ANALYSIS
Methodology note: Historical index values on this Pulse are normalized to the current methodology for comparability. The original as-published record remains unchanged in the canonical archive.
Verdict: the strongest evidence since the previous Daily is an unusually broad cluster of execution-grounded AI-for-Science results. Google Co-Scientist is the clearest signal, extending a Gemini-based multi-agent system into real-world research loops with physical or experimental interaction across multiple domains. Separate preprints show frontier models designing operations-research algorithms, controlling atomic-force microscopy workflows and performing closed-loop protein-folding model development. Tencent also released Hy4 preview as a major open-weight agentic model. The breadth is notable; the verification profile is not yet strong enough to move readiness. ASI Readiness remains 59.03/100, Singularity Readiness 53.19/100, and the ASI forecast remains 30% at 5 years / 81% at 10 years with a 2032–33 central estimate.
AI for Science — direct ASI component
Google Co-Scientist / real-world closed-loop science — Impact 9.1/10, evidence 80/100, primary confirmed. A Google-led preprint reports Gemini-based agents operating across end-to-end scientific workflows that include interaction with real experiments and empirical validation rather than stopping at hypothesis generation. This is the strongest trajectory signal in the current candidate set because it moves AI-for-Science toward execution-grounded closed loops. Caveat: the paper remains first-party, heterogeneous across case studies and not independently reproduced; some scientific outputs remain unconfirmed. Policy outcome: no AI-for-Science maturity, ASI Readiness or forecast change.
GPT-5.6 Sol near-optimal operations-research algorithm design — Impact 8.7/10, evidence 81/100, primary confirmed. A preprint reports that a single untuned GPT-5.6 Sol query produced interpretable class-level algorithms competitive with specialized methods across inventory, queueing and assortment optimization, with frozen algorithms evaluated on held-out instances. This is a strong algorithm-design signal, but the evaluation is author-run, unreviewed and not independently reproduced; contamination and broader generalization remain unresolved. No readiness or forecast movement.
EPFL agentic atomic-force microscopy control — Impact 8.5/10, evidence 78/100, primary confirmed. EPFL researchers report MCP-connected agents executing checked commands, assessing images and tuning a live atomic-force microscope across bounded workflows. This is real scientific-instrument control rather than a software-only agent demonstration. The strongest caveats are narrow workflow scope, small human-comparison sample, human-defined safety bounds and no independent end-to-end reproduction. No readiness or forecast movement.
AgentFold closed-loop protein-folding model design — Impact 8.4/10, evidence 79/100, primary confirmed. AgentFold reports a multi-agent system completing long-horizon model design, debugging, training and evaluation in a protein-folding codebase and outperforming matched proposal/random controls. It is execution-grounded scientific ML work, but it remains an author-run preprint on a bounded substrate without independent reproduction. No readiness or forecast movement.
Live AI-assisted brain-tumour surgery at UCL/UCLH — Impact 8.3/10, evidence 86/100, independently confirmed. UCL and UCLH report first live use of a learned vision system providing real-time anatomy guidance during brain-tumour surgery. The deployment and patient outcome are independently reported, but this is one patient, the surgeon retained control and no controlled comparison isolates the AI contribution. This is a material clinical deployment signal, not yet evidence of general autonomous scientific or medical capability. No readiness movement.
AI-co-developed rapid-mixing proof on girth-five graphs — Impact 8.2/10, evidence 78/100, primary confirmed. A new spectral local-to-global principle lowers a longstanding graph-girth restriction in spin-system mixing, with a substantial disclosed AI role. The theorem is material but partial, unreviewed and lacks independent proof verification. No AI-for-Science or forecast movement.
Full replica-symmetry-breaking preprint — Impact 8.7/10, evidence 64/100, primary confirmed. A July preprint newly surfaced in the candidate set claims a full zero-temperature result for the Sherrington–Kirkpatrick spin glass with an unusually direct ChatGPT 5.6 role. It is retained in the full Daily because it was discovered after the previous full review, but not promoted into the X top five: the underlying result is older, the Lean artifact is conditional on analytic inputs, contradiction review found scope concerns and no independent full verification exists. No readiness movement.
Additional AI-linked scientific results above the Compass threshold include a polynomial-time stable-matching result for network hypergraphs (7.6), Fröberg-conjecture slices (7.5), a longstanding Heyting-algebra realization question (7.4), a dual futile-cycle positivity certificate (7.6), an entanglement-of-formation counterexample (7.3) and an improved every-genus hyperbolic-surface systole bound (7.7). These broaden the AI-for-Science evidence base, but none clears the independent-verification threshold for readiness movement.
Intelligence / frontier models
Tencent Hy4 preview — Impact 8.8/10, evidence 86/100, independently confirmed release. Tencent released a 770B-total / 49B-active open-weight agentic model with a live API and long-context support. Release state and availability are independently confirmed. Capability comparisons remain first-party and unreproduced, with preview limitations still material; the capability penalty therefore remains active and Intelligence does not move.
Compute and ecosystem overlays
Anthropic / Nscale reported $45B compute lease — Impact 8.2/10, evidence 72/100, preliminary. Reporting describes a very large commitment for planned West Virginia AI capacity. It is strategically important but concerns future capacity, lacks first-party confirmation and is not evidence of current compute or model capability. Compute remains an overlay; no readiness movement.
Other material overlay signals include SK hynix beginning construction of its $4B Indiana HBM packaging facility (7.2), NVIDIA reportedly pausing some AI-cloud revenue-sharing deals (7.3), Moonshot reportedly discussing Kimi K3 distribution/revenue sharing with major hyperscalers (7.8), and termination of the Solstice / Element Solutions merger (7.1). These are system-structure or future-capacity signals, not demonstrated frontier capability.
Open Problem Closure Radar
The period added or updated several AI-linked mathematical problem records, but none advanced to independently verified full closure. The highest-significance newly surfaced item is the full replica-symmetry-breaking claim at 8.7/10, with full claim scope but no independent complete proof verification. The rapid-mixing, stable-matching, Fröberg and large-systole results remain partial-scope; the Heyting-algebra result is full-scope but only AI-assisted and unverified independently.
Readiness and forecast
ASI READINESS: 59.03/100 — Δ0.00 today
SINGULARITY READINESS: 53.19/100 — Δ0.00 today
ASI FORECAST: 5Y 30% | 10Y 81% — Δ0 pp
Central estimate: 2032–33 — unchanged.
Bottom line: the important pattern is breadth across execution-grounded scientific work: real experimental loops, physical instrument control, algorithm design and scientific-model development are all appearing in the same review window. That is stronger evidence of a changing AI-for-Science regime than any one isolated demo. But almost all of the strongest capability claims are still first-party or preprint-stage. Independent reproduction is now the clearest gate between a rapidly broadening evidence base and a justified readiness increase.