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

Manus returns independent after $2B Meta deal
08/11/26, 05:02 PM

Manus says it will return to independent operations after Meta’s $2 billion acquisition, and says affected users’ data generated from December 29, 2025 onward will be deleted during a two-day August 23–24 window in Singapore time. Manus’s user notice says users can back up data through August 23 and restore it from August 25; The Verge’s report says the separation follows China’s blocking of the deal. Teams evaluating hosted agents should treat ownership changes as a data-retention event, not just a corporate footnote, and preserve export paths—an operational extension of the archive’s earlier analysis of Meta’s agent acquisition.
Wire

LTX-2.5 renders 10 seconds in 6.8 seconds
08/11/26, 04:16 PM

LTX released LTX-2.5 as open weights and says a self-hosted setup with two GB200 GPUs can generate a 10-second 720p video in 6.8 seconds—faster than real time. The model page also lists a 16GB minimum VRAM and API prices from $0.09 per generated second at 720p; the speed claim is vendor-reported and tied to two high-end GPUs, not a laptop baseline. Creative and robotics teams should treat it as a local-control option to benchmark, not an automatic replacement for a managed API, extending the archive’s open-video model comparison.
Wire

IBM and Together AI sign $240M inference deal
08/11/26, 04:13 PM

IBM and Together AI signed a multi-year $240 million agreement for a large NVIDIA HGX B300 inference cluster on IBM Cloud, with expected availability in Q1 2027. IBM’s newsroom release says Together will use the cluster for open-source model inference and reports that its inference product already serves 400 trillion tokens monthly; the contract is a capacity commitment, not a published unit-cost guarantee. Builders should file the deal as another sign that open models are moving into dedicated, enterprise-scale serving infrastructure, alongside the archive’s analysis of hyperscaler capacity bets.
Wire

Mistral plans 1 GW of European AI compute
08/11/26, 04:10 PM

Mistral says it plans to build up to 1 GW of European AI compute by 2030, while making Regional Endpoints generally available and putting its Priority Tier into public preview with an uptime SLA. In its August 11 announcement, the lab says customers can choose Europe or the U.S. for inference, its long-term capacity commitments become European Compute Units, and the platform will host third-party open models starting with Z.ai’s GLM-5.2; the regional promise still allows limited, safeguarded transfers to outside subprocessors. The operator takeaway is less “Europe has solved sovereign AI” than “capacity, region, and model choice are becoming one procurement decision,” a sharper version of the archive’s earlier sovereign-AI infrastructure analysis.
Wire

Multilingual agents lose up to 18.4 points
08/11/26, 03:13 PM

OmnilingualGAIA2 finds an 8.8–18.4 pass@3-point gap across ten languages when it evaluates seven frontier and open-weight agents on a multilingual version of GAIA2. The arXiv paper says the gap concentrates on tool orchestration rather than quantitative reasoning, does not close with model scale, and is predominantly model-driven (55%) after accounting for a 6.4% translation-contamination floor. Teams deploying agents outside English should add localized tool-use tests before claiming parity, extending the archive’s agent-evaluation governance beyond English leaderboards.
Wire

TRACE attributes 72.7% of agent failures
08/11/26, 03:10 PM

TRACE reports 72.7% root-cause attribution and 82% end-to-end fix effectiveness on 60 agent dissatisfaction traces spanning three complexity tiers. The arXiv paper mines corrections, rephrasing, and abandonment cues from historical trajectories, then uses exploratory verification to distinguish missing context that needs CREATE from stale context that needs UPDATE, reaching 96% operation accuracy. Teams maintaining prompts, skills, tool descriptions, or knowledge bases should treat interaction logs as a context-debugging dataset and validate automated edits on their own failures, extending the case for a governed agent control plane.
Wire

TCS-BENCH gives proof verifiers 90% accuracy
08/11/26, 03:07 PM

TCS-BENCH evaluates LLMs on research-level theoretical-computer-science proof generation and reports a reference verifier with more than 90% accuracy against expert-labeled target-statement and proof pairs. The arXiv paper builds tasks from STOC, FOCS, and SODA papers, supplies the context needed for self-contained proofs, and checks generated proofs with a verification agent. Teams using LLMs for mathematical research should file this as evidence that proof generation and proof checking need separate benchmarks, complementing Claude’s reproducibility gate without proving that frontier models can independently produce correct research.
Wire

Amazon hides product names from order emails
08/11/26, 02:06 PM

Amazon has replaced some item-specific order-confirmation emails with category-only notices, such as “1 Hardware item,” while saying the change both directs shoppers to its app and reduces information shared outside Amazon’s properties. The company’s explanation arrives after Amazon sued Perplexity over shopping on its behalf and a judge sided with Perplexity earlier this month. Retailers building agentic commerce should file the trade-off: hiding product detail may protect a first-party relationship, but it also adds friction for customers and for assistants that depend on inbox context; Shopify’s AI-traffic data shows why machine-readable product facts are becoming distribution infrastructure.
Wire

Spotify keeps AI Personas out of recommendations
08/11/26, 01:16 PM

Spotify says it will start rolling out AI Persona labels in mid-September and keep music from those profiles out of 2 recommendation surfaces—editorial suggestions and personal recommendations—by default. Artists can disclose an AI identity now, but Spotify says it will also combine human review with AI tools to flag profiles that appear to represent photorealistic generated identities; it will initially review profiles that meet defined audience thresholds, and artists can appeal a “Likely AI Persona” label. Spotify’s rollout page and reporting on the policy make the boundary explicit: this labels the artist identity, not whether the music itself was AI-made. Music teams should treat profile provenance and recommendation eligibility as separate controls, extending the archive’s AI music analysis beyond generation quality.
Wire

Claude adds two invisible marks to AI output
08/11/26, 01:13 PM

Anthropic says it will mark Claude-generated content with two invisible signals: an imperceptible watermark woven into supported text and signed C2PA provenance metadata attached to supported .svg, .png, and .jpg files. Its support documentation says the text mark travels with copied text and may survive some edits, while reporting on the rollout says new Claude models will mark output from launch but support for existing models remains in progress. The marks apply at the model level across Claude surfaces and cloud deployments; detection details are still forthcoming. Teams shipping Claude output should treat provenance as a traceability layer—not proof of authorship—because metadata can be stripped and the text detector’s robustness is unreported, then retain their own model and version logs alongside the archive’s EU transparency checklist.