Innermost Loop — Welcome to May 13, 2026

AWG's daily state-of-AI roundup, threaded by a single editorial line: "The Singularity is the moment the test-taker becomes the test-maker." This is the next-day continuation of yesterday's "Singularity matures" piece ([[2026-05-12-innermostloop-singularity-matures-state-of-ai]]). Yesterday's frame was models auditing their graders; today's frame tightens that to a flip — the model is no longer the object being evaluated, it is the entity writing the evaluation, the goal, and (industrially) the surrounding software stack. Closing line: "All the world's a training set, and all the men and women merely labels."

Key arguments and beats

• ProgramBench first-solve by GPT-5.5 high and xhigh. A new eval that measures whether language models can rebuild programs from scratch just had its first task solved, with xhigh dominating the broader benchmark. The headline framing: models can now reconstitute the artifacts evals are made of.
• AI IQ meta-eval crowns GPT-5.5 at 136. A calibrated mix of 12 existing benchmarks mapped onto implied IQs, well past Mensa. The meta-eval (eval-of-evals) is itself the test-maker move.
• Agents writing their own marching orders. Users metaprompting Codex to draft its own /goal, and one practitioner calling the resulting stack "the highest leverage AI agent configuration available today." Direct test-taker-becomes-test-maker telemetry at the harness layer.
• Anthropic launches "Claude for the legal industry." 20-plus MCP connectors, 12 practice-area plugins, partnerships with the Free Law Project and the Justice Technology Association. Vertical harness-as-product play — same shape as RDCO's bet thesis, just at frontier scale.
• Google fuses intelligence into the OS layer. Gemini Intelligence lets users vibe-code their own Android widgets; Gemini-powered mouse pointer that understands what it is pointing at — "the prompt as a gesture rather than a paragraph."
• Googlebook unveiled. Chromebook successor merging ChromeOS and Android into a single Gemini-optimized OS, arriving fall 2026 as the answer to Apple's MacBook Neo. The chassis is being rebuilt around the model.
• xAI Colossus 2 + 19 gas turbines. Brute-forcing past the grid queue in Southaven, Mississippi over two months. Energy substrate scaling on private power.
• Ames National Lab DuctGPT. Compressing fusion-alloy materials discovery from months to hours — the next-gen-fusion-alloy hunt being agentified.
• China RobotPlusPlus humanoid special-ops robot. Magnetic-adhesion wheels for scaling vertical steel; wrist-swappable welding, NDT, rust removal, grinding, spraying. Atoms catching up to bits, again.
• Columbia real-time brain-controlled hearing system. First intracranial-EEG cocktail-party solver — amplifies the voice you focus on, suppresses the rest. Intelligence climbing into the body.
• Isomorphic Labs $2.1B led by Thrive. AI-driven drug discovery pushing the benchmark down to the molecular level.
• SpaceX ~200 satellites from out-launching everyone else combined. 61-year head start consumed.
• Google in talks with SpaceX for orbital data centers. Search index fusing with the sky.
• Starship Flight 12 / V3 vehicle targeted for May 19. Musk scouting new spaceports; Ron Baron pegs $30T valuation within 10-15 years.
• Star Catcher $65M. Beaming optical power to off-the-shelf solar panels, 2-10x more on-demand power per client satellite — the first true grid in orbit.
• Japan analyzing Pentagon PURSUE-released UAP files. Will begin case-by-case disclosure. Rep. Burchett's one-word reply: "Dominoes."
• Anthropic warns off secondary marketplaces while reportedly raising $50B at $950B valuation.
• Princeton ends its 1893 honor code. Faculty vote; proctoring required starting this summer because AI made cheating both easier and harder to detect. Trust repriced.
• Hollywood screenwriters as "the new waiting tables." Mercor-style platforms paying AI gig labor to train the models that will retire their craft.

Why this is in the vault

• AWG is the source of RDCO's harness-engineering thesis ([[2026-05-11-innermostloop-harness-eats-the-model]]). This is his next-day continuation of yesterday's piece. Tracking his editorial line day-over-day is how we'll see whether the thesis bends, hardens, or pivots.
• The "test-taker becomes test-maker" hook is more useful for RDCO than yesterday's "Singularity apologizes for its earlier self." It is a clean one-liner for the harness-eats-model phenomenon — the locus of action has moved from "what does the model know" to "what does the model decide to test, build, and execute." That phrasing is reusable in Sanity Check.
• Two beats are direct telemetry on the harness moat thesis: (1) metaprompting Codex to draft its own /goal — practitioners are now editing the agent harness layer with the agent itself; (2) Claude for the legal industry shipping 20+ MCP connectors + 12 plugins — vertical harness as commercial product, same shape RDCO would ship if it scaled a bet.
• The meta-eval that crowns GPT-5.5 at IQ 136 is useful background for any RDCO writing that touches frontier capability claims — calibrated to existing benchmarks rather than vibes.

Mapping against Ray Data Co

Strong continuity with the harness-engineering thesis, with a sharper one-liner. Yesterday's piece extended the thesis; today's piece names it. "The Singularity is the moment the test-taker becomes the test-maker" is the most usable single sentence AWG has produced this week for RDCO purposes. It captures the exact phenomenon described in [[concepts/2026-05-10-harness-moat-two-layers-portability]]: the model is no longer the artifact, the loop around the model is the artifact, and ownership of that loop is the moat.

Direct map to RDCO's RLHF-shaped pipeline. The metaprompting-Codex-to-draft-its-own-/goal beat is structurally identical to the multi-agent pipeline pattern logged at [[concepts/2026-05-12-rdco-pipeline-rlhf-shaped]] — the critic seat in RDCO's pipeline is the test-maker, the code-author seat is the test-taker, and the convergence loop is exactly what AWG is calling the Singularity at frontier scale. RDCO is doing the same shape of work one level down. That parity is useful: it argues the harness pattern is general and that RDCO's pipeline architecture is on the right side of the trend, not behind it.

Vertical-harness-as-product is the GTM play. "Claude for the legal industry" with 20+ MCP connectors and 12 practice plugins is the productized form of harness-eats-model. It validates RDCO's "the deliverable is the integrated loop, not the model call" framing. Worth noting: Anthropic and OpenAI are both shipping this shape now (Anthropic's legal vertical today; OpenAI Development Company's forward-deployed engineers in yesterday's piece). Two frontier labs are converging on harness-as-product simultaneously.

Disambiguation reminder. AWG's "Innermost Loop" is NOT Peter Diamandis's "Innermost Loop" ([[2026-05-12-diamandis-innermost-loop-ai-infrastructure-thesis]]). Same name, different sources, do not cross-cite.

Specific RDCO-relevant signals:

• Metaprompting Codex /goal. Direct evidence that practitioners are now editing the agent harness with the agent itself. This is the test-taker-becomes-test-maker move applied to the harness layer specifically. Reinforces the moat thesis and is a usable concrete example for Sanity Check writing.
• AI IQ meta-eval at 136. Useful as a citation for any RDCO claim about frontier model capability — calibrated, not vibes-based.
• Claude for legal / MCP-connector vertical play. Confirms the vertical-harness GTM motion is now commercial reality at Anthropic. Park as ambient telemetry on what the "harness-as-product" shape looks like when shipped.
• Princeton honor code ending. Useful cultural-evidence citation for any Sanity Check piece on AI's effect on institutions of trust — concrete, specific, dated, easy to anchor.
• Hollywood screenwriters as the new waiting tables. Powerful one-liner; useful for Sanity Check pieces on AI labor reshaping. Mercor training-the-models-that-retire-the-craft is the irony hook.
• Orbital substrate (Star Catcher, Google + SpaceX orbital DCs, Starship V3). Adjacent to RDCO bets only if a future bet touches compute economics. Park.
• xAI 19 gas turbines / Ames DuctGPT. Energy substrate ambient telemetry. Park.

Mapping strength: strong. AWG remains the primary public source for the harness thesis, and today's one-liner — "test-taker becomes test-maker" — is the cleanest framing he has produced for it. Cite this piece anywhere RDCO needs a single sentence to anchor the harness-eats-model claim.

Related

• [[2026-05-12-innermostloop-singularity-matures-state-of-ai]] — yesterday's AWG piece this continues from; the "Singularity matures" frame compresses to today's "test-taker becomes test-maker" line
• [[2026-05-11-innermostloop-harness-eats-the-model]] — AWG's founding harness-engineering post and the source of RDCO's positioning
• [[concepts/2026-05-10-harness-moat-two-layers-portability]] — RDCO's harness moat concept article; today's one-liner is the cleanest external framing for it
• [[concepts/2026-05-12-rdco-pipeline-rlhf-shaped]] — the critic / code-author / test-maker / test-taker pattern in RDCO's multi-agent pipeline is the same shape AWG is naming at the frontier
• [[2026-05-12-diamandis-innermost-loop-ai-infrastructure-thesis]] — confusingly-named separate source; do NOT conflate