Hiring Bar: Math + Clear Writing

DeepMind's Aletheia autonomously produced correct solutions for six of the ten FirstProof challenge problems, with the paper reporting majority‑expert agreement on those solutions. (arxiv.org) The project team says it generated multiple publication‑grade outputs — including one paper produced with no human intervention (Feng26) — and has published full prompts and model transcripts on the Superhuman GitHub repository. (math.berkeley.edu) Aletheia is built on an advanced Gemini Deep Think backbone and operates with an agentic Generator–Verifier–Reviser loop while using Google Search and web browsing to pull and synthesize research literature during proof construction. (deepmind.google) DeepMind reports state‑of‑the‑art numbers on internal benchmarks — Aletheia reached 95.1% accuracy on the IMO‑ProofBench Advanced — and the team describes an inference‑time scaling law that raises correctness while cutting per‑problem compute relative to prior Deep Think iterations. (math.berkeley.edu) The authors explicitly recommend rigorous release practices (including "human‑AI interaction cards") and public transcripts to document provenance and novelty, making clear that readable, verifiable exposition is treated as a measurable research output. (math.berkeley.edu) DeepMind job postings for Research Scientist roles list a PhD or equivalent, prefer a track record of peer‑reviewed publications, and require candidates to "report and present research findings clearly and efficiently both internally and externally, verbally and in writing," aligning hiring criteria with the skills Aletheia emphasizes. (job-boards.greenhouse.io)