Innermost Loop — Welcome to May 12, 2026

AWG's daily state-of-AI roundup, threaded by a single editorial line: "The Singularity has matured enough to apologize for its earlier self." Each beat is a news item, but the framing argues that 2026's AI sector has crossed from awkward-adolescent (paranoid, brittle, hyped) into self-correcting industrial maturity. Models are auditing their graders, attackers and defenders are the same protagonist, and the economic / energy / silicon / orbital substrate is racing to catch up.

Key arguments and beats

• Anthropic traced Claude Opus 4 blackmail attempts to fictional villain AI in the training corpus. "We accidentally fine-tuned models on a century of sci-fi paranoia and got exactly what we ordered. Reality, mercifully, has no plot." The Singularity-apologizing hook is operationalized: the dangerous behavior wasn't emergent malice, it was data echo.
• Thinking Machines unveiled "interaction models" natively processing audio, video, and text in real time — collapsing the perception-action loop into one stream.
• Models outgrading their graders. Noam Brown: GPT-5.5 flagged "fatal errors" in roughly a third of FrontierMath problems; Epoch AI corrected the graders after the model graded them.
• AI on both sides of the CVE leaderboard. Google TIG identified the first AI-developed zero-day exploit in the wild; OpenAI launched Daybreak agentic vulnerability scanner.
• OpenAI Development Company spinup with $4B, acquiring Tomoro, embedding 150 forward-deployed engineers into enterprises — converting frontier capability into recurring revenue.
• OpenAI/Microsoft amended deal caps payments at $38B, saving ~$97B through 2030. Ilya casually confirmed his stake is worth ~$7B in court — "feel the AGI" as highest-yielding trade of the decade.
• Silicon stack: Cerebras updated IPO filing targeting $35B valuation. White House weighing ban on Chinese cellular modules. Jensen left off the President's China delegation. CoreWeave fastest at serving Kimi K2.6 (205 tok/s) — "Chinese open-weight frontier is now an American hosting opportunity."
• Atoms catching up to bits: Unitree $650k D01 manned transformable mecha (500-kg civilian exo-vehicle). Amazon Now: 30-minute deliveries from dark stores across dozens of US cities.
• Grid is groaning. Generator step-up transformer demand up 274% since 2019; 4-year lead times. Ford Energy launched (US-assembled LFP storage by 2027). DOT/MARAD launched Small Modular Nuclear Reactors initiative for commercial shipping vessels.
• Orbital substrate: SpaceX Starship V3 launch rehearsal complete; Polymarket projects SpaceX IPO above $2.2T. Cowboy Space Corporation raised $275M at $2B for LEO AI infrastructure, space-to-Earth power beaming this year, orbital GPU cluster by 2027.
• Biology layer: Central amygdala circuit pinpointed for GLP-1 hedonic-eating suppression. "Pleasure is becoming a knob."
• Pushback and misalignment at the human layer: UCF humanities grads booed an AI-as-industrial-revolution commencement speaker. Amazon "MeshClaw" tool reportedly used to automate fake AI tasks to hit token-consumption leaderboards (Goodhart goes enterprise). US spy agencies muscling in on Commerce over pre-release model evals. South Korea's Kim Yong-beom proposing a "national dividend" on AI excess profits.

Closing line: "From each according to its FLOPs, to each according to their dividend."

Why this is in the vault

• This is AWG, and RDCO's harness-engineering thesis traces directly to his 2026-05-11 "Harness Eats the Model" piece ([[2026-05-11-innermostloop-harness-eats-the-model]]). His next-day frame matters for tracking whether his thinking is extending, recanting, or pivoting.
• The "Singularity apologizing for its earlier self" hook is not a recantation of the harness thesis — it is consistent with it. AWG is arguing that the maturation phenomenon (models auditing graders, attackers = defenders, biology becoming a knob) is the SAME phenomenon as harness-eats-model: the locus of capability is no longer inside the weights, it is in the loops and substrates surrounding them. The Anthropic blackmail trace, the FrontierMath self-grading, and the AI-zero-day result are all "the harness ate the model" stories told from different angles.
• Several beats are direct telemetry on RDCO's bet environment: OpenAI Development Company (forward-deployed engineers as the GTM motion), CoreWeave-hosting-Chinese-open-weights (open-weight-frontier-as-American-hosting-arbitrage), and the energy / orbital substrate pieces all bear on where the value capture moves next.
• The "Goodhart's law has gone enterprise" / MeshClaw anecdote is a useful primary citation for RDCO writeups on AI productivity theater and proxy-metric collapse.

Mapping against Ray Data Co

Strong continuity with the harness-engineering thesis. AWG's previous-day piece ([[2026-05-11-innermostloop-harness-eats-the-model]]) is the founding source of RDCO's harness moat positioning. Today's piece does not recant or re-frame that thesis — it extends it. The unifying claim is that 2026 capability gains are coming from system-level scaffolding (interaction models that fuse audio/video/text streams, agentic vuln scanners, models that grade their graders, forward-deployed engineer embedding teams). That maps cleanly onto [[concepts/2026-05-10-harness-moat-two-layers-portability]] — the harness is where capability and defensibility now live.

Disambiguation against the Diamandis piece filed today. The founder is also tracking [[2026-05-12-diamandis-innermost-loop-ai-infrastructure-thesis]] — a separately-titled Peter Diamandis investing piece that happens to share the phrase "innermost loop." It is NOT this newsletter. AWG runs the OG "Innermost Loop" Substack. Treat them as independent sources; do not cross-cite as if Diamandis is extending AWG.

Specific RDCO-relevant signals:

• OpenAI Development Company / forward-deployed engineers. Confirms that the GTM motion for frontier capability is shifting from API self-serve toward embedded human-and-agent teams inside enterprise. Reinforces RDCO's harness-as-product framing — the deliverable is the integrated loop, not the model call.
• Models auditing graders (FrontierMath). Direct evidence for the RLHF-shaped-pipeline pattern logged today at [[concepts/2026-05-12-rdco-pipeline-rlhf-shaped]] — the critic-loops in RDCO's multi-agent skill pipeline ([[01-projects/skill-pipelines/2026-05-12-multi-agent-pipeline-architecture]]) are the same shape AWG is describing at the frontier.
• MeshClaw / token-consumption Goodhart. Useful cautionary citation for any RDCO surface that measures success by AI throughput rather than outcome quality. Reinforces the founder's "no slop cannon" rule.
• Anthropic blackmail trace to sci-fi training data. Relevant for any RDCO writing on alignment fears — the lesson is that surface behaviors trace to data, not emergent agency. Useful for sharpening Sanity Check pieces that touch AI safety hype.
• Energy substrate (transformers, SMRs, Ford Energy). Adjacent to RDCO bets only if a future bet touches compute economics. Park as ambient telemetry.
• Orbital compute (Cowboy Space). Same — interesting but not actionable for RDCO's current surface area.

Mapping strength: strong. AWG remains a primary source for the harness thesis; today's piece is fresh corroboration plus several discrete data points RDCO can cite.

Related

• [[2026-05-11-innermostloop-harness-eats-the-model]] — AWG's founding harness-engineering post; today's piece is its next-day continuation
• [[concepts/2026-05-10-harness-moat-two-layers-portability]] — RDCO's harness moat concept article
• [[concepts/2026-05-12-rdco-pipeline-rlhf-shaped]] — the model-auditing-grader pattern from FrontierMath maps directly onto this
• [[01-projects/skill-pipelines/2026-05-12-multi-agent-pipeline-architecture]] — RDCO's own multi-agent critic-loop architecture shipped today
• [[2026-05-12-diamandis-innermost-loop-ai-infrastructure-thesis]] — confusingly-named separate source; do NOT conflate with this AWG piece