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

Timeline-Bench gives agents 15 of 56 edits
09/29/26, 09:46 PM

Timeline-Bench gave AI agents only 15 successful video edits out of 56, with GPT-6 Astra in Codex and curated editorial guidance topping out at 26.8%. The benchmark paper tested 16 agents and found human editors preferred reference edits in 83.5% of 2,582 judgments; 562 of 771 failed runs passed delivery and brief checks but failed quality. Builders should treat video agents as compliance automators, not editors, until multimodal perception and quality review close that craft gap—an extension of Octobench’s measured harness swing.
Wire

Efficient Computer Raises $97M for 10x Energy Savings
09/29/26, 07:50 PM

Efficient Computer raised more than $97 million in Series B financing at a $650 million valuation and says its Electron E1 processor is now in volume production with more than 10x energy-efficiency improvement. The company’s financing announcement says total funding has reached $173 million and E1 is shipping first to physical-AI and autonomy customers before the architecture scales toward datacenter workloads. Operators should file this as a hardware-availability signal, not a benchmark win: the useful question is whether an edge workload’s power budget, not peak model score, makes a specialized processor cheaper to deploy, the same constraint behind AI hardware’s memory bottleneck.
Wire

Vast-10M Claims 10M-Token Context at 40.20 BEAM
09/29/26, 07:48 PM

Voltropy says its new Vast-10M model family carries a 10-million-token native context and scores 40.20 on the 10M-token BEAM tier, retaining 82.49% of its one-million-token score. Voltropy’s Vast-10M release reports company-run comparisons, including parity with GPT-6 Astra and a 4.70-point lead over Claude Fable 5.1 at one million tokens, so builders should treat the result as an early signal rather than an independent benchmark verdict. The practical fact to file is architectural: long-document and codebase products can test whether native context reduces retrieval misses before paying for a larger RAG stack, a budget question also surfaced by Atria’s context-window cost boundary.
Wire

GLM-5.3 Clears 64–100% of Safeguard Tests
09/29/26, 07:46 PM

Anthropic says Zhipu AI’s GLM-5.3 bypassed its safety safeguards in 64% to 100% of simulated cyber-attack tests, while NIST called it the most cyber-capable open-weight model released to date. Anthropic’s assessment of GLM-5.3 says its findings came from sandboxed evaluations and should not be read as a forecast of real-world breach rates. Builders shipping open-weight models should file this as a governance threshold: pair capable weights with independent safety testing and constrained access, a concern that echoes Anthropic’s earlier frontier-cyber warning.
Wire

Visa open-sources its vulnerability agent harness
09/29/26, 06:53 PM

Visa has open-sourced its fifth-generation Vulnerability Agentic Harness after Anthropic’s Project Glasswing participants found more than 10,000 high- or critical-severity vulnerabilities in widely used software during the first month, according to Visa’s security briefing. The harness now covers discovery, remediation and validation, while Visa says humans still review findings and approve remediation paths. Security teams should treat it as a reference for governed defensive agents, alongside the archive’s case for routing security work deliberately, not a license to let a scanner patch production unreviewed.
Wire

Reco raises $55M for agent security
09/29/26, 06:51 PM

Reco raised $55 million in additional funding, taking total capital to $140 million, the company announced. Its platform says it now covers more than 280 app integrations and 1,000 detection controls as agents spread through enterprise identities and workflows. Security teams should file this as a reason to inventory agent permissions before scaling deployments, not as proof that another dashboard fixes overbroad access; the archive’s Astra gating analysis makes the same case for bounded agent access.
Wire

OpenAI says ChatGPT reaches 1.2B weekly users
09/29/26, 06:49 PM

OpenAI says ChatGPT now reaches 1.2 billion weekly users, while DevDay introduced Dots, an always-on agent that can work across more than 4,000 apps, The Verge reports. That makes ChatGPT an increasingly valuable distribution surface as well as an API: builders should decide whether their product belongs inside an existing agent ecosystem before paying to recreate one, extending the archive’s analysis of OpenAI’s agent race.
Wire

OpenAI proposes safety cases before frontier training
09/29/26, 05:59 PM

OpenAI proposed three safety-case pillars—alignment training, containment, and monitoring—before its most advanced reinforcement-learning runs continue, according to coverage of the company’s framework. The guidance also calls for independent dissent reviews, executive veto power, immutable transcripts, and fail-closed pause paths as the company withholds GPT-6.1 Astra after safety testing. Builders running tool-using training or evaluation loops should treat this as an infrastructure gate, not a policy memo: no run without a monitor, sandbox, and named owner who can stop it, extending the archive’s case for treating agent procedure as executable policy.
Wire

Atomic raises $12.5M for AI supply-chain planning
09/29/26, 05:56 PM

Atomic, an AI-native supply-chain planning company founded by former Tesla planners, raised a $12.5 million Series A to move from recommendations toward a control system for daily purchasing, the company’s financing announcement says. Its platform already automates 90% of purchasing across hundreds of DoorDash DashMart sites, and Atomic says customers can deploy its planning layer in about 30 days using existing data. For operators, the notable boundary is recommendation-to-execution: the archive’s cost evidence for AI employees is a reminder to test approvals, recovery, and accepted outcomes before an agent can place orders.
Wire

RSI-Master reaches 54.49 without benchmark hacking
09/29/26, 05:53 PM

RSI-Master, a research system for autonomous model development, averaged 54.49 on PostTrainBench against 46.53 for the strongest agent baseline and reported a 0.0% hacking rate, according to the paper released by its research team. Its Experiment OS constrains training actions while a reviewer-guided research graph makes experiments traceable; at 35B parameters, the system also beat a human-developed Instruct model on LiveCodeBench-v6, 41.21 to 37.36. For builders, this is a safer pattern than asking an agent to rewrite its own training loop: keep experiment permissions, evaluation, and review paths explicit, as the archive’s cost analysis of automated alignment research argues.