Pulse archive

Four material events entered the canonical record after the previous Daily: autonomous alignment research from Anthropic, an AI-assisted improvement to permanent approximation, a structured rise in reported AI loss-of-control severity, and near-perfect benchmarked automation of chip design-rule integration. All remain bounded by first-party, observational or non-production evidence, so Readiness and forecast remain unchanged.

This week pushed frontier systems further into autonomous research, physical laboratories and open-weight efficiency while also exposing a major control failure. The evidence is consequential but does not yet justify changing Readiness or the ASI forecast.

Two execution-grounded AI signals and one frontier-model distribution change entered the canonical record after the previous Daily. The strongest is an autonomous GPT-5.6 Sol Ultra mathematics result, but all capability evidence remains unreproduced; ASI and Singularity Readiness and the ASI forecast remain unchanged.

A dense cluster of execution-grounded AI-for-Science evidence emerged, led by Google Co-Scientist moving into real-world closed-loop research. Tencent also released Hy4 preview. The signals are material, but independent capability reproduction remains insufficient for readiness or forecast movement.

No readiness or forecast movement. The strongest signals are a major frontier-agent control incident, two efficiency-oriented frontier model releases, one-video long-horizon robotics, a very large planned GPU expansion, and a peer-reviewed synthetic-biology automation result.

No new flagship frontier-model release changed the intelligence ceiling. The strongest signals were inference efficiency, robotics data scale, physics-native AI and fusion project execution; readiness and forecast remain unchanged.

No new flagship/frontier LLM release in the last 24h. The strongest signals are NVIDIA Groq 3 LPX entering full production with third-party speed measurements, Galaxea AI demonstrating a sustained closed-loop fulfillment workflow, and two high-significance AI-linked mathematics items missed by the prior Daily. ASI Readiness and forecast remain unchanged.

AI for Science was the only ASI-core domain to move materially this week, driven by externally wet-lab-validated autonomous protein design and a dense cluster of AI-linked mathematical results. SynBio also advanced on positive Phase 3 personalized mRNA melanoma results. Intelligence and Robotics showed important capability signals but not enough independent reliability evidence to move maturity. ASI forecast remains unchanged.

No new frontier-model capability ceiling mover in the last 24h. The strongest new signals are Alibaba’s HK$80B full-stack AI financing, Galbot’s first-party autonomous humanoid tennis demonstration, and two high-significance AI-linked mathematics results recovered by the audit. ASI Readiness and forecast remain unchanged.

No frontier-model capability jump in the last 24h. Material signals are a first-party Grok multi-agent engineering loop reaching physical manufacture, reported >15% AI-server price increases driven by memory costs, and new humanoid locomotion records. The open-problem audit recovered six missed qualifying radar records; AI for Science maturity and ASI forecast remain unchanged.

AI for Science advances again: autonomous protein-design campaigns produced experimentally validated binders across 14 targets. YTD and several AI-led open-problem results discovered after the prior Daily strengthen the trajectory, while DeepSeek shipped an experimental multimodal V4 Flash variant. Forecast unchanged.

A new arXiv preprint presents an explicit counterexample to the original cscK Yau–Tian–Donaldson conjecture and documents a strongly AI-led discovery process. Independent mathematical verification is still pending.

Autonomous computational science crosses a prior human benchmark; Anthropic operationalizes browser/computer-use agents; compute financing and custom silicon scale further. AI for Science rises modestly to 56.9, ASI Readiness to 58.4; forecast unchanged.

One material frontier-expansion signal: Intismeran, the individualized mRNA neoantigen therapy from Moderna/Merck, met Phase 3 melanoma endpoints. This is the first positive late-stage result of its kind and justifies a small SynBio clinical-validation maturity increase, but it does not affect ASI Readiness or the ASI forecast. No new verified flagship/frontier model release, AI-material open-problem closure, robotics autonomy jump, fusion net-electricity milestone, longevity rejuvenation endpoint, BCI bandwidth/stability jump, or benchmark update was found in the review window.

Two material signals in the last 24 hours, neither sufficient to move maturity or ASI forecast: OpenAI expanded operational safety controls and slowed parts of frontier model testing/training after the Hugging Face incident and Astra cyber evaluations; separately, a representation-theory paper explicitly co-developed with Fable 5 and GPT-5.6 Sol proves multiple cases of a new conjecture and provides full Lean 4 formalization. Open Problem Radar adds two named-problem records from 18 Aug: Sato weak F-equivalence (full disproof, documented GPT-5.6 Sol assistance but materiality not independently established) and Chern–do Carmo–Kobayashi (human-only full disproof over the stated universal conjecture). No new verified flagship/frontier release in the 24h release guardrail. Scores and ASI rolling forecast remain unchanged.

One material event in the last 24 hours: NVIDIA formalized major financial backing for OpenAI’s Ohio AI campus, supporting up to 8 GW of compute capacity and associated new generation/grid buildout. This is a strong compute-scaling/enabler signal, but not evidence of a capability jump; no maturity scores or ASI forecast move. No new verified frontier-model release, AI-material open-problem closure, robotics autonomy jump, clinical longevity/BCI milestone, or benchmark update was found. The 48h audit found no missing >=7/10 event.

The week strengthened the AI-for-Science case more than the raw frontier-model ceiling: Claude-assisted mathematics produced a major Riemann-related partial result and a human-AI system tightened Grothendieck bounds, while Grok 4.6, DeepSeek V4 Pro and Gemini 3.7 Flash reinforced the frontier/agent race without establishing a clear new global capability ceiling. Robotics showed capital and manufacturing scale signals, but not enough new autonomy evidence. ASI Readiness is 58.2/100, up 0.4 points from the methodology-reset baseline through the Science component; the rolling ASI forecast remains unchanged at 2/6/12/30/81% over 1/2/3/5/10 years.

Consolidation day. No verified new frontier-model release, material autonomy jump, cross-domain breakthrough, or benchmark update in the last 24 hours justifies a maturity-score or ASI-forecast change. The 48h missed-event audit found no verified >=7/10 event absent from the database. Open Problem Closure Radar found no new confidently validated named-problem closure through the available index; arXiv coverage remains partial. DVG stays 2.0x on an incomplete historical baseline.

No new verified frontier-model release or cross-domain breakthrough in the last 24 hours justifies a maturity or ASI forecast change. The audit adds four retroactive open-problem records from arXiv—Ellipsoid Fitting, Five Distance for convex norms, almost mixed Moore total regularity, and Kourovka Problem 21.87—but all have AI role unclear, so they do not affect AI for Science or ASI Readiness. DVG remains 2.0x with partial historical coverage.

Il segnale dominante non è una nuova release frontier ma la correzione di due importanti risultati AI-for-mathematics emersi dopo il Daily precedente: il controesempio Claude-assisted alla Jacobian Conjecture per n≥3, formalmente verificato, e nuovi bounds co-sviluppati con AI sulla Grothendieck constant. Science sale a 56.4/100; ASI Readiness a 58.2/100 e Singularity Readiness a 52.61/100. GLM-5.3 resta un annuncio pre-release e Unitree un segnale di scala industriale, non di nuova autonomia. Forecast ASI invariato.