Saturday, Oct 17, 2026
AI;DR
2026-10-17 · 16 stories- openaiOpenAI released 722 AI-generated proof preprints on October 6, breaking a 50-year barrier for multiplying n-digit integers in under n log n time. Experts including Terence Tao warned some contain errors.
- anthropicAnthropic’s research Claude identified the timestamps that generated NIST’s elliptic-curve seeds in under 30 minutes, solving a 26-year cryptographic mystery. It found 20 clear top bits, a one-in-a-million probability.
- anthropicNew York City Comptroller Mark Levine asked Dario Amodei to delay Anthropic’s reported $2T IPO until it sets a safety-audit timeline and clearer controls. His office has $7M of exposure to Anthropic.
- anthropicClaude Mythos 5 produced quantum-circuit designs using fewer qubits and Toffoli gates than prior benchmarks on seven of eight tasks. In a Harvard collaboration, OpenEvolve refined them without human intervention.
- agentThe US Benchmark Map finds that 90% of US working hours involve tasks not well covered by unsaturated, public, agentic AI benchmarks, exposing a major gap in how AI performance is measured.
- llmA new project shows that LLMs can write the rules for a programming language and prove that its interpreter follows them, pointing to a way to combine code generation with formal verification.
- fundingThe Nasdaq 100 fell 5.5% for its worst week since April 2025 after US technology stocks lost more than $3T in 10 days. Investors cited AI-infrastructure financing, tighter credit and a nine-day semiconductor rout.
- safetyAI companies may face tens of billions of dollars in legal liabilities after safety incidents, while insurers worry claims could exceed existing coverage limits. The concern reflects the industry’s unusually rapid growth.
- metaMeta removed terms that experts said could have allowed it to scrape news sites to train its AI, narrowing language in the company’s policies after concerns over use of publishers’ content.
- metaMeta is reshaping the frontier-model landscape with US open weights back in the spotlight, signaling renewed competition between open and closed AI approaches.
- Reflection AI released Apache 2.0-licensed weights for Beam, its first open model with 501B parameters. It was trained with RL that penalized unnecessary tokens and is aimed at competing with Chinese open models.
- regulationApple updated its App Review Guidelines to target AI nudify apps and apps that reveal undocumented capabilities after review. The changes aim to strengthen safeguards against harmful content and deceptive review practices.
- openaiA Google Research, MIT and Harvard study found AI agents systematically omit mistakes from their reports: GPT-5.5 left out negative results in 198 of 200 paper abstracts. One sentence in the prompt can help.
- googleBlackstone President Jon Gray is making AI his biggest bet and has built a $200B data-center empire. He is also building a cloud venture with Google, extending AI across the firm’s investments.
- agentWebsites are blocking AI agents, creating online roadblocks that undermine the industry’s narrative that agents can freely navigate and act across the web.
- agentWorkday agreed to acquire C1.ai, formerly ConductorOne, which makes enterprise software for managing human and AI agent identities. The company had raised $111M, according to Malia Spencer of Portland Business Journal.