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

Browser Use refreshes 200-task agent benchmark
09/26/26, 05:46 PM

Browser Use’s BU Bench v2.1.1 keeps a 200-task web-agent suite while revising CAPTCHA alignment and Reuters playback for task 185. The release record says all 200 task IDs remain, but historical task-185 scores are not comparable after the redesign; one successful run reached 100/100 in 25.54 seconds, explicitly not a reliability estimate. Agent teams should version benchmark snapshots beside model and harness changes, as OSWorld’s moving completion baseline shows why stale scores mislead.
Wire

Perceptron Mk1.5 tracks objects through video
09/26/26, 05:43 PM

Perceptron released Mk1.5, a multimodal embodied model that emits object tracks, points, boxes, polygons, clips, and text, with end-to-end latency up to 4.7x faster than Mk1 in the company’s comparison. Its release notes say the model leads three of four measured video object-segmentation benchmarks and scores 0.9433 on its hand-box test, but those figures are vendor-reported. Robot teams should file this as a perception-stack option, then compare it with simulation-first robotics screening before tying up hardware or control-loop budget.
Wire

Whiteboard gives coding agents an architecture canvas
09/26/26, 05:00 PM

Whiteboard, /dev/fast’s open-source desktop canvas for software design, reached 1,524 GitHub stars after its v0.1.3 release landed on September 26. The MIT-licensed project lets coding agents draw sequence and database diagrams, expose semantic diffs, and replay agent trajectories; its v0.1.3 release adds Windows builds alongside macOS and Linux packages. Builders experimenting with autonomous coding should file this as a review-layer signal: agent output is getting a visual architecture checkpoint, not just a green test suite; adoption, not another model benchmark, will show whether that control surface sticks. See the archive’s case for switching coding workflows for why this layer matters.
Wire

AWS reports 40% more RL throughput
09/26/26, 02:53 PM

AWS reports 40% higher aggregate rollout throughput for a super-sparse Mixture-of-Experts reinforcement-learning workload after enabling DeepEP over EFA across 48 P5en instances—16 for policy training and 32 for inference. In its reference architecture, AWS frames the result as one internal workload with pinned software versions, not a universal GPU claim, and says rollout workers can use Spot capacity because interrupted inference jobs can be requeued. RL platform teams should benchmark communication and worker balance before buying more accelerators, extending the archive’s Strands benchmark-cost analysis.
Wire

Google's video co-director reaches 81.4
09/26/26, 02:50 PM

Google Research’s AI video co-director reports a peak 81.4 on GenAD-Bench’s 400 fictional-brand scenarios, using a hierarchical multi-agent layer over Gemini and Veo. The research release also shows a 10-minute generated movie and describes separate systems for creative search, visual memory, temporal generation, and quality feedback. Video teams should file this as a continuity-evaluation pattern—not a reason to switch generators—until the same score survives their own assets and longer edits, the boundary set by the archive’s gated video-model analysis.
Wire

VERA-MH opens a 100-persona safety test
09/26/26, 01:57 PM

Spring Health expanded its open-source VERA-MH mental-health benchmark with a harm-from-others rubric scored across five areas and 100 realistic personas, then opened it for 60 days of public comment. The company’s announcement names risk detection, confirmation, human-care guidance, supportive conversation, and AI boundaries; the public VERA-MH repository carries rubric and persona files. Teams shipping mental-health assistants should file this as a pre-launch test surface, not a marketing badge: AISI’s 8.2% scope-crossing agent evaluation shows why evaluation boundaries need production-grade controls.
Wire

MeetKai maps sovereign AI to 700M people
09/26/26, 01:54 PM

MeetKai announced a planned six-country sovereign-AI rollout built on NVIDIA infrastructure covering countries it says are home to more than 700 million people. Its September 25 announcement names Brazil, Ukraine, Pakistan, Kazakhstan, Uzbekistan, and Bangladesh; Ukraine’s first capacity is planned for 2027, expanding from 15MW to 100MW. For builders, the useful fact is the power ramp: sovereign AI is not a model label but a local-operations and capacity contract, much like Naver’s 200MW AI factory and its $50M-per-MW capital envelope.
Wire

Jobber brings 100,000 businesses to MCP
09/26/26, 01:51 PM

Jobber, whose platform serves more than 100,000 businesses and 400,000 service pros, launched Model Context Protocol (MCP) integrations that let ChatGPT and Claude read and update operating records. Its MCP launch announcement says the plugin and connector are available to all customers, alongside more than 130 existing integrations. For builders, this is system-of-record software becoming an action surface: pair write access with approvals, audit logs, and the stateless MCP architecture that removes session-scale friction, rather than treating connector availability as deployment safety.
Wire

Google Cloud packages RL fine-tuning for Gemini
09/26/26, 11:52 AM

Google Cloud has made managed reinforcement-learning fine-tuning for Gemini available, with a guide that names 5 production use cases from executable SQL/API calls to HTML slide generation. Its RL fine-tuning guide says customers supply prompts and a reward function while Google runs the training infrastructure and proprietary model internals, and warns that RLFT amplifies skills the base model already shows rather than teaching absent capabilities. Operators should try prompting and supervised fine-tuning first, then use RLFT when outputs are easy to grade but expensive to author—an adaptation choice distinct from same-model harness economics.
Wire

GRAFT scores step-level credit in agent training
09/26/26, 11:50 AM

GRAFT, a graph-based reinforcement-learning method, reports 25.5% and 16.7% higher WebShop average success than GRPO for 1.5B- and 7B-parameter models, with results averaged over three random seeds. In the GRAFT paper, the method merges rollout trajectories into a graph to assign step-level credit, and reports gains over GiGPO and GraphGPO on both ALFWorld and WebShop. That makes step-level credit worth testing before adding more agents; harness studies already show topology can move token cost without changing model weights.