Notes RSS feed

The Wire — short items from the day's overflow, verdict follow-ups on earlier calls, and changelogs for the paper's maintained data assets. When a story earns a full argument, it runs as an article instead.

Wire

Ouroboros evolves its harness to 86.74%
08/12/26, 05:42 AM

Ouroboros, a coding-agent harness that changes its tools, prompts, context assembly, and core implementation through reviewed commits, reports an audited 86.74% on Terminal-Bench 2.1 and 90.69% on OSWorld-Verified with Claude Opus 5. The arXiv paper also describes a 161-day live deployment in which one agent evolves under governed human communication across seven surfaces, while warning that self-development can expand autonomy or select stronger model APIs. Engineering teams should treat the result as a harness-and-model claim requiring reproducible traces and immutable control boundaries, not as evidence that self-modifying agents are ready for unsupervised production; it sharpens the archive’s case that agent harnesses are part of the capability stack.
Wire

Team8 raises $365M for AI-native enterprise
08/12/26, 05:40 AM

Team8 closed $365 million in new capital—$265 million for its third fund and more than $100 million for follow-on investments in portfolio companies. The fund announcement says the fund targets Seed and Series A companies in cybersecurity, software infrastructure, fintech, and digital health, bringing Team8 to nearly $2 billion across eight funds since 2014; it also reports that 97% of surveyed organizations had started adopting AI agents while only 2.32 out of 5 was the average confidence in securing them. Builders should read the round as a capital signal for the enterprise-control layer—not proof that every funded category has demand—and compare new agent bets with the archive’s analysis of enterprise AI adoption economics.
Wire

Riot signs a $9.1B AI data-center lease
08/12/26, 05:38 AM

Riot Platforms signed a 20-year lease for 191 MW of critical IT capacity at its Rockdale, Texas campus with an unnamed frontier AI lab, worth about $9.1 billion over the base term. Riot’s results release says the deal could reach $16.1 billion with two five-year extensions, with 96 MW due in December 2027 and the full deployment by June 2028; Morgan Stanley is providing a $573 million interim financing facility. Operators should file the lease as evidence that approved power is becoming contracted AI capacity, while remembering that the first megawatt is still more than a year away and the tenant remains undisclosed; it extends the archive’s analysis of sovereign AI-factory economics.
Wire

CoreWeave backlog reaches $104B
08/12/26, 02:42 AM

CoreWeave reported $2.575 billion in second-quarter revenue and a $104 billion revenue backlog as quarterly capital expenditure reached roughly $6.4 billion. Its Q2 results show the infrastructure race’s strange arithmetic: a huge contracted future, but a $626 million quarterly net loss and a still larger financing burden behind the GPUs. Cloud buyers should treat backlog as a capacity signal, not proof of cheap or portable supply, extending the archive’s analysis of CoreWeave’s GPU-cloud bet with a closer look at delivery and counterparty risk.
Wire

Noosphere Labs raises $10.25M in stealth
08/12/26, 02:39 AM

Noosphere Labs, a Seattle-area stealth startup led by former Meta AI research director Kevin Carlberg, has raised $10.25 million to build what it calls “human-centered physical intelligence.” GeekWire’s report links the financing to an SEC Form D and says Trilogy Equity Partners led it with major participation from Madrona; the company has not yet disclosed a product. The useful signal is not a demo but the capital’s direction: operators tracking physical AI should watch for systems that help people interpret and act in the world, not just robots that replace a screen workflow, alongside the archive’s analysis of Prometheus’s physical-AI bet.
Wire

Energy Vault signs 1.25 GW AI deal
08/12/26, 02:36 AM

Energy Vault says it signed an agreement to deploy 1.25 GW of integrated power infrastructure for a hyperscaler AI data center, with $500 million–$600 million of expected revenue through the end of 2027. The company’s Q2 results and AI-infrastructure announcement pair batteries, generation, and control software with Caterpillar gas engines; Energy Vault also reported a roughly $2 billion backlog, though it says only about 40% should convert in the next 12–18 months. Builders should file this as evidence that “time to power” is becoming a bundled procurement category, alongside the archive’s analysis of Naver’s capital-heavy sovereign AI factory.
Wire

SWE-Bench ProMax caps agents at 41.2%
08/12/26, 01:41 AM

SWE-Bench ProMax gives the best tested coding agent a 41.2% resolve rate on 170 multilingual refactoring tasks averaging 11.4 modified files and 261.6 changed lines. The arXiv paper spans Python, Java, TypeScript, Go, C, C++, and Rust, and says its expert-curated tasks are designed to avoid flawed tests and training-set patch reproduction. Engineering teams should file the score as a reminder that multi-file behavior-preserving work remains unsaturated, then benchmark their own repositories rather than extrapolating from easier coding leaderboards, alongside the archive’s harness-focused evaluation analysis.
Wire

Mojo reaches 1.0 after 1,100 pull requests
08/12/26, 01:38 AM

Modular released Mojo 1.0 as a stable systems language after nearly 200 contributors landed more than 1,100 pull requests changing over 200,000 lines in its open-source standard library. The company’s release note says Mojo now underpins MAX and Modular Cloud in production, while its compiler and toolchain remain slated for open-sourcing later in 2026. AI-infrastructure teams can now assess Mojo as a longer-lived CPU/GPU programming layer, but should price in the still-closed toolchain and evaluate ecosystem portability before moving performance-critical code, alongside the archive’s analysis of AI runtime control.
Wire

ChatGPT desktop reaches five Linux releases
08/12/26, 01:35 AM

OpenAI launched its ChatGPT and Codex desktop app in preview for five named Linux releases: Ubuntu 24.04 and 26.04 LTS, Debian 13, and Fedora 43 and 44. The launch report says the worldwide release bundles ChatGPT, ChatGPT Work, and Codex, while downstream distributions based on those releases should also be compatible. Linux engineering teams can now evaluate OpenAI’s desktop workflow without a browser wrapper, but should treat the preview as a portability milestone—not a production endpoint—and keep a managed CLI or web fallback for developer environments, alongside the archive’s analysis of Codex’s hands-free workflow.
Wire

NVIDIA targets $500B for AI infrastructure
08/12/26, 12:38 AM

NVIDIA signed memorandums with six financial firms to create compute-financing platforms intended to mobilize more than $500 billion of third-party capital for AI infrastructure over time. The announcement carried by Brookfield says the plan remains subject to final agreements, so the figure is a target rather than committed cash; infrastructure buyers should watch whether financing converts into energized, customer-accepted capacity rather than count the headline as supply. The archive’s Firmus capital-stack analysis makes the same distinction between a funded project and a delivered accelerator-hour.