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: AI for Science is the material mover. Anthropic reports autonomous 24–48h de novo protein-binder campaigns with no human input into design decisions; two independent CROs synthesized and tested the delivered designs. Claude produced binders against 14 of 15 interpretable targets, with 354/1,320 designs binding (27% hit rate). This is stronger evidence for sustained autonomous scientific workflow plus external physical validation, but humans still chose targets, placed synthesis orders and interpreted assays, so it is not a closed autonomous wet-lab loop.
The previous Daily closed before the AI-led Yau–Tian–Donaldson disproof was discovered. That result, plus the subsequent protein-design validation, lifts AI for Science from 56.9 at the prior Daily to 59.0 today. The YTD claim remains a preprint without independent proof validation, so the ASI forecast is unchanged.
Frontier release coverage: DeepSeek released V4-Flash-Vision-Exp on 21 Aug as an experimental multimodal/vision-agent API model and it is recorded in events/model_landscape as Tier A preview. No other verified flagship/frontier release was found in the 24h two-pass checks across OpenAI, Anthropic, Google DeepMind, xAI, DeepSeek and Qwen/Alibaba.
Anthropic also expanded Mythos 5 into Claude Security repository scanning. This is new operational deployment of already-known frontier cyber capability, not a new capability-ceiling event. CHIVE, published 21 Aug, found no predictive uplift from three activation-reading interpretability tools on its counterfactual behavior evaluation; this reinforces interpretability/control as a bottleneck but does not move readiness.
Open Problem Closure Radar: 11 new records entered since the previous Daily, 1 existing record received a material verification/status update, and 5 of the new records have documented material AI roles with significance >=7. The most important is the AI-led YTD disproof (9.5/10, full claim, independent verification pending). Other notable AI-linked records include the smooth random fast dynamo (8.5/10, partial because the deterministic smooth problem remains open), Big-Line-Big-Clique first open case (7.6/10), exact linear-extension counting below 2^n (7.3/10), and Marton's inner-bound capacity question (7.4/10). The radar also recorded a rank-30 elliptic curve with AI provenance still unclear. DVG remains 1.5x (1.01 discoveries/month vs 0.68 verified/month), coverage partial.
Robotics: no verified new autonomy-duration/intervention-rate/generalization result changes maturity. Generalist GEN-1.5, discovered after the prior Daily, is a useful one-shot physical-learning signal but remains first-party, short-horizon and pending independent reproduction. Robotics stays 47.4.
Compute & Infrastructure: no new 24h anchor-changing event beyond already-ingested infrastructure signals. Compute stays 75.1.
Synthetic Biology: GenBio AIDO Cell is a material virtual-cell modeling release but novel wet-lab validation is pending; SynBio stays 66.8. The autonomous protein-binder result is scored under AI for Science rather than double-counted into SynBio maturity.
Energy/Fusion, Advanced Materials, Longevity and BCI/HMI: no verified 24h event changes their maturity anchors. Energy 46.8, Materials 62.4, Longevity 36.3, BCI 42.8.
Benchmark Observatory: no new official METR Time Horizon, HLE or FrontierMath point requires ingestion. Benchmark health remains a concern as public tasks saturate or become exposure-sensitive; no score movement today.
Readiness: Intelligence 63.7; AI for Science 59.0; Robotics 47.4. ASI Readiness 59.03. Singularity Readiness 53.19. Rolling ASI probabilities remain 1y 2%, 2y 6%, 3y 12%, 5y 30%, 10y 81%, central estimate 2032–33.
Bottom line: the strongest evidence today is not a new LLM benchmark score but a 24–48h autonomous scientific-design workflow whose outputs survived independent physical testing. The key remaining bottleneck is closure of the entire research loop—especially prospective experiments, replication and verification—without humans bridging the physical stages.