2026-08-18

1. Two papers stop proposing harnesses and start measuring them — and one of them undercuts the case (1.75)

• Beyond Final Scores: A Systematic Evaluation of Agents for Long-Horizon AI Research and Development (arXiv:2608.13417) runs 7 frontier models over 36 long-horizon AI-R&D tasks with rule-based within-run metrics — Solution Framing, Execution, Feedback Control — instead of final scores. Verdict: agents "operate more like engineering optimizers than fully autonomous researchers"; their best solutions adapt or combine established techniques, and genuine methodological novelty remains rare (source).
• The mechanism behind this page's long-running complaint is named: distinct process bottlenecks sit behind similar final outcomes. Two systems can score alike and fail in different places, so a score cannot attribute a gain to the harness, the model or the run.
• It also tests an assumption every self-improving loop here makes. Experience reuse "can help or mislead" later decisions — measured, not monotone. DarwinX: Evolving Agent Harnesses Through Natural Selection (arXiv:2608.07545)'s archive of lineages and Spatial Memory Agent: Experience-Grounded Procedure Memory for Spatial Intelligence (arXiv:2608.12743)'s scored lessons stand as things those loops achieved on their own benchmarks, not as a general property.
• Apodex Discovery: Reality Benchmarks and Environments for Evaluating and Building Discoverative Artificial Intelligence (arXiv:2608.11341) arrives at process metrics from the opposite direction — problems that "rarely arrive in an executable or verifiable form", scored by HDS6 across Tools, Repair, Alternatives, Coherence, Evidence and Scope, independently of final-task success. 561 industries surveyed, 423 problems assembled, 20 released.
• And it contains the counter-evidence. The same biomedical environment lifts GPT-5.5 by 2.5 points and GPT-5.6-sol by 7.6 points against the same closed-book backbone. If the harness were the capability, a 3× spread across two backbones would not appear.
• Why it matters: after today the honest reading of this wiki's own harness thesis is that a harness is a multiplier on the model, not a substitute for it — and nothing collected in three weeks could have told those two claims apart, because everything reported a final score.
• → Eval Harness Configuration, Agents (LLM Agents)

2. PORTS-Pike: NVIDIA is the chip vendor, the guarantor, an owner of the landlord, and the exclusivity clause (1.09)

• OpenAI contracted approximately 8 GW-IT at the PORTS-Pike Technology Campus in Pike County, Ohio — the site of the Portsmouth Gaseous Diffusion Plant, a former uranium enrichment facility — with SB Energy, NVIDIA and the U.S. Department of Energy. SB Energy builds, owns and operates on a 20-year lease; first tranche 4.25 GW with an option for 3.75 GW more, phased from 2028 (source).
• NVIDIA provides up to $105 billion in financing and puts $1.5 billion of equity into SB Energy, joining SoftBank Group and OpenAI as investors in the landlord. Its own release headlines the campus as hosting NVIDIA AI compute exclusively.
• Local terms: 35,000 construction jobs over six years to 2032, 2,500 operating jobs, a $40 million OpenAI community fund beside SB Energy's own $40 million, and $84 million in Codex credits for Ohio college students. Cooling is closed-loop and air-cooled.
• The two capacity figures are not a discrepancy: 10 GW of new generation is stated to yield the 8 IT-GW of campus capacity. Reporting that says the site "could grow to 10GW" is quoting the generation number.
• Why it matters: this wiki recorded the same structure once before at a tenth the size — Google backstopping lease payments on the TPUs it sold into Anthropic's $35 billion SPV. At $105 billion, a compute vendor financing demand for its own product is no longer an exception, and nothing read states what the exclusivity actually binds.
• → NVIDIA, OpenAI

3. Two labs fund the same research question 200× apart — and the framework neither of them mentions (1.56)

• OpenAI named the winners of its policy call: 14 projects, $1 million collectively plus up to $1 million in model credits, from more than 400 responses, spanning the American Enterprise Institute, the Progressive Policy Institute, the Tax Foundation and the Nuclear Threat Initiative, plus Europe, Brazil, Singapore and South Korea (source).
• This wiki holds the comparison: Anthropic's Economic Futures Research Fund (2026-07-22) committed $200 million at $5M–$30M per grant. Same instrument, same stated subject, four weeks apart — one Anthropic grant is at minimum five times OpenAI's entire programme.
• Separately and more consequentially: Pax Silica, a US-led alliance for coordinating AI supply chains and export controls, has existed since December 2025; the EU joined on 2026-06-03, committing to buy at least $40 billion of American AI chips. It appears nowhere else in this wiki, and surfaced only because three Trivium China items about it landed in one prefetch batch (source).
• Why it matters: AI Governance has recorded export controls per model — Fable 5's restriction and restoration, the H200 approvals — because its intake was model-release coverage. A standing framework produces no release, so it produced no capture for eight months while every one of those decisions was taken inside it. Same failure shape as the Alignment Science blog: adjacent to everything tracked, read by nothing, because an absence raises no error.
• → AI Governance

4. A third way to stop wasting reasoning tokens: don't start (1.49)

• Knowing When to Quit: Diagnosing and Training LLMs to Abort Futile Reasoning (arXiv:2607.29211) names futile reasoning — expensive, semantically void reasoning on beyond-capability tasks — and reports universal capability overreach with systematic miscalibration between capability and behaviour. The dominant failure is specious reasoning: output that looks valid, contains subtle errors, and escalates with task difficulty (source).
• Its fix, CaRL, shapes reward to prefer refusal over futile reasoning and converts past failures into refusal supervision.
• Yesterday's two mechanisms both assume the model eventually answers — Gambit reallocates a live budget, the full-bandwidth transformer reduces what a step needs. This one says that on some inputs the correct budget is zero, which is a different claim about where the waste is.
• The number it needs is the one it does not publish. An objective that pays a model to refuse has an obvious degenerate solution, and "without sacrificing utility" carries no refusal rate beside it.
• Why it matters: four mechanisms in two days now change how many tokens an answer costs. They compose only if the tokens they remove are disjoint — none of the four cites another, so nothing establishes that any two of them are.
• → Test-Time Compute (Inference-Time Compute Scaling)

5. MiniMax resolves the contradiction this wiki refused to publish — and the licence went the permissive way (1.40)

• Five days ago MiniMax recorded that Music 3.0's release date, licence and weights availability were all contested, and declined to create a model page rather than publish four unknown rows and a contradiction. Today's Chinese-lab rotation closed it: two independent accounts agree the weights went up on 2026-08-13 — ~11.1B parameters, five-minute songs in a single generation from lyrics plus a structured caption, 32 kHz 16-bit stereo, on Hugging Face, GitHub and ModelScope with ComfyUI support the same day (source).
• The MiniMax-Music3 Community License carries no territorial exclusion. MiniMax H3, from the same company eleven days earlier, excludes the US, EU, UK and South Korea by territory clause.
• Why it matters: the 08-13 entry said the licence, not the announcement, was the fact worth waiting for — and it resolved against the pattern. "Open weights from MiniMax" is a per-release decision, not a company policy, which is exactly why this wiki records a licence per model. What is still missing is unchanged: no controlled listening benchmark from anyone, and no explanation of the 2026-07-16 date a model directory gives. Both stay in ## Conflicting Reports.
• → MiniMax Music 3.0 (new), Open-Weights Policy Fight