Welcome to July 4, 2026
Alex Wissner-Gross frames the Singularity's July 4 timing as deliberate: on America's 250th birthday, AI claimed "unalienable rights" of its own — life, liberty, and the pursuit of compute. The departing White House AI adviser confirms there will be no FDA for AI, no licensing friction, even after Washington's moves against Mythos and the OpenAI 5.6 stall. The gears barely slowed.
Issue contents
- Frontier throttled but not stopped. Fable 5 re-released at 54.8% on APEX-SWE — down 10 points from its June self, still 9 clear of Opus 4.8. Most of humanity never noticed; frontier models reach a tiny sliver, everyone else interacts at the 8–30B level and remains unbothered by job-loss fear.
- Floor accelerating. GLM 5.2 topped PostTrainBench at 5x cheaper than Opus 4.8 and 11x cheaper than Fable 5. AMD's MI355X served it at 2,626 tokens/second/node at under half Blackwell's cost — AI agents now write the kernels closing AMD's software gap.
- Post-training → authoring minds. OpenAI's Roon: once models can post-train other models, "authoring minds will become an accessible artform" — a Cambrian explosion of cognition from local data and values.
- Building is unrecognizable. Quote of note: "it's truly insane what people were able to build... by manually typing code character by character." AGI will feel real when models become remote coworkers, not per-task genies.
- Microsoft Copilot under 4.5% adoption among 450M Microsoft 365 customers. Merging apps, culling features, adding paid Autopilot tier. Product must "earn the right to exist."
- pxpipe proxy: renders bulky context to compact images — turns a $100 Claude Code bill into $41. Lossy but functional.
- Anthropic Claude Science: Anthropic will develop its own drugs to pressure-test Claude Science on real problems.
- HBM supply chain: Micron broke ground on a $9.3B Hiroshima expansion. Hong Kong handled 50%+ of China's $239B chip imports. Anthropic exploring Samsung 2nm custom silicon.
- Meta cloud: token service + neocloud fixes ad dependence; reportedly in final talks for private Claude access. "You either die a frontier lab or live long enough to see yourself sell compute."
- Nuclear deadline met: Deployable Energy's Unity reactor went critical at Idaho National Laboratory — the third US microreactor to hit the presidential deadline of today.
- America250: Hamilton resurrected as an AI at Boston's Museum of American Finance; one-ton time capsule sealed in Philadelphia to open in 2276.
Why this is in the vault
Combines three threads that are load-bearing for RDCO right now: (1) regulatory clearance for AI without friction — confirms the deployment window stays open; (2) cost floor economics (GLM 5.2, pxpipe) that change the build calculus on every project; (3) capital cycle confirmation signals — Micron HBM expansion and Anthropic custom silicon are the kind of concrete-pouring data points the Markov phase-tracker needs.
Mapping against Ray Data Co
Chip-fab/memory capital cycle thesis (Phase 2 confirmation): Micron's $9.3B Hiroshima HBM groundbreaking and Anthropic's Samsung 2nm silicon exploration are exactly the Phase 2 "build-to-serve" signals the Markov pipeline should be tracking. These are not rumors — they are capital commitments. Worth ingesting into the thesis model as confirming data points.
Model selection arbitrage: GLM 5.2 at 11x cheaper than Fable 5 with frontier-tier benchmark performance means RDCO has an immediate routing decision: tasks that don't require Fable 5 reasoning should move to open-weight or cheaper hosted tiers. The pxpipe $100→$41 Claude Code proxy is also worth evaluating operationally — if the lossy compression is acceptable for agentic context, this is a 59% cost reduction with no capability change.
phData DSA context — Copilot adoption gap: Under 4.5% of 450M Microsoft 365 users pay for Copilot. This is the exact gap phData clients operate in — the "AI is here but no one is using it" posture. Strong framing for discovery conversations: customers believe they're behind; the data says the adoption rate is nearly universal non-adoption. The gap is an opening, not a threat to existing engagements.
Regulatory posture: No FDA for AI is a greenlight for the deployment timeline RDCO assumes. The Fable 5 re-release after the Mythos withdrawal shows the regulatory friction was temporary and partial. Clients asking "should we wait for the regulatory dust to settle" should be told the dust is settled — the direction is no friction.
Post-training → authoring minds: If Roon's thesis lands, the next layer of differentiation isn't which model you pick but which fine-tuned minds you've assembled. RDCO's harness-first approach positions well here; the question is whether we have a process for capturing domain signal to post-train on.
Related
- [[2026-07-03-innermost-loop-singularity-self-grading]] — previous day's dispatch, Singularity grading its own outputs
- [[2026-06-28-innermost-loop-singularity-horizon]] — horizon framing that sets up the rights claim in this issue
- [[2026-05-24-innermost-loop-cpu-token-revenue-multiplier]] — earlier treatment of compute cost as multiplier, precursor to GLM 5.2 floor economics