SuperintelligenceMONITOR
Published by Center for Artificial WisdomLatest newsletter: Issue #1 · draft v1.11

Saved daily version · 2026-10-31

Daily evidence notes

OpenAI’s new Rosalind-6 account describes a restricted biology deployment of Astra’s existing weights, not a newly trained frontier model. Newly reviewed Epoch studies show an open-weight model exploiting a software-release gap and estimate rapid growth in chip capacity sold. The evidence points to changes in access, practical capability and infrastructure—not a demonstrated safe path to superintelligence.

Sources checked and limits of this review

Reviewed since the October 30 briefing, checking original research and announcements from METR, ARC Prize, Epoch AI, Google DeepMind, OpenAI and Anthropic, plus independent research and reporting leads. OpenAI’s Rosalind-6 addendum is dated October 30. The two Epoch reports are newly reviewed here with their October 29 publication dates preserved; they were not included in the previous edition. No additional material development was selected from the other watchlist sources checked. This is a bounded review, not an exhaustive internet search. Epoch’s biology-preprint insight appears as October 30 on its listing but October 21 on the article; it was not treated as new evidence. Routine product and website-control announcements were not treated as capability breakthroughs. No forecast or overall wisdom-trajectory verdict is offered. Repository refresh failed; editorial and updater instructions were read from the previously verified version-25 checkout, matching the Site’s current version. Website briefing only; no newsletter campaign sent.

Evidence note 1 · Lab-reported deployment and evaluation; SI relevance is editorial interpretation

Rosalind-6 changes research access, not the underlying model weights

OpenAI’s October 30 addendum says GPT-Rosalind-6 uses Astra’s weights, without added training data or post-training for this release. It offers fewer refusals on some dual-use biology questions through approved research access. OpenAI classifies biological/chemical capability as High, below Critical, under its own framework.

Our interpretation. Editorial interpretation: usable scientific capability depends on deployment rules as well as model training. Whether expanded research assistance benefits people depends on experimental validation, access decisions and effective human oversight—not a model name.

Main limitation. These are lab assessments, not independent confirmation of scientific benefit or safety. The four Critical biological evaluations did not reach their indicative thresholds, but those thresholds are provisional. OpenAI says it is still repeating biosecurity red-teaming of the shared monitoring and enforcement process after deployment.

Next evidence. Independent wet-lab validation, completed safeguard testing and evidence that required human review catches consequential errors.

Evidence note 2 · Independent capability experiment, newly reviewed; SI relevance is editorial interpretation

Newly reviewed: a public model turns unshipped fixes into browser exploits

Epoch reports that GLM 5.3 compromised six Firefox releases using public fixes not yet shipped to users, plus an unpublished flaw it rediscovered. The researchers supplied an offline environment and release-time information. Marginal API cost was about $1,000 per version after more than $100,000 of Firefox research spending.

Our interpretation. Editorial interpretation: practical cyber capability can spread beyond restricted frontier systems. Defenders’ release processes and patch timing help determine consequences; this is not simply a model-ranking story.

Main limitation. Independent research, not an independent replication. The result depends on the supplied tools and environment. Mozilla patched the unpublished flaw on October 27; the researchers had not confirmed success against that release. Comparable Chrome attempts failed. The study does not establish the overall offense–defense balance.

Next evidence. Replication on updated software, broader targets and tests of whether shorter public-fix-to-release delays reduce exposure.

Evidence note 3 · Independent infrastructure estimate, newly reviewed; not a forecast; SI relevance is editorial interpretation

Newly reviewed: chip-capacity additions grow rapidly, but sales are not deployment

Epoch estimates that quarterly AI-chip capacity sold grew about 2.5 times annually from early 2024 through mid-2026: a fitted doubling every 9.1 months. It converts sales from six designers into H100-equivalent computing capacity. This describes additions each quarter, not a measured doubling of intelligence.

Our interpretation. Editorial interpretation: hardware supply can support larger training runs and more inference, but turning chips into useful capability also requires power, operational data centers, software and effective research.

Main limitation. This is an independent estimate of a historical trend, not a forecast. Chips can take months after sale to become operational. The 8.4–10.0-month statistical interval excludes uncertainty in underlying sales estimates. Ten quarters of data do not establish an enduring growth law.

Next evidence. Revised sales estimates, actual energized capacity and evidence connecting extra compute with reliable performance rather than spending alone.