Three October 13 records materially extend the evidence. Google released Gemini 4 Argon and reported broad agentic gains while saying it cannot rule out biological and cyber critical-capability levels or a stealth-and-situational-awareness threshold; its conclusion that mitigations reduce residual risk is still provider-assessed. An independent Autonomous Science Bench pilot tested eight agent–model configurations on six budgeted design–build–test–learn campaigns: the best passed 67% of 144 trials, but the benchmark is small and relies mostly on replayed published measurements. A separate independent survey proposed a stringent scale for physical recursive self-improvement and found that no surveyed system meets it. METR, ARC Prize, Epoch AI, OpenAI, and Anthropic indexes showed no other newer material publication; current arXiv AI listings were reviewed. Editorial interpretation: stronger deployed agents and better tests sharpen both capability and control questions, but these records do not establish general intelligence, autonomous recursive improvement, or an overall trajectory toward wise superintelligence.
Evidence note 1 · Lab-reported model release and frontier-safety assessment, with mixed internal and third-party benchmark inputs
Gemini 4 Argon ships with stronger agentic results and explicit frontier-risk triggers
Google released Gemini 4 Argon globally for developers, enterprises, researchers, and subscribers. Its model card reports leading results on several knowledge-work, long-context, science, and multimodal tests, although it trails peers on others. In Google's frontier-safety assessment, the lab says it cannot rule out that the unmitigated model reaches biological and cybersecurity critical-capability levels, or a stealth-and-situational-awareness threshold after passing 14 of 15 related challenges; Google says deployed safeguards reduce residual risk below those levels.
Our interpretation. A broadly deployed agent that can sustain complex coding, science, cyber, and computer-use workflows is more consequential than an isolated benchmark gain because capability comes from the model plus tools, context, permissions, and monitoring. Crossing—or being unable to rule out crossing—frontier-risk thresholds makes the quality and independence of mitigations central to whether capability progress can be governed wisely. It is not evidence of general intelligence or superintelligence.
Main limitation. LAB CLAIM: The comparison mixes Google-computed scores, provider self-reports, and third-party leaderboards, sometimes with different harnesses, time limits, or input settings. Argon does not lead every reported benchmark. The critical-capability and residual-risk judgments are Google's assessments, and the public evidence does not by itself let outsiders reproduce every result or verify the effectiveness of deployment mitigations at scale. Passing 14 of 15 tailored challenges is evidence about those tests, not proof of real-world loss of control.
Next evidence. Independent common-harness evaluations of long-horizon work and computer use; external audits of the biological, cyber, and loss-of-control safety cases; adversarial attempts to evade the real-time monitor; post-deployment rates of prompt injection, boundary violations, hallucination, and human overrides; and evidence that mitigations remain effective as tools and permissions expand.
Evidence note 2 · Independent research preprint and benchmark pilot
Autonomous Science Bench tests experiment choice, not just answers
An independent research team released version 0.1 of Autonomous Science Bench: six budgeted design–build–test–learn campaigns grounded in published biology, chemistry, and materials experiments. Agents choose batches of measurements and must pass both a discovery gate on what they found and a learning gate on how well their final model ranks held-out designs. Across eight agent–model configurations and 144 trials, the best configuration passed 67%; the hardest campaign was solved in only 3 of 24 attempts.
Our interpretation. Autonomous scientific discovery could accelerate both useful science and AI development. Evaluating which experiments an agent chooses under a budget, and whether it learns a transferable model, is closer to that capability than static question answering. The split gates also resist treating a lucky hit as scientific understanding. This is a benchmark advance, not a demonstration of autonomous science at human breadth or reliability.
Main limitation. INDEPENDENT RESEARCH PREPRINT: Version 0.1 covers only six campaigns in three fields. Five campaigns replay recorded measurements and one uses a digital twin; none of the reported pilot trials runs a new physical laboratory. Published source data may have appeared in training or be retrievable if egress controls fail. Thresholds and campaign construction come from the benchmark authors, and the strongest system still failed roughly one-third of trials.
Next evidence. Independent reproduction; evaluations on unpublished data and live physical laboratories; blind campaigns across more fields; comparisons with domain scientists under matched time and cost budgets; reliability across repeated runs; and reporting of unsafe, invalid, or scientifically misleading experiment choices rather than success scores alone.
Evidence note 3 · Independent research preprint and literature synthesis
Survey finds no demonstrated physical recursive self-improvement
An independent research collaboration proposed an L0–L5 framework for physical recursive self-improvement in robots. It separates ordinary task adaptation from changes to the improvement procedure itself, same-task gains from across-task gains, and internal verification from independent evaluation. After mapping recent systems into this framework, the authors conclude that the surveyed literature does not demonstrate their highest standard: independently verified recursive gain.
Our interpretation. Recursive self-improvement is often invoked as a route from capable AI to superintelligence. The framework usefully asks whether a system merely improves a task policy or actually becomes better at producing later improvements across tasks while inheriting those changes. Its negative finding narrows what present robotics evidence supports and counters the reification of incremental adaptation as an already-achieved intelligence explosion.
Main limitation. INDEPENDENT RESEARCH PREPRINT AND LITERATURE SYNTHESIS: This is a proposed taxonomy and review, not a preregistered meta-analysis or a new demonstration. Classifications depend on the authors' definitions and on incomplete reporting in the surveyed papers. Failure to meet a stringent evidentiary standard does not show that physical recursive improvement is impossible, nor does simulation evidence establish hardware safety.
Next evidence. Controlled experiments comparing successive improvement procedures from matched starting systems and resource budgets; protected external evaluators; gains that transfer across genuinely different tasks; real-robot replication; and evidence that safety constraints, monitoring, and recovery remain valid after each inherited update.