Thursday, June 11, 2026

📡 Jin Miao Signals — Morning Brief · 2026-06-11

1. Top 5 — what actually matters today

• Anthropic kills its own Fable 5 "self-sandbagging" safeguard after researcher revolt — what changed since Tuesday: the covert throttle on frontier-AI research we flagged is gone, walked back within ~24h once researchers called it covert sabotage of competitors — a live precedent on whether a lab can quietly degrade its model for some users; matters to anyone building on a single model family WIRED .
• World Model Self-Distillation: video world models that solve tasks without text crutches — distillation + RL elicits planning straight from an unlabeled scene, no paired task-execution videos to collect; the scalability unlock world-model researchers have been waiting for, and squarely on the paradigm Jin tracks paper .
• Almond Robotics ships Axol — an open-source dual-arm humanoid at $7,999 — 860mm reach, 6.5kg payload, 500Hz control, US-assembled, open Python SDK + VR teleop; if real, this drops the embodied-AI hardware on-ramp by an order of magnitude for builders and labs (Rumor — single social source, see §5) @CyberRobooo .
• Google DeepMind funds a $10M program on what happens when millions of agents interact — with Schmidt Sciences, Cooperative AI, ARIA; the first serious money on emergent collective agent behavior, not single-agent alignment — the risk surface tech workers shipping agent fleets are about to inherit @GoogleDeepMind .
• EvolutionaryScale drops ESMFold2 + ESM Atlas, reportedly beating AlphaFold — announced on NoPriors with Zuckerberg/Rives/Chan alongside the Biohub "cure all disease" push; a new flagship bio-foundation model is the kind of capability jump that resets a whole science vertical (Rumor — pod announcement, awaiting benchmarks, see §5) @saranormous .

2. New-direction sparks

• Calibration Drift Under Reasoning — past a task-specific CoT budget, more "thinking" makes models systematically overconfident on wrong answers. Non-obvious because it inverts the "reason longer = safer" reflex everyone bakes into agent loops arXiv .
• The Structural Attention Tax — RAG format (KG triples) hijacks 2–3× the attention per token vs. semantically equal prose, independent of relevance. A retrieval-design lever almost no one is tuning arXiv .

3. Threads worth watching

• World models / embodied AI — directly moved by two fresh, independent results: World Model Self-Distillation and Embodied-R1.5 , a 15B-token unified embodied foundation model with a plan→ground→correct loop. The research substrate under the humanoid hardware wave is hardening fast.

4. Contrarian watch

• China's InP chokehold as the AI-datacenter single point of failure [OUTLIER] — Reuters reportedly flags indium phosphide control threatening the buildout; consensus prices GPUs and power, not the optical-substrate supply chain ($AXTI named) @aleabitoreddit .
• Transformers may structurally lack executive control [OUTLIER] — a PNAS Nexus result argues attention has deficient executive control; cuts against the "scale + reasoning budgets fix it" consensus that the Calibration-Drift paper also dents PNAS Nexus .

5. Verification flags

• ⚠️ Almond Robotics Axol $7,999 open-source humanoid — do not act on yet; needs primary source (single X post, no spec sheet/site) @CyberRobooo .
• ⚠️ ESMFold2/ESM Atlas "beating AlphaFold" — do not act on yet; podcast claim, no published benchmark @saranormous .
• ⚠️ OpenAI weighing drastic token-price cuts to win Anthropic customers — do not act on yet; "people familiar" sourcing only @kimmonismus .

Markets context only — not financial advice.