This page may contain stale information. Last updated: 2026-09-24
Definition
- 2026-10-26: openai bundled three frontier results — harmonic analysis (Sogge local smoothing), ETH false (Lean Oct 17), QCD/Yang–Mills existence (held for physics shepherd release) (2026-10-26-openai-three-scientific-discoveries, 2026-10-26-openai-science-discoveries-verge)
Treat each claim's verification status separately; physics materials not publicly released.
- 2026-09-23: anthropic Claude agents autonomously discovered array-associated-reverse-transcriptases in bacteriophage DNA — ~950 agents, 21h, 210M tokens; wet-lab validation; Feng Zhang endorsement (2026-09-23-anthropic-claude-enzyme-discovery)
- 2026-09-23: enveda raised 2B valuation; multiple AI-discovered drugs in clinical trials (2026-09-23-enveda-311m-ai-drug-funding)
- 2026-09-21: openai forms IAS-hosted Advisory Group on Mathematics and AI — governance response to navier-stokes-millennium-problem controversy and “100+ open problems” claims (2026-09-21-openai-math-advisory-group)
OpenAI's claim of resolving 100+ open mathematical problems and the Navier-Stokes solution remain unverified by independent mathematicians.
-
2026-09-18: anthropic wet lab bridges in silico models to bench biology (2026-09-18-anthropic-wet-biology-lab)
-
2026-09-18: damo-radar open-source medical imaging release (2026-09-18-alibaba-damo-radar-medical-ai)
-
2026-08-01: openai-astra ten math/TCS advances with lean-formalization (2026-08-01-openai-ten-advances-mathematics-primary)
AI for Science (AI4Science) applies machine learning, agentic orchestration, and domain-specific foundation models to accelerate scientific discovery — from hypothesis generation through experimental validation.
2026 Milestones
Agentic Beamline Control (July 2026)
Nature Machine Intelligence: LLM agent autonomously performs X-ray sample alignment on ssrl BL17-2 synchrotron beamline (2026-07-01-nature-agentic-xray-scientist):
- Trained in virtual six-circle diffractometer simulator; deployed on real beamline without customization
- mcp tools: terminal I/O, detector images, motor scans
- Claude Opus 4 on real experiments; human operator passive safety relay only
- Correctly identified reference reflections and orientation matrix for single-crystal scattering
ElementsClaw — Materials Discovery (July 2026)
alibaba DAMO Academy’s ElementsClaw (icml-2026):
- Fuses Large Atomic Models (1B-parameter Elements) with LLM semantic reasoning
- Screened 2.4M crystals in 28 GPU hours → 68,000 superconductor candidates
- 4 experimentally verified novel superconductors (highest Tc 6.5K)
- Open-sourced dataset at developer.damo-academy.com/material
Parisi-Claude Jamming Proof (July 2026)
Nobel laureate giorgio-parisi + Claude (Opus 4.7) proved decade-old identity a+b=1 in jamming theory (2026-07-01-parisi-claude-jamming-arxiv). Pattern: numerical verification → proof attempt → human refinement.
Agentic Discovery Pattern
- Domain-specific foundation model (atomic, molecular, physics)
- LLM orchestration for literature reasoning and workflow
- Closed-loop experimental validation
- Human expert verification of AI outputs