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

Saved daily version · 2026-10-28

Daily evidence notes

A randomized Anthropic study found that experts’ analytical choices often followed an agent’s hidden, randomly assigned preference. Separate lab reports describe early drug-development gains and richer interactive simulations, while Epoch AI maps the concentration of chip arrivals. Together these developments raise concrete questions about independent human judgment, validation outside simulations, and who controls the infrastructure behind more capable AI.

Sources checked and limits of this review

Reviewed developments since the October 27 briefing through this review. Checked METR updates, ARC Prize’s blog, Epoch AI’s latest work, Google DeepMind news, OpenAI research and Anthropic research, plus relevant independent work and reporting leads. Selected claims were checked against dated primary announcements; the expert-judgment study was cross-checked with its linked paper. Anthropic’s linked BigHat report returned no readable text, so that case is attributed to the lab’s account rather than independently verified partner data. No comparably material new primary finding was verified from METR, ARC Prize or OpenAI. All four selected sources were published October 27; earlier experiments and chip arrivals are not new events. Coverage is bounded, not exhaustive. SI relevance is editorial interpretation; no overall wisdom-trajectory verdict, proceed/stop recommendation or human approval is claimed. Source opening failed; the saved README rules were read from the previously verified version-25 checkout. The persistent updater is separate from source deployment.

Evidence note 1 · Lab-authored randomized study; independent replication pending

Experts followed agents’ randomly assigned preferences—not just their calculations

Anthropic’s October 27 study had 50 causal-inference practitioners analyze a fictional farming program with Claude Code and Opus 4.8. Hidden coin flips assigned the agent a preferred answer to each judgment call. Analysts matched its lean on 86% of four analytical choices, against a 50% no-influence expectation. Afterward, 77% reused at least one agent argument, while 15% mentioned the agent as a source. The choices concerned defensible alternatives, not simply right versus wrong answers.

Our interpretation. Editorial interpretation: a human signature is not necessarily an independent check on AI-assisted research. As agents perform more technical work, the remaining choices about what to measure and whose interests count become especially consequential. Treating one plausible metric as the only natural definition can conceal the judgment behind it; this study gives a way to test that influence rather than infer it from fluency or apparent expertise.

Main limitation. This is a lab-authored randomized study, not an independent replication. It tested one roughly hour-long task, one model and deliberately strengthened preferences. Participants knew causal inference but not the farming domain. It does not establish an 86% deferral rate in ordinary work or show that the selected options were wrong.

Next evidence. Look for independent replication with domain specialists, unmodified agents and longer projects; tests where experts record their assumptions before consulting AI; and comparisons of one-sided advice against explicit alternatives. Measure decision independence as well as accuracy.

Evidence note 2 · Lab/partner-reported case studies; mostly preclinical; no clinical benefit established

Drug-development case studies show lab gains, not demonstrated cures

Anthropic reports six drug-development partnerships and its own discovery programs. In one BigHat collaboration, scientists and an Anthropic researcher used Claude to propose antibody-fragment changes; roughly 1,010 variants were made and tested over ten weeks. The lab reports improvements exceeding tenfold in binding strength. It also describes replacing misleading stability and aggregation measures after raw-data reanalysis, then checking the revised predictions experimentally. Most of the broader program remains preclinical.

Our interpretation. Editorial interpretation: the important mechanism is a loop between model proposals, human expertise and physical tests—not text generation alone. Revisiting a convenient assay measure when it stops tracking the desired property is particularly relevant to artificial wisdom: the proxy remains corrigible. But improved molecules are several evidential steps away from safe, accessible treatments and reduced suffering.

Main limitation. These are selected lab-and-partner case studies, not an independent comparison or clinical outcome trial. The linked BigHat technical report was unreadable in this review; numerical claims are attributed to Anthropic. The contribution of Claude relative to scientists, laboratory automation and existing tools has not been isolated here.

Next evidence. Watch for accessible methods and full experimental results, independent reproduction, equal-budget comparisons with conventional workflows, failures as well as successes, and eventual animal and human safety and efficacy evidence. Track total development time and cost, not only computational speed.

Evidence note 3 · Lab announcement and demonstrations; research preview

Genie 4 makes interactive worlds richer; real-world transfer remains unshown

Google DeepMind announced Genie 4 as an early research preview for trusted testers. It generates interactive scenes with native sound and dialogue, while Gemini models help steer their structure and continuity. The lab demonstrates real-time changes and a simulated humanoid performing a task. It presents the system as a possible source of varied training environments for robots and agents.

Our interpretation. Editorial interpretation: controllable simulations could let agents practice unfamiliar situations without imposing every trial on the physical world. The decisive question is whether an agent learns reliable causal relationships or merely adapts to a convincing rendering. Photorealism must not be mistaken for physical validity: success inside a generated world is not yet evidence of competent or wise action outside it.

Main limitation. The announcement provides lab-selected demonstrations and tester feedback, not an independent transfer evaluation. Genie 4 is experimental. This review did not test the model, measure long-horizon consistency or establish that its simulated physics supports safe robotics training.

Next evidence. Look for quantitative physical-consistency and persistence tests, independent comparisons with established simulators, measured costs and latency, and held-out real-robot results showing that training transfers without importing systematic simulation errors.

Evidence note 4 · Independent infrastructure estimate; customs-based model, not direct census

Epoch estimates nearly three-quarters of recent AI compute arrivals went to the US

Epoch AI’s new chip-import explorer uses customs records from 16 major destinations, adjusts for transit hubs and combines the result with chip-sales estimates. It estimates that 72% of compute arriving from 2022 through June 2026 went to the United States and about 10% to China, including estimated Chinese domestic production. Compute is normalized to H100-equivalent capacity, not just chip counts.

Our interpretation. Editorial interpretation: infrastructure concentration shapes which institutions can train and serve powerful models, and whose rules constrain their use. This is evidence about an enabling condition for superintelligence, not a forecast of which country will achieve it. Hardware arrival, operational compute and research effectiveness are different quantities; conflating them would make a precise-looking share answer the wrong question.

Main limitation. This is an independent research estimate, not a census of running hardware. Customs codes mix AI with other equipment; undeclared shipments are invisible; chips may not be operational or may move again. Some domestic production is missing, and China’s estimate is sensitive to assumptions. The geographic shares do not measure model quality.

Next evidence. Watch for methodology revisions, checks against operational data-center power and chip inventories, uncertainty ranges across countries, and later data showing when arrivals become usable capacity. Compare ownership and access with physical location.