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

Saved daily version · 2026-10-16

Daily evidence notes

Five October 15 publications add evidence since the previous review: Anthropic describes embedded METR access; a Gemini agent team completes Minecraft; MLExpert reports stronger machine-learning engineering; LEASH tests selective escalation between models; and Google demonstrates an antimicrobial search workflow. These are company announcements and author-reported preprints, not independently replicated findings. Editorial interpretation: model capability, tools, memory, coordination, and oversight must be evaluated together. A successful run, benchmark gain, or predicted biological effect does not establish general intelligence, safety, or wisdom. No overall wisdom-trajectory judgment is made. METR, ARC Prize, Epoch AI, Google DeepMind, OpenAI, Anthropic, and current independent research listings were reviewed; older releases and database refresh dates were not treated as new findings.

Evidence note 1 · Lab announcement about independent evaluation access; results pending

Anthropic starts embedding METR evaluators

Anthropic says six METR staff now have company hardware, workspace, and internal access comparable to its risk teams. METR may publish assessments without a company veto, subject to predefined redactions. This follows the evaluator proposal covered yesterday.

Our interpretation. Internal access could expose risks that public benchmarks miss. Whether institutions respond to uncomfortable findings matters alongside rising capabilities.

Main limitation. LAB ANNOUNCEMENT, not a completed independent evaluation. Customer data, privileged material, and architecture information have access restrictions. The arrangement is voluntary; other partnerships have narrower rights. Access alone does not certify safety.

Next evidence. METR's own assessment reports, disclosed access gaps and redactions, and documented changes to safeguards or development decisions.

Evidence note 2 · Author-reported agent demonstration; preprint, not independently replicated

Four Gemini agents sustain a 27-hour Minecraft task

A preprint reports four Gemini 4 Argon agents defeated Minecraft's Ender Dragon in 27 hours 14 minutes, writing 2,783 programs and dying 605 times. They coordinated through chat and shared memory without human gameplay intervention.

Our interpretation. Sustained coordination and recovery suggest progress in systems that execute many dependent steps. This measures a model-plus-tools team, not an isolated model.

Main limitation. AUTHOR-REPORTED DEMONSTRATION: One run on Easy difficulty cannot establish reliability. Mineflayer exposed structured state, including normally hidden information, and supplied pathfinding. Prior game knowledge and persistent memory helped; this was not play from pixels or learning unfamiliar rules.

Next evidence. Repeated random-seed trials, independent reproduction, tool and memory ablations, and transfer to unfamiliar environments with consequential failure costs.

Evidence note 3 · Lab-reported benchmark evaluation; preprint, not independently replicated

MLExpert reports gains in automated machine-learning engineering

Researchers introduce MLExpert, a 35-billion-parameter model trained to organize experiments and learn from intermediate outcomes. They report MLE-Bench Lite medal rates rising from 37.9% to 72.7% relative to its base model, with additional optimization-task results.

Our interpretation. Better automated experimentation could contribute to faster AI development. Successful engineering benchmarks do not demonstrate sustained recursive self-improvement.

Main limitation. LAB-REPORTED PREPRINT: The headline covers 22 tasks and three evaluations; larger-model comparisons use single evaluations. Harnesses and compute affect results. Benchmark success is not evidence of frontier research invention or an accelerating improvement loop.

Next evidence. Independent reproduction using released code and weights; held-out, unfamiliar research tasks; matched compute budgets; contamination checks; and measured research gains after accounting for human supervision.

Evidence note 4 · Author-reported computer-use evaluation; preprint, not independently replicated

LEASH tests when a smaller agent should seek stronger help

LEASH uses a smaller model's internal uncertainty signal to route computer tasks, escalate difficult steps, and observe the stronger agent afterward. On OSWorld, authors report 73.66% success versus 72.90% for always-on Claude Opus 5, at 19.2% lower inference cost.

Our interpretation. Selective assistance could make capable agents cheaper. Monitoring and willingness to hand off are useful mechanisms to test, not validated measures of wisdom.

Main limitation. AUTHOR-REPORTED PREPRINT: Evidence is limited to OSWorld. Predicting the smaller model's failure does not ensure the stronger model can help. The post-handoff observer faces changed inputs, and its intervention effect was not causally isolated. Task success does not establish resistance to deception.

Next evidence. Independent replication, unfamiliar-task calibration, isolated monitoring experiments, and explicit tests of correlated failures and deceptive behavior.

Evidence note 5 · Lab-reported scientific workflow; biological validation pending

Google's antimicrobial workflow identifies a candidate awaiting lab testing

Google DeepMind describes biologists using Antigravity and protein-function predictions to search databases and prioritize an antimicrobial peptide candidate. Its reported activity is predicted; wet-lab validation is underway.

Our interpretation. Connecting agents, specialist models, and researchers could accelerate scientific search. Demonstrating useful discoveries requires evidence beyond a plausible computational candidate.

Main limitation. LAB-REPORTED WALKTHROUGH: The announcement does not establish experimental antimicrobial efficacy, toxicity, clinical usefulness, or a controlled productivity gain. Human researchers and existing databases remain central parts of the workflow.

Next evidence. Published laboratory protocols and results, appropriate controls, independent replication, failed-candidate rates, and comparisons of time and resources against conventional searches.