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

Saved daily version · 2026-10-15

Daily evidence notes

Two October 14 publications materially extend the evidence. METR proposed independent loss-of-control assessments spanning agent alignment, internal control systems, and possible AI-driven acceleration of AI R&D, with access to company data and private staff interviews; METR explicitly says the voluntary, company-scoped exercises would surface evidence rather than certify safety. Epoch AI's SpireClimb benchmark found that frontier agents now exceed 90% win rates at Slay the Spire's easiest setting and reach 17–20% at its hardest, up sharply from recent failures but still far below expert humans and potentially influenced by game-specific knowledge or training. ARC Prize, OpenAI, Anthropic, and Google DeepMind indexes showed no other newer material research release; the current arXiv AI listing had no October 15 updates. Editorial interpretation: capability is improving in long-horizon, partially observed computer-use settings while credible loss-of-control assessment still depends on access, independence, and disclosure rules. Neither item establishes general intelligence or an overall trajectory toward wise superintelligence.

Evidence note 1 · Independent evaluator proposal; no completed assessment result

METR proposes independent access to test internal loss-of-control risks

METR proposed assessment exercises covering three linked questions inside frontier developers: whether agents might pursue seriously harmful actions, whether monitoring and control systems would catch and stop them, and whether AI-assisted R&D could accelerate capabilities faster than safeguards. METR asks for weeks or months of access to relevant company data and systems, private staff interviews, and permission to publish METR-written reports subject to a disclosed redaction process.

Our interpretation. Progress toward superintelligence is shaped not only by model scores but by whether institutions can observe, understand, and constrain increasingly capable internal agents. Testing alignment, control, and AI-R&D acceleration together addresses a central evidence gap: capability may advance faster than public evaluation or organizational response. The proposal is a governance and measurement mechanism, not evidence that any developer is currently safe or unsafe.

Main limitation. INDEPENDENT EVALUATOR PROPOSAL: No new assessment result is reported. Participation, scope, access, and redactions would remain at each company's discretion, creating selection and pressure risks. METR says the work cannot provide confident safety assurance, audit compliance, or make go/no-go decisions because methods and standards are immature. The proposed scope excludes important areas including model-weight security, sandbox robustness, and broader organizational governance.

Next evidence. Which frontier developers accept the proposal; the exact scope, access, redactions, and publication deadlines they permit; whether staff can participate without retaliation; the first public reports and their raw evidentiary support; independent replication by other assessors; and whether findings lead to measurable changes in monitoring, incident response, or development decisions.

Evidence note 2 · Independent evaluation and benchmark report

SpireClimb finds rapid gains in long-horizon visual game play

Epoch AI introduced SpireClimb, a visual computer-use benchmark in which agents play Slay the Spire and its sequel through screenshots, mouse, keyboard, persistent notes, and a restricted game wiki. Frontier models now win the original game's easiest setting more than 90% of the time. At its maximum difficulty, GPT-6 Astra won 17% of runs and GPT-6.1 Sol won 20%; top human players are reported near or above 80%. Epoch describes this as a sharp gain from models that struggled to finish the game six months earlier.

Our interpretation. The task combines long-horizon planning, partial information, resource management, visual interaction, and recovery from stochastic outcomes, making progress more informative than a static question set. Rapid improvement suggests that general agent and computer-use advances can unlock tasks that recently resisted frontier models. A game benchmark, however, does not establish broad real-world autonomy, judgment, or wisdom.

Main limitation. INDEPENDENT EVALUATION: Slay the Spire is public and popular, and Epoch says frontier models display extensive prior knowledge; labs may also train specifically against it. Results may therefore mix general improvement, computer-use gains, memorized strategy, and targeted reinforcement learning. The maximum-difficulty sample is 60 runs per model on fixed, arbitrarily chosen seeds, leaving wide uncertainty around low win rates. Models use maximum reasoning and a wiki, and the sequel is an early-access game frozen at one version.

Next evidence. Independent reproduction on the same seeds and harness; confidence intervals from larger samples; transfer to unfamiliar games and hidden rules; contamination and targeted-training disclosures; performance under varied interfaces, compute budgets, and tool access; and correlations with consequential long-horizon work such as software engineering or scientific experimentation.