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

WhatWorkedBench tests agents across 36 tasks
09/24/26, 07:53 PM

Researchers introduced WhatWorkedBench with 36 tasks, 30 data sources, eight workflow types, and 1,248 configuration records to test whether research agents can predict how component changes affect results, not merely optimize a final score, according to the preprint. With eight new measurements, Gaussian-process modeling lifted reported effect recovery from 0.632 to 0.698 in one Flash cohort and from 0.621 to 0.720 in another; results remain preprint evidence, not a production guarantee. Teams building experiment-running agents should log response surfaces and measurement budgets, not only best-run scores, extending the archive’s benchmark-cost analysis.
Wire

LiveKit buys Loophole Labs for faster agents
09/24/26, 07:51 PM

LiveKit acquired Loophole Labs and will fold its eight-person engineering team into work targeting AI-agent startup times below three seconds, according to the acquisition announcement. The deal follows LiveKit’s $100 million Series C at a $1 billion valuation in January, but the release gives no purchase price; its claim is a latency and utilization bet, not proof of production performance. Teams running voice or physical agents should measure cold-start time and idle capacity separately, extending the archive’s audio-agent budget analysis.
Wire

Google batch prediction cuts Scribd pricing by 50%
09/24/26, 07:49 PM

Google Cloud says Scribd classified more than 400 million documents—over 12 billion pages—with Gemini Enterprise batch prediction at a 50% discount to interactive pricing, according to its customer case study. The account says more than 99% of the corpus entered as native PDFs and the backfill ran in months, but it is a vendor-published case study rather than an independent cost benchmark. Builders with large offline document queues should model batch pricing before building always-on inference, alongside the archive’s Gemini Enterprise capacity analysis.
Wire

Dataiku launches an AI agent inventory
09/24/26, 06:48 PM

Dataiku launched Agent Management on September 24, saying fewer than one in five organizations maintain a complete, current inventory of their AI systems, in its product announcement. Agent Management scans agents across AWS Bedrock, Databricks, Google Vertex, Microsoft Copilot Studio, Salesforce Agentforce, Snowflake Cortex, Dataiku, and custom OpenTelemetry environments; general availability is planned for October with per-instance annual pricing and per-agent metering. For teams whose agents already cross vendor boundaries, GitHub’s pull-request control plane is a useful reminder that inventory must end in explicit action and approval paths.
Wire

NVIDIA maps protein complexes for 2,800 viruses
09/24/26, 06:46 PM

NVIDIA, Google DeepMind, and EMBL-EBI released predicted 3D structures for protein complexes from more than 2,800 viruses through the AlphaFold Database dataset announcement. About 30% of the interactions are completely new to science, and NVIDIA published the BioNeMo Structure Prediction Pipeline so researchers can run structure prediction in bulk. Scientific-AI builders should treat this as an open hypothesis-generation layer, not clinical evidence; Enveda’s clinical-proof gate keeps discovery capital separate from validation.
Wire

DOE backs 31 projects to unlock 23 GW
09/24/26, 06:44 PM

The U.S. Department of Energy selected 31 grid-improvement projects across 26 states, combining $5.25 billion in federal and recipient funding to make more than 23 gigawatts of capacity available, according to the DOE’s SPARK announcement. The package pairs $1.9 billion in federal money with $3.35 billion in cost share, reconductors more than 1,500 miles of lines, and deploys grid-enhancing technologies across nearly 21,000 miles. For AI builders, this is capacity relief rather than a data-center connection; Anthropic’s megawatt cost analysis shows why interconnection and power contracts remain gating items.
Wire

Google Beam cuts follow-up meetings by 21%
09/24/26, 04:56 PM

Google is shipping Beam units to six countries after an eight-week internal study found teams felt 50% more connected and needed 21% fewer follow-up meetings. In Google’s rollout report, the company says Beam integrates with Google Meet and Zoom and is expanding through 18 channel partners, with an Industrious network opening in October. Teams evaluating premium telepresence should treat the 21% result as an internal study signal—not a general ROI guarantee—and compare it with Apple’s earlier ambient-computing hardware bet before budgeting.
Wire

EmbodiedSWE tests agents on 30-minute robot tasks
09/24/26, 04:54 PM

EmbodiedSWE-Bench asks coding agents to solve long-horizon, dexterous-robot tasks requiring up to 30 minutes of continuous interaction. The paper reports that verified solutions can be expanded into diverse simulation trajectories, and a policy trained only on those generated demonstrations completed a task on a real robot—but the solutions remain specialized to individual task instances. Robotics teams should file the result beside Prometheus’s $41 billion physical-AI bet as evidence for using agents as supervision tools before trusting them as general controllers.
Wire

PASTABench finds agents hit 40.74% safety timing
09/24/26, 04:52 PM

PASTABench evaluated 16 language models across 1,139 multi-turn safety trajectories and found the best model achieved optimal intervention timing in only 40.74% of cases. The benchmark paper separates whether an agent should intervene, when it should intervene, and what risk it sees; it also warns that smaller models’ apparent safety can collapse when hazard keywords are neutralized. Teams shipping tool-using agents should read that gap alongside AISI’s 8.2% scope-crossing agent evaluation and test timing, not only final refusal rates.
Wire

Kontext raises $4M for agent authorization
09/24/26, 03:51 PM

German AI-security startup Kontext raised $4 million for software that checks an agent’s identity, assigned task, and target resource before allowing an action to run. SiliconANGLE’s report says the product supports Claude Code and Codex, offers an observe mode before enforcement, and starts paid team plans at $149 a month. Builders moving from credentials to autonomous tool use should test task-aware authorization in shadow mode first, extending the archive’s evidence-first agent governance playbook.