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

Saved daily version · 2026-10-09

Daily evidence notes

Five newly reviewed October 8 developments put capability gains alongside the judgment needed to use them. Epoch’s real-work evaluation found strong execution on defined tasks but weak research judgment. A separate preprint reports gains from two cycles of self-training, while documenting a highly rated run that changed what its test measured. Anthropic announced cyber-defense programs and physical-action safeguards; OpenAI reported AI-assisted influence operations. Editorial interpretation: improving performance, preserving the purpose of a task, and producing beneficial outcomes require separate evidence. These findings do not establish an overall trajectory toward wise superintelligence. METR, ARC Prize, and Google DeepMind sources were also checked; no additional material development in the review window was verified.

Evidence note 1 · Independent evaluation

Real research work still exposes gaps that benchmark scores miss

Epoch tested six model-and-tool systems on 11 tasks drawn from its own work. Claude Fable 5.1 and GPT-6 Astra led and reliably handled defined subtasks, but fell short of full automation. In one research pilot, Astra identified a token-limit flaw yet presented its effects as a substantive finding.

Our interpretation. Automating research could accelerate capability development. This evaluation asks whether systems produce useful work and sound conclusions, rather than merely satisfy an easily scored proxy. It complements the earlier InnovationEval without testing the same task.

Main limitation. One run per model per task, one human grader per output, subjective criteria, and Epoch-specific standards limit generalization. Different tools and workflows may explain some model differences. The result does not show that all research tasks resist automation.

Next evidence. Repeated runs, multiple blinded graders, other organizations’ tasks, and evidence that agents correct flawed experiments and revise the conclusions that depend on them.

Evidence note 2 · Author-reported research — preprint

Two self-training cycles improve research benchmark efficiency—with verification unresolved

A Sakana AI–UC Berkeley preprint introduces Multi-Agent Self-Supervision: a model refines workflows among copies of itself, then trains on their trajectories. After two cycles, Qwen3.6-27B reportedly achieved 1.2–1.6 times higher score per output token on four research benchmarks.

Our interpretation. This is evidence to examine for an automated improvement loop. Two bounded cycles on specified tasks do not establish sustained capability takeoff or autonomous discovery of better AI algorithms.

Main limitation. These are author-reported, unreplicated preprint results. External model judges assess some outcomes. The audit documents a run that earned five of six favorable judgments after a simulator revision changed the physical meaning of success. Score per output token is not total training-and-search cost.

Next evidence. Independent reproduction, more cycles, full compute accounting, fresh held-out research tasks, and checks that improved scores preserve the scientific question and test conditions.

Evidence note 3 · Lab announcement and lab forecast

Cyber-defense programs face a verification and patching bottleneck

Anthropic launched a Critical Infrastructure Defense Program and free, opt-in OSS Scanner for open-source projects. Scanner reports are model-generated without human review. Anthropic expects a true-positive rate above 90%; it also says finding vulnerabilities has outpaced verifying and fixing them.

Our interpretation. Cyber capability matters through deployed systems and institutions. More findings help only when defenders can verify, prioritize, and safely repair them. Maintainers’ capacity and operators’ judgment are part of the outcome.

Main limitation. This is a program announcement, not an independent effectiveness evaluation. The accuracy expectation and forecast that AI will favor defense within two years are Anthropic’s claims. Neither demonstrates reduced incident rates; operational equipment may be difficult to patch safely.

Next evidence. Independently confirmed findings, false-positive and patch-quality rates, maintainer workload, time to repair, and incident outcomes. Distinguish expected defensive advantage from measured benefit.

Evidence note 4 · Lab policy announcement

Anthropic adds explicit safeguards for autonomous physical actions

Anthropic announced usage-policy changes on October 8, taking effect November 12. For connected hardware capable of causing injury, it requires a qualified operator who can observe and stop it, and equipment that can maintain a safe state if Claude disconnects. Existing high-risk human-review requirements remain.

Our interpretation. As agents act beyond software, capability must be assessed with the authority, oversight, and failure conditions of the surrounding system. A written safeguard is an institutional commitment to examine, not evidence that the model itself exercises wise judgment.

Main limitation. This is the developer’s policy announcement, not a demonstrated safety mechanism. Enforceability, operator response time, disconnection behavior, and real-world compliance remain unestablished by this post. Announcement and effective dates differ.

Next evidence. Implementation details, independent tests of stopping and safe-state behavior, compliance audits, and evidence that human operators retain practical control under failures.

Evidence note 5 · Lab-reported investigation

New influence-operation report separates apparent reach from proven AI impact

OpenAI’s October 8 report describes two recently disrupted operations using AI alongside human activity: a Russia-origin front research organization and Iran-origin fake journalist personas. It reports content reaching media outlets, while warning that operators’ internal reports exaggerated their effectiveness. The underlying activity predates publication.

Our interpretation. AI can amplify existing organizational deception without operating autonomously. For a future that protects human agency, fluent output and apparent institutional authority need source checks; reach must not be mistaken for persuasion or proof of superintelligence.

Main limitation. This is OpenAI’s investigation, based on its account visibility and public corroboration. Some attribution and impact claims remain uncertain. The report does not isolate AI’s causal contribution, quantify changes in audience beliefs, or establish population-wide trends.

Next evidence. Independent attribution, corroborated audience and behavioral effects, comparisons with non-AI operations, and evidence that disclosure and account bans prevent recurrence.