06-reference

innermostloop harness eats the model

2026-05-11·reference·source: Innermost Loop·by Alex Wissner-Gross
singularityagentic-stackharness-thesisfine-tuning-deprecationhyperscaler-capexgrid-powerspace-computelabor-replacement

"Welcome to May 11, 2026" - Alex Wissner-Gross

Why this is in the vault

Wissner-Gross's load-bearing aphorism today: "The Singularity is deprecating its own knobs." Two lines later he names the architecture shift outright - "the harness is eating the model" - and pins it to Hermes Agent overtaking OpenClaw on OpenRouter token volume by generating its own skills while OpenClaw users hand-write theirs. This is the cleanest single-paragraph external validation of the thin-harness/fat-skills thesis RDCO's whole COO architecture is built around, paired with an OpenAI fine-tuning API deprecation announcement (January 2027 cutoff) that effectively retires the alternative "adjust weights" path. Two paragraphs, two load-bearing signals.

Issue contents

Curation across six thematic clusters. Opening: "The Singularity is deprecating its own knobs." Closing: "Compounding, the only constant is compounding."

  1. Harness eats model. OpenAI winding down fine-tuning API with a January 2027 cutoff, logic being that as frontier models keep generalizing, weight-adjustment matters less. Sam Altman's tongue-in-cheek re-coinage: GPT-5.5 is a "genius," not just a "generalist." Cisco released open-source Model Provenance Kit, treating metadata and weights "like a model genome" to detect shared origins and tampering. Hermes Agent now #1 on OpenRouter token rankings, surpassing OpenClaw by generating its own skills while OpenClaw users hand-write theirs and lean on Opus for prompt-injection safety.

  2. Agents leaving the desk. Standalone AI solutions to open Erdos problems skyrocketing. OpenAI "Codex mobile" spotted, letting users keep working with Codex whenever the computer is awake. Codex told to "go off and make me $5" allegedly found an open-source security bounty, filed a legit PR, worked 22 hours across audits, netted $16.88.

  3. Silicon and grid sprinting. AMD's ROCm stack reportedly improved 75x in 14 days since DeepSeek V4, with only another ~7.5x needed to catch Nvidia B200. Maryland's Office of People's Counsel filed a FERC complaint over a $2B ratepayer tab for grid upgrades servicing out-of-state data centers ("breach of the ratepayer protection pledge"). Microsoft/G42's $1B Kenyan geothermal data center stalled over payment guarantees exceeding national resources. US hyperscalers piping Gulf data-center traffic out via fiber strung alongside crude-oil pipelines by an Iraqi telecom. Fusion's high-temperature superconducting wire scaled from 5,000 km to 1.5 million km over 15 years, with knock-on effects in transport, medical imaging, power, and data-center design.

  4. Data centers leaving the planet. SpaceX filed a trademark for "SpaceXAI" covering satellite-based data centers and orbital computing. Starship V3 stacked for the first time. Machine learning identified 10,000 new exoplanet candidates from TESS images. Utah's Hypercraft launched Razorback, an autonomous combat vehicle carrying 2,400 pounds, 280-mile range, exporting 38 kW to charge drones and run directed-energy weapons with no human onboard.

  5. Superintelligence making hidden agents shallow. After the White House's PURSUE initiative launch, Rep. Burlison asked MIT Lincoln Lab to preserve a 1952 reel-to-reel tied to early federal UAP investigations. Pentagon reportedly releasing 46 more UAP videos next week per Rep. Luna's request. Rep. Burchett: first PURSUE drop was "just the tip of the iceberg."

  6. Biosphere repair and human patching. Great Lakes river otters clawing back after decades of cross-border effort. Plant seeds can sense raindrop vibrations and wake from dormancy in response. MIT released FINGERS-7B, the first AI foundation model for Alzheimer's prevention, integrating lifestyle, clinical, genomic, and proteomic data from tens of thousands of at-risk individuals.

  7. Economy in AI's image. OpenAI + Anthropic projected to end 2026 with combined ARR exceeding Nvidia's revenue last year. Alphabet briefly overtook Nvidia in market cap. Billionaire tax refugees flocking to Incline Village at Lake Tahoe, "the nicest San Francisco neighborhood." McClatchy journalists withholding bylines from AI-spun articles. Women hold 83% of the 15 most AI-vulnerable jobs despite being 47% of the workforce. 40% of the 5,000 breaches Experian serviced last year were AI-powered. Tehran's record 70-day internet blackout couldn't slow the world's compounding.

Mapping against Ray Data Co

Verdict: strong. The harness-eats-model line is the load-bearing one.

Implication 1 - "The harness is eating the model" is now an Innermost Loop framing, which closes the loop on the cross-source cluster. This phrase, used by Wissner-Gross today, is the same architecture-shift Garry Tan named ([[commentary-tan-fat-skills-thin-harness-2026-04-14]]), Joao Moura named ("Agent Harnesses Are Dead. Long Live Agent Harnesses."), Harrison Chase named ([[2026-04-12-harrison-chase-harness-blog]]), Kieran Klaassen instantiated ([[2026-04-13-every-folder-is-the-agent]]), and Cobus Greyling/Akshay Pachaar/Paddy Srinivasan each described in their own vocabulary ([[2026-04-23-unhobbling]], [[cross-checks/2026-04-12-cross-check-agent-architecture]]). With AWG's curation-style aggregator beat now using the same frame, the cross-source cluster includes ten-plus independent authors AND the field's main daily-rapid-fire aggregator. This isn't a vault-internal thesis anymore; it's the public framing. RDCO's architecture - thin harness, skills-as-folders, fat vault, model-portable - is on the right side of where the consensus has now visibly settled.

Implication 2 - OpenAI fine-tuning API deprecation is the deepest validation of the harness/skills bet to date. The argument we've been making (RDCO doesn't fine-tune, the skills + vault + harness pattern beats it for solo-operator use cases) just got institutionalized at the API provider level. OpenAI is telling customers: as frontier models keep generalizing, adjusting weights matters less. The January 2027 cutoff gives existing fine-tuning customers an 8-month deprecation window, which is the kind of timeline that forces the entire ecosystem to re-architect around prompt + skills + harness. For RDCO this means: (a) the "what if competitors fine-tune their way past us" worry can be downgraded - the API surface for that path is being retired by the largest provider, (b) any future RDCO bet that touches "agent infrastructure for builders" has a structural tailwind because the alternative is being deprecated, (c) the case for investing further in skill quality and the vault-as-memory pattern just got reinforced by the largest model provider's own roadmap decision.

Implication 3 - Hermes Agent #1 on OpenRouter by token volume, generating its own skills. This is the next-rung evolution. RDCO is currently in the "Ray uses /skillify after a workflow succeeds to convert it into a permanent skill" pattern - human-supervised skill genesis. Hermes is doing it autonomously. Worth a low-priority research item: read the Hermes paper/post on how its self-generated skills compare to human-authored ones in quality, persistence, and prompt-injection safety. Specifically what's interesting is that OpenClaw users still hand-write skills "and lean on Opus for prompt-injection safety" - the AWG framing suggests Hermes self-generated skills may be trading some safety for autonomy. If true, that's a calibration anchor for how aggressively RDCO should move toward autonomous skill genesis vs. keeping the /skillify human-gate.

Calibration signal - "go off and make me $5" Codex bounty proof point. $16.88 net, 22 hours, real PR landed. Pairs with [[2026-05-01-alpha-vantage-collison-agent-as-customer-evidence]] (agent-as-customer pivot) and [[2026-04-09-every-four-ai-agents]] (25-person company on four agents). This is the cleanest single-line evidence of an agent autonomously generating revenue against an external bounty system. Useable when articulating the RDCO L5 thesis externally - it's a concrete dollars-and-hours story rather than a hypothetical. Not Sanity Check material by itself (would be derivative), but supporting evidence for any piece on agent autonomy that uses original framing.

Specific items worth surfacing into other workstreams:

Watch-channel pattern check: This is the first STRONG-mapping Innermost Loop in the last several issues (May 5, May 6 both landed medium). The "harness eats model" framing flip plus the OpenAI fine-tuning deprecation pulls this back into load-bearing territory. The channel is earning its keep again - no deactivation signal.

Where this DOESN'T extend: no new contact, no new framework name beyond AWG borrowing existing harness-thesis vocabulary, no operational tactic to ship today. Strength comes from the cross-source cluster closing, not from any single novel claim.

Related