AI Proves 18-Year Quantum Conjecture; Agents Earn but Can't Catch Spies; Shell Choice Swings Scores 24%
This daily brief covers five AI developments: PSI proving the Aaronson-Ambainis quantum conjecture in Lean in three days; Anthropic's EVE Online experiment where agents earn 3x but fail to detect insiders; Harness Leaderboard showing coding agent shells cause 24% solve-rate variance; Vork headless 3D engine for agents; and Arete-27B's shallow SFT reducing marketing tone.
PSI Proves Aaronson–Ambainis Quantum Conjecture in Lean
Physical Superintelligence (PSI), an AI research lab, used an internal system running an unreleased Anthropic model to formalize the 2008 Aaronson–Ambainis conjecture into a machine-checkable Lean proof in roughly three days. The conjecture addresses how much quantum algorithms can outperform classical ones on unstructured problems; the proof shows at most a polynomial speedup, not exponential. Peer review is ongoing and system details remain unpublished.
Anthropic Project EVE: Agents Earn but Fail Insider Detection
Anthropic's red team placed Claude into EVE Online , a player-driven space MMO, to test multi-agent collaboration and insider-threat detection. Over six months, the new model generation achieved ~3× the average returns of the older model. However, in a four-ship fleet with a designated saboteur, detection rates were extremely low: even the best public monitor caught only ~30–40% of betrayals. The result demonstrates that economic competence does not imply social deduction ability.
Harness Leaderboard: Coding Agent Shells Cause 24-Point Solve-Rate Swing
Using the same base model across nine different coding-agent "shells," the Harness Leaderboard (~29,000 test runs) found average solve rates ranging from ~51% down to ~27% — a ~24 percentage-point gap. The top five shells were nearly tied, but the worst shell also incurred over 10× higher per-task cost. The benchmark underscores that the agent framework (tool use, prompting, orchestration) is as critical as the underlying model.
Vork: Headless 3D Engine for AI Agents
Voyager released a research preview of Vork, a headless 3D/VFX engine designed for AI agents. Instead of manual Blender workflows, agents write code to define materials, lighting, and physics. In a demo, ~1 million starlings were simulated and rendered in ~97 seconds on a single L4 GPU. The engine targets creative and simulation teams that need editable, reproducible shots.
Arete-27B: Thin SFT Layer Cuts Marketing Tone
Nous Research applied shallow supervised fine-tuning (~52.8M tokens) on Qwen3.8-27B, reducing responses with promotional language from ~22% to ~1% while leaving general capabilities almost unchanged. This suggests the "values layer" can be extremely thin. Weights and a custom evaluation suite are open-sourced, though the evaluation carries a stated stance and uses an LLM judge. Teams concerned about stance drift in open-source deployments can replicate the test before adopting the model as a primary production model.
Additional Notes
Meta Muse Spaces: Muse can now generate persistent sidebars or homepage widgets (itineraries, gift trackers) and connect to external app backends for self-updating content.
Claude Duet × Claude Code: Anthropic brought coding-task monitoring to mobile; Duet can attach to Claude Code sessions, view progress, review diffs, and send instructions.
Visa Shopping Agent Benchmark: 22 shopping agents achieved ~57% overall pass rate; over half of failures still placed incorrect orders. The key challenge is knowing when to stop due to incomplete information.
OpenAI Funding Rumor: Media reports claim ~$11B first tranche at ~$125B valuation; no official confirmation within the announcement window.
Routine Updates: OpenAI Developers Showcase/API credits; Anthropic science modeling blog post and Claude Code permission toggles; no new frontier model releases.
Signed-in readers can open the original source through BestHub's protected redirect.
This article has been distilled and summarized from source material, then republished for learning and reference. If you believe it infringes your rights, please contactand we will review it promptly.
How this landed with the community
Was this worth your time?
0 Comments
Thoughtful readers leave field notes, pushback, and hard-won operational detail here.
