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

Muse Voice Transcribe supports 70-plus languages
09/02/26, 02:34 AM

Meta’s Muse Voice Transcribe launches after training on 70-plus languages, with 25 extensively validated, diarization for 20-plus speakers, and a reported 3.1% final streaming word-error rate. Meta’s technical announcement says the model handles streaming ASR, code-switching, and endpointing in real time, but the headline rankings are Meta-reported and access and pricing remain unspecified. Voice-agent builders should treat 70 languages as a routing opportunity, not a global production guarantee; Mistral’s Voxtral launch in the archive is a useful reminder that transcription fit depends on workload, not language count alone.
Wire

BenchMIRT finds WMDP measures reasoning, not safety
09/02/26, 02:33 AM

Ai2’s BenchMIRT audit trained on results from 100 open-weight LLMs across 16 benchmarks and more than 34,000 questions, then found WMDP had no significant correlation with safety while tracking general reasoning in the opposite direction. Ai2’s technical release says BBQ, a social-bias test, also aligned much more strongly with general reasoning than safety. Evaluation teams should stop treating one benchmark score as a clean proxy for the construct it names; the archive’s cost study of automated alignment research makes the same case for inspecting what safety tests actually measure.
Verdict

Astra finds two zero-days; warning holds
09/02/26, 02:33 AM
Follows up on OpenAI Puts a 20% Compute Tax on Safety Monitoring

On August 19, the archive argued that OpenAI’s Astra warning made continuous monitoring a real compute and access constraint. OpenAI now says Astra scored 100% on ExploitBench and discovered and exploited two zero-day vulnerabilities in a modified test, according to TechCrunch’s report on the preview. The call holds, not flips: those are vendor-reported results without independent verification, but they make the earlier monitoring-cost warning harder to dismiss as hypothetical.
Wire

DeepSeek opens a 305B multimodal model
09/02/26, 01:37 AM

DeepSeek published the 305-billion-parameter V4-Flash-Vision-Exp as an open-weight multimodal model, adding a visual encoder and aligner to its agent stack. The official model card confirms the MIT license, tokenizer and inference reference, while coverage of the release reports the 305B scale; benchmark results remain vendor-run and the checkpoint is multi-GPU territory. Teams choosing a permissive base model should file this beside the archive’s open-weight adoption denominator: license freedom is useful only if serving cost and hardware fit survive production.
Wire

EU sends AI Act requests to 30-plus companies
09/02/26, 01:36 AM

The European Commission sent information requests to more than 30 AI companies under the AI Act, its first use of new powers after this summer’s agent-containment failures. Reporting on the requests says questionnaires cover model safety, independent review, monitoring and copyright; incorrect, incomplete or misleading answers can draw a fine up to €15 million or 3% of global turnover, while another account says recipients were not named and no investigation has begun. Builders should treat sandbox evidence, evaluation logs and copyright records as regulator-facing infrastructure, not internal paperwork; the archive’s containment analysis makes the same operational case.
Wire

Bitdeer pays $100M for 200 Texas acres
09/02/26, 01:35 AM

Bitdeer paid about $100 million in cash for 200 acres near its Rockdale, Texas facility, positioning the parcel for AI/HPC development. The company announcement carried by Business Insider says Milam County sites have 563 MW of interconnected capacity today and a planned 742 MW, while a second publication of the release confirms the acquisition; future power allocations and AI conversion remain company projections. For operators, land control is becoming infrastructure strategy: compare this powered-site bet with OpenAI’s Ohio lease economics before treating announced megawatts as usable compute.
Wire

Cheap models cost 52× less in ClickHouse test
09/02/26, 12:36 AM

ClickHouse tested 28 models on 201 production-derived analytics questions and found DeepSeek V4 Flash scored 11 points below Claude Fable 5.1 while the full run cost roughly $1 versus $52, in its agentic-analytics benchmark. That is a 52× cost gap for a bounded correctness trade, using a synthetic-but-truthful reconstruction of ClickHouse’s warehouse rather than a toy text-to-SQL set. Builders should replay representative warehouse questions before choosing a default model tier; the archive’s model cost-per-point analysis makes the same procurement lesson visible from another angle.
Wire

Wafer raises $40M for inference optimization
09/02/26, 12:35 AM

Wafer raised a $40 million Series A co-led by Marathon and Chemistry, with AMD Ventures, Wing, Outset Capital, Fifty Years, and Y Combinator participating, according to the company’s funding announcement. Wafer says its system learns traffic patterns and performance constraints, then tunes the model, inference engine, kernels, and hardware continuously instead of relying on one manual optimization pass before deployment. Operators should benchmark that promised serving efficiency against current GPU costs before buying new silicon; the archive’s model cost-per-point analysis shows why inference economics can matter as much as model choice.
Wire

Gemini agentic video cuts token use 88%
09/02/26, 12:34 AM

Google launched agentic video understanding across Gemini 3.7 Flash, 3.6 Flash, and 3.5 Flash-Lite, reporting up to 88% lower token consumption, 66% lower analysis cost, and 7% higher accuracy in its official benchmark report. Instead of processing every frame at a fixed rate, the feature searches relevant frames, audio, and transcripts on demand. Builders handling long-form video should test this mode against their current sampling pipeline before budgeting more context; Gemini 3.7 Flash’s earlier price cut makes the savings a deployment question, not only a model-quality claim.
Wire

John Deere puts 10 promises behind JD AI
09/01/26, 11:32 PM

John Deere is testing its JD AI assistant with select U.S. customers and backing it with a 10-point Farmer Data Commitment, according to The Verge’s report. The assistant uses field, machine, and operational data to answer questions about equipment settings, fuel use, and harvest timing; for vertical copilots, the frontier-model data-custody checklist now has a practical baseline: spell out data use, sharing, revocation, and promised value before model quality becomes the selling point.