Anthropic Ditches Coding Interviews as Genie 4 Launches and AI Claims ETH Disproof
Latent Space · rss · 2026-10-28
Lead: Anthropic ends coding interviews
Anthropic hiring manager Felix Reiseberg stopped using traditional coding interviews: "Claude is now better at both React challenges and assembly quizzes." His advice to junior engineers: pick business impact, choose one radical idea, build a prototype over a weekend, focus on a single good idea, and lean into fresh user empathy.
Google DeepMind ships Genie 4
- Real-time interactive world model built on Gemini Omni, with Gemini planning multi-step events in the loop
- First Genie with native audio (ambient sound, music, live dialogue); claims aerodynamics, collisions, natural human motion
- Positioned as safe virtual environments for testing agents and robots; research preview live for trusted testers, then AI Ultra subscribers via Project Genie, first API availability through Gemini API
AI for Math
- Axiom completed a full Lean auto-formalization of the Classification of Finite Simple Groups: an agent swarm produced 50M lines of Lean in 24 days, reviewing each other's work via GitHub and reportedly fixing hundreds of errata—the largest computer-verified proof ever
- OpenAI claimed results including a refutation of the Exponential Time Hypothesis; Anthropic says Claude independently found the key ETH-disproving algorithm with 14M output tokens and no human input
- Ten Claude agents proved in Lean that seven electrons on a sphere form a pentagonal bipyramid (Thomson problem)
Independent evals
- AA Intelligence Index v5: 60% private/held-out sets; Claude Opus 5.5 (max) leads at 58, GPT-6 Astra 57, GPT-6.1 Sol 55, Gemini 4 Argon 50; open-weights gap widens from 3 to 12 points; GPT-6.1 Terra costs $0.96/task vs $13.04 for Claude Fable 5.1
- FrontierCode 2.0 (Cognition): top score 45% with underspecified prompts graded on merge-worthiness; 80% of PRs rejected during construction; top 20 all OpenAI/Anthropic plus Grok 4.7
- GPT-6.1 Terra: beats Astra on vision tasks while 47% faster and 92% cheaper
- Others: Microsoft's MAI-Image-2.7-Preview tops 4/7 image-editing actions; open-weight Ling 3.1 Flash (560B/25B active, 1M context, MIT) scores 44 but uses 40%+ more output tokens
More from Companies & People
- How to Land a Research Internship Without Any Connections: A 5-Step Playbook — cneuralnetwork · 2026-10-28
- Agentic Tourism: 56% of US Travelers Now Use AI to Plan Trips, Reshaping Distribution — robertonappe · 2026-10-28
- Dev explains Microsoft's four Copilots as nested if, for, while and while(true) — pswider · 2026-10-28
- DeepSeek means to Liang Wenfeng what SpaceX means to Musk, not X — teortaxesTex · 2026-10-28
- Anthropic's Stainless repo deletion cascade-wiped customer's private fork with no notice — kingslyj · 2026-10-28
- UMA_Robots Scales Robot Fleet, Hiring Software Engineer in Paris for Fleet Ops — RemiCadene · 2026-10-28