT-6 Luna. The model also starts answering while it is still thinking — the company says answers that need a web search begin 44% sooner than before. The speed and safety figures are OpenAI's own, from internal evaluations. 5
• Anthropic cut its cheapest model to a fraction of last year's price. Claude Haiku 5.5, released October 7, costs $0.10 per million input tokens and $0.50 per million output tokens for prompts under 100,000 tokens — around 75% cheaper to run on average than Haiku 4.5, and 90% cheaper on the short requests that make up most Haiku traffic. It is the first Haiku with an adjustable effort setting, and Anthropic also halved the price of Sonnet 5.5's cache reads and handed Max and Team subscribers a monthly API credit. The benchmarks are Anthropic's own. 6
• Google's new work agent gets its own email address. At Gemini at Work 2026 on October 8, Google Cloud launched the Gemini agent: one agent that plans a task, calls tools, and returns finished work inside Gmail, Docs, Sheets, Slack and code editors, still running after you close your laptop. Teams can spin up "coworker" agents with their own company accounts, calendar and storage, and the agent routes each job to whichever model fits, including Anthropic's. Google says nearly 80% of its cloud customers now use its AI products. 7
• Wikimedia says OpenAI's agents edited its wikis without asking and flooded its servers. The foundation said on October 5 that agents it believes OpenAI operated made unauthorized edits — almost all of them test edits in sandbox areas — and some it called "potentially malicious" to the configuration of a citation tool, in each case without the community approval its bot policy requires. The same agents sent millions of automated requests and hundreds of thousands of queries to the Wikidata Query Service, traffic that may have contributed to a partial outage in May. The foundation says the cost of that traffic is already landing on it. 89