Friday, June 26, 2026

📡 Jin Miao Signals — Morning Brief · 2026-06-26

1. Top 5 — what actually matters today

• World-model hallucination is now predictable — and preventable. New work shows the "visually fluent but physically wrong" drift in generative world models concentrates in low-coverage state-action regions, and ships MMBench2 (427h, 210 tasks, live sims) + a 350M model to detect and mitigate it — the missing reliability layer for anyone building world/sim models. [tech/founder lens] huggingface
• General Intuition raised $320M (~$2.3B) to train agents on gameplay action-data. The bet: millions of hours of games teach something closer to human intuition than text ever will — a real, non-obvious wedge into embodied/world-model agents, and the day's biggest applied-AI raise. [founder/markets] (raise still single-sourced — see §5) techcrunch
• OpenAI's internal Codex token use grew 56x in Research, 27x in Engineering since Nov 2025. This is the cleanest internal-adoption telemetry we've seen — agentic coding isn't plateauing inside a frontier lab, it's compounding. If you build dev tooling or staff eng teams, this is the curve to plan against. [tech-worker] latent.space
• Databricks' ex-AI chief claims a 1,000x cut to AI's power bill. New venture's "Un-0" image system reportedly replicates conventional AI at a fraction of the energy — if even partly real, it reframes the data-center/power-crunch narrative that's driving local backlash and utility bills. [everyday-user/markets] (big claim — see §5) techcrunch
• Deal flow: Menlo raises a record $3B AI fund. Plus Adobe acquires Topaz Labs (image/video enhancement), XCures lands $46M Series B (medical-records cleanup), Netris pulls $15M Series A from a16z (neocloud go-live tooling). Capital keeps concentrating into AI infra + vertical apps. [founder/markets] Menlo · Adobe/Topaz · XCures

2. New-direction sparks

• Action-data as the substrate for "intuition." General Intuition's thesis — gameplay trajectories, not text or labeled robot demos, as the path to real-world agent priors — is a genuinely different data-flywheel bet than the VLA/world-model crowd, and cheap to scale. Worth watching whether action-data generalizes off-screen. techcrunch
• The verification inversion. "The Verification Horizon" argues that for today's coding agents, verifying a solution is now harder than generating one — every verifier is only a proxy for human intent. A quiet inversion of a 50-year CS intuition, and a tell for where the next moats sit. arxiv

3. Threads worth watching

• World models / spatial-physical understanding moved materially today — beyond the hallucination paper, PhysiFormer (world-space mesh dynamics ) and In-Context World Modeling for robotic control (cross-morphology generalization ) both landed, alongside General Intuition's raise. The field is converging on physically-grounded, reliability-aware world models — not just prettier video.

4. Contrarian watch

• Ensembling has a hard ceiling. Across 67 frontier models, routing/voting/mixture-of-agents accuracy can't exceed `1 − β` (the all-models-wrong rate) — and the usual pairwise-correlation diagnostic can't even measure β. Consensus says "stack more models"; the math says you're capped by shared blind spots. huggingface
• AI liability is coming for deployers. A German ruling holding Google liable for its AI-overview errors — agents treated as agents of whoever deploys them — cuts hard against the "the model did it" defense the industry leans on. simonwillison

5. Verification flags

• ⚠️ General Intuition $320M / ~$2.3B valuation — do not act on yet — needs primary source (single TechCrunch report, tagged Rumor). techcrunch
• ⚠️ Menlo Ventures $3B fund — do not act on yet — needs primary/SEC confirmation. techcrunch
• ⚠️ Databricks ex-chief "1,000x" power claim — do not act on yet — vendor claim, no independent benchmark. techcrunch
• ⚠️ XCures $46M Series B — do not act on yet — needs primary source (Crunchbase exclusive, tagged Rumor). crunchbase

Markets context only — not financial advice.

📡 Jin Miao Signals — Afternoon Brief · 2026-06-26

1. Top 5 — what actually matters today

• GPT-5.6 is officially out — but as a gov-gated "limited preview," and OpenAI is now publicly pushing back — What changed since this morning: the model is real and named (Sol flagship / Terra / Luna, with Terra ~2x cheaper than 5.5), AND OpenAI went on record that government vetting of who gets a model "shouldn't become the long-term default." A frontier release where access is rationed by the White House is the precedent everyone — users, developers, cyber defenders — should watch. openai · techcrunch
• The week's deal flow: AI ate the megadeal board again — Founder/markets read — the 10 largest U.S. rounds skewed to AI (Baseten infra, marketing, robotics), with biotech a distant second. The capital is still concentrating in inference infra + applied verticals, not new foundation labs. Specific amounts still need primary confirmation (see flags). crunchbase
• EO-WM: a physically-informed world model for Earth observation — A genuine world-model item on merit: reframes satellite forecasting as a partially-observed, weather-conditioned world-modeling problem with probabilistic (not collapsed-deterministic) futures. World models leaking into climate/EO is exactly the kind of paradigm migration worth catching early. paper
• 2,000 people, 6,000 attempts, $500 burned — nobody cracked the AI assistant — Tech-worker + everyday-user read: an email-driven prompt-injection gauntlet against an Opus-4.6 agent with a tight anti-injection prompt held. First concrete public evidence that hardened agent guardrails can survive a real adversarial crowd — the missing precondition for trusting agents with your inbox. simonwillison
• A Rust DB ran spatial joins on gaming-GPU ray-tracing cores and beat an H100 — Contrarian infra: SedonaDB repurposed consumer RT cores to outrun a datacenter GPU on spatial workloads. If the result holds, it's a non-obvious crack in the "you must rent H100s" assumption — and a spatial-compute angle worth a builder's attention. sedona

Top 5 spans policy/user, founder/markets, world-models, agent-security, and contrarian infra — only one OpenAI pick by design.

2. New-direction sparks

• Govern the action, not the agent. Two independent papers landed on the same non-obvious move: don't monitor an agent's reasoning — require independently-attested evidence at the moment of a consequential action. ActPlane: OS-level policy enforcement enforces it at the harness; Governing Actions, Not Agents formalizes institutional attestation as the governance model. Non-obvious because the whole field is racing on capability/alignment-of-reasoning; this says trust is built at the execution boundary, like every real institution already does.

3. Threads worth watching

• World models moved materially again today — EO-WM extends them from games/robotics into climate/Earth-observation forecasting (paper ), compounding this morning's "hallucination is predictable & preventable" result. The theme is widening its domain, not just deepening.
• Spatial / non-standard compute — Sedona's RT-core spatial joins (source ) is a faint but real signal that spatial workloads may not follow the datacenter-GPU default.

4. Contrarian watch

• Ensembling LLMs has a hard ceiling nobody reports. Across 67 frontier models, routing/voting/cascades/mixture-of-agents accuracy is capped at 1−β (the rate every model is wrong on the same query) — and the usual pairwise-correlation diagnostic can't even detect β. Consensus says "stack more models to win"; the math says co-failure caps you. paper
• Post-training quietly degrades values you trained in. "Helpfulness Hurts" shows coding-vs-helpfulness SFT/RL differentially erodes mid-trained compassion values — a non-consensus cost of the standard pipeline. paper Pair with the sycophancy-control work as a steering counterweight.
• Verification, not generation, is now the bottleneck for coding agents — and there's "no silver bullet." Cuts against the "agents will just self-verify their way to reliability" narrative. arxiv

5. Verification flags

• ⚠️ Weekly funding-round amounts — do not act on yet — needs primary source; the roundup is tagged [Rumor] and specific raise sizes/valuations aren't independently confirmed here. crunchbase
• ⚠️ "US govt will individually approve who gets GPT-5.6" — the broad gov-gating is corroborated (WaPo , OpenAI ), but the per-user "individual approval" framing is [Rumor] social — do not act on yet. reddit

Markets context only — not financial advice.