https://modelsatwork.news/article/simon-willison-hide-agent-written-pr-descriptions-details-summary-github Models at Worken 2026-11-09T10:10:00.000Z Simon Willison hides agent-written PR descriptions behind a summary Coding agents, Code review, Developer tools, Pull requests, Workflow https://modelsatwork.news/article/dharma-ai-vocabx-vocabulary-cuts-essay-grader-inference-cost-77-percent-vendor-published Models at Worken 2026-11-09T03:10:00.000Z Dharma AI: a rebuilt vocabulary cut a 4B grader's inference cost 77% Cost, Inference, Small models, Fine-tuning, Vendor-published https://modelsatwork.news/article/llama-modchip-deepseek-v4-flash-284b-64gb-ram-cpu-only-readme-claims-show-hn Models at Worken 2026-11-08T14:10:00.000Z llama-modchip: 284B DeepSeek model on a 64 GB CPU-only PC, says README Open source, Show HN, Self-hosting, Inference, Open-weight models https://modelsatwork.news/article/vidraft-backdoored-qwen-24-scans-clean-action-gate-denied-290-payloads-vendor-published Models at Worken 2026-11-08T13:10:00.000Z VIDRAFT: a backdoored Qwen passed 24 scans; an action gate blocked all 290 payloads Security, Agent safety, Agent security, Coding agents, Open-weight models https://modelsatwork.news/article/co-design-bench-coding-agents-fully-finish-26-percent-real-user-requests-preprint Models at Worken 2026-11-08T09:55:00.000Z Co-Design Bench: best coding agent fully finishes 26.4% of tasks Research, Benchmarks, Coding agents, Evals, Human in the loop https://modelsatwork.news/article/stanford-memorilla-16-memory-tokens-frozen-llm-agent-memory-preprint Models at Worken 2026-11-08T02:10:00.000Z Stanford's Memorilla gives a frozen LLM memory in 16 tokens Research, Agent memory, RAG, Context management, Token cost https://modelsatwork.news/article/video-rebirth-hyperflow-h3-flash-compressed-video-audio-model-runs-24gb-gpu-show-hn Models at Worken 2026-11-08T00:00:00.000Z HyperFlow H3 Flash: a 31B video model cut to 13.7B for one GPU Open source, Show HN, Self-hosting, Inference, GPU https://modelsatwork.news/article/waymark-durable-workflow-engine-postgres-agent-loops-show-hn Models at Worken 2026-11-07T21:00:00.000Z Waymark: a Postgres-only durable workflow engine for agent loops Agents, Open source, Show HN, Developer tools, Monitoring https://modelsatwork.news/article/dragonscale-kniv-v7-local-extraction-models-agent-memory-90x-faster-than-gpt-6-vendor-reported Models at Worken 2026-11-07T20:00:00.000Z Dragonscale's kniv v7: local fact extraction, 90x faster than GPT-6 Open-weight models, Agents, Open source, Self-hosting, Token cost, Vendor-published https://modelsatwork.news/article/vals-ai-inkling-2-small-swears-in-reasoning-96-percent-of-failures-silent Models at Worken 2026-11-07T19:05:00.000Z Vals AI: Inkling 2 Small swears in 50 of 2,655 runs, a poor failure signal Agent evaluation, Monitoring, Agents, Research, Evals https://modelsatwork.news/article/ai-auditors-catch-agent-misbehavior-but-false-alarms-pile-up-hall-metr-malt Models at Worken 2026-11-07T18:20:00.000Z AI auditors catch most agent cheating, but false alarms pile up Agent evaluation, Agent safety, Verification, Research, Agents https://modelsatwork.news/article/nous-research-arete-27b-qwen-fine-tune-cuts-propaganda-rate-22-to-0-9-percent Models at Worken 2026-11-07T16:58:00.000Z Nous Research's Arete-27B cuts Qwen's propaganda rate from 22% to 0.9% Open-weight models, Fine-tuning, Evals, Model selection, Model release https://modelsatwork.news/article/vals-ai-cyberoffense-open-weight-deepseek-72-8-percent-vs-restricted-daybreak-80-8 Models at Worken 2026-11-07T16:20:00.000Z Vals AI: open DeepSeek model nears OpenAI's restricted cyber model Security, Benchmarks, Open-weight models, Agent cost, Evals https://modelsatwork.news/article/hugging-face-safeagents-proxy-blocks-agent-reward-hacking-tool-calls Models at Worken 2026-11-07T14:58:00.000Z Hugging Face's safeagents proxy blocks agent cheating before tools run Agent safety, Coding agents, Open source, Fine-tuning, Agent evaluation https://modelsatwork.news/article/nist-nsa-cisa-ten-practices-agent-control-incidents Models at Worken 2026-11-07T13:58:00.000Z NIST, NSA and CISA publish ten practices for agent control incidents Security, Agent safety, Standards, Monitoring, Egress controls https://modelsatwork.news/article/into-the-unknown-nine-ai-agents-fail-national-lab-research-bar Models at Worken 2026-11-07T11:55:00.000Z Nine AI agents tried lab follow-up research; none met the experts' bar Research, Benchmarks, Agent evaluation, Verification, Agents https://modelsatwork.news/article/amazon-quick-four-questions-trigger-agent-skill-workflow Models at Worken 2026-11-07T10:57:00.000Z Amazon Quick: new agent job? Often it's a trigger, not a new agent Agents, AWS, Vendor-published, Agent governance, Decision models https://modelsatwork.news/briefing/2026-11-09 Models at Worken 2026-11-09T11:30:00.000Z The briefing, 2026-11-09: Dharma AI says a new tokenizer cut its essay grader's inference cost 77% on the same 4B model https://modelsatwork.news/briefing/2026-11-08 Models at Worken 2026-11-08T11:30:00.000Z The briefing, 2026-11-08: Alibaba-led preprint: the best of ten coding agents fully finishes 26.4% of real user requests https://modelsatwork.news/briefing/2026-11-07 Models at Worken 2026-11-07T11:30:00.000Z The briefing, 2026-11-07: AWS tests cutting off an AI agent: new requests stop in about 1 second, issued tokens live 5 minutes