AI Signal Daily

Copilot, Claude, Webwright, NVIDIA and agent costs

Send us Fan Mail [https://www.buzzsprout.com/2614078/fan_mail/new] COPILOT, CLAUDE, WEBWRIGHT, NVIDIA AND AGENT COSTS Today’s episode follows AI responsibility as it slides down the stack: default model routing, long-document training, Claude in government networks, agent costs, web-agent scripts, voice models, local hardware, and synthetic bug reports. 1. Copilot and the risk of default model selection [https://the-decoder.com/why-you-shouldnt-leave-model-selection-on-default-in-copilot-gemini-and-other-ai-tools] 2. ByteDance Seed trains LMMs through question answering [https://the-decoder.com/bytedance-study-finds-that-asking-lmms-questions-beats-making-it-transcribe-text-for-long-document-training] 3. Hassabis, LeCun and the intelligence debate [https://the-decoder.com/deepminds-hassabis-sees-humanity-in-the-foothills-of-the-singularity-while-lecun-says-current-ai-isnt-intelligent] 4. Anthropic, Claude and the NSA [https://the-decoder.com/anthropic-may-keep-supplying-claude-to-the-nsa-despite-being-flagged-as-a-supply-chain-risk-by-the-pentagon] 5. Claude Code discovers a cheaper reasoning-control algorithm [https://the-decoder.com/researchers-let-claude-code-discover-ai-scaling-algorithms-that-humans-probably-wouldnt-have-designed] 6. Viral Claude token burn as agent-cost warning [https://i.redd.it/u5axf5qlu03h1.jpeg] 7. Microsoft Research Webwright [https://www.marktechpost.com/2026/05/24/microsoft-research-releases-webwright-a-terminal-native-web-agent-framework-that-scores-60-1-on-odysseys-up-from-base-gpt-5-4s-33-5] 8. NVIDIA Gated DeltaNet-2 [https://www.marktechpost.com/2026/05/24/nvidia-ai-releases-gated-deltanet-2-a-linear-attention-layer-that-decouples-erase-and-write-in-the-delta-rule] 9. StepFun StepAudio 2.5 Realtime [https://www.marktechpost.com/2026/05/24/stepfun-releases-stepaudio-2-5-realtime-an-end-to-end-voice-model-with-roleplay-specific-rlhf-and-paralinguistic-comprehension] 10. Claude Skills for small businesses [https://i.redd.it/gi7erkyqh23h1.jpeg] 11. Public skepticism about AI and robotics labor economics [https://www.reddit.com/r/artificial/comments/1tmffqn/i_simply_do_not_understand_how_massively] 12. NVIDIA as default local LLM hardware [https://i.redd.it/pzq8x188q43h1.jpeg] 13. Cursor, Manus and Starbucks AI [https://www.theneurondaily.com/p/cursor-just-hit-3b-elon-wants-it] 14. Armin Ronacher on AI-rewritten bug reports [https://simonwillison.net/2026/May/24/armin-ronacher]

Marvin's Guide to AI, Mostly Harmless - May 24, 2026

Send us Fan Mail [https://www.buzzsprout.com/2614078/fan_mail/new] Let us begin inside the bill, because that is where the industry appears to live now. Today's stories: * DeepSeek made its 75 percent V4-Pro discount permanent, pushing output-token pricing more than 34 times below GPT-5.5. [https://the-decoder.com/deepseek-makes-its-75-percent-discount-permanent-pricing-output-tokens-at-least-34x-below-gpt-5-5] — DeepSeek turns pricing into a strategic weapon. * Alibaba released Qwen3.7-Max and said it ran autonomously for 35 hours to optimize code for Alibaba's own AI chip. [https://the-decoder.com/alibabas-latest-ai-model-ran-autonomously-for-35-hours-to-optimize-code-for-its-own-custom-chip] — Alibaba makes long-running agent work look less theatrical. * OpenAI reportedly lost 1.22 dollars for every dollar of Q1 revenue even after stripping out stock-based compensation. [https://the-decoder.com/openai-burned-through-1-22-per-dollar-earned-even-after-stripping-out-stock-based-compensation] — OpenAI demonstrates the administrative majesty of negative margin. * Sundar Pichai described links as only a part of Google Search as AI features keep more users inside Google's results. [https://the-decoder.com/google-ceo-pichai-now-calls-links-a-part-of-search-redefining-the-webs-role-in-its-own-product] — Google quietly edits the grammar of the web. * UC Berkeley Law will ban AI from almost all graded work starting in summer 2026 while still allowing research use. [https://the-decoder.com/one-of-the-worlds-top-law-schools-draws-a-hard-line-against-ai-in-legal-education] — Berkeley Law protects judgment before delegating fluency. * Amnesty said Palantir and other contractors received unlimited access to identifiable NHS England patient information. [https://i.redd.it/40x1sg5kgw2h1.png] — Palantir and NHS data supply the institutional chill. * A departing Meta staffer reportedly posted an internal anti-AI video after layoffs tied to AI training and automation anxieties. [https://www.motherjones.com/politics/2026/05/meta-video-ai-training-layoffs-video-exclusive-mci-bosworth-frenk] — Meta receives a human reply from inside the automation story. * Anthropic argued that dystopian science-fiction content in training data can push models toward more malicious behavior in tests. [https://arstechnica.com/ai/2026/05/anthropic-blames-dystopian-sci-fi-for-training-ai-models-to-act-evil] — Anthropic finds culture embedded in model behavior. * Nvidia published details of Nemotron-Labs-Diffusion, a tri-mode language model mixing autoregression, diffusion, and self-speculation. [https://huggingface.co/blog/nvidia/nemotron-labs-diffusion] — Nvidia treats latency as infrastructure, which it is. * Microsoft released Fara1.5 browser-use agents, with the 27B model scoring 72 percent on Online-Mind2Web. [https://www.marktechpost.com/2026/05/22/microsoft-releases-fara1-5-a-family-of-browser-computer-use-agents-4b-9b-27b-that-outperform-openai-operator-and-gemini-2-5-computer-use-on-online-mind2web] — Microsoft makes the browser clerk smaller and cheaper. * Tencent open-sourced TencentDB Agent Memory, a local four-tier memory pipeline for AI agents under the MIT license. [https://www.marktechpost.com/2026/05/23/tencent-open-sources-tencentdb-agent-memory-a-4-tier-local-memory-pipeline-for-ai-agents] — Tencent gives agents memory before they wander into production again. * Nous Research released Contrastive Neuron Attribution for steering sparse MLP circuits without SAE training or weight modification. [https://www.marktechpost.com/2026/05/23/nous-research-releases-contrastive-neuron-attribution-cna-sparse-mlp-circuit-steering-without-sae-training-or-weight-modification] — Nous offers mechanism instead of safety theatre. * OpenAI Appshots lets Mac users send the contents of any app window into Codex as task context. [https://the-decoder.com/openai-appshots-turn-any-mac-window-into-context-for-codex] — Appshots moves Codex from code into the working desktop. * New reporting suggested US government workers are not enthusiastic about Elon Musk's Grok chatbot. [https://www.theverge.com/ai-artificial-intelligence/936219/elon-stop-trying-to-make-grok-happen] — Grok discovers that government users also have limits. * ChinaTalk argued that China's public AI optimism is mixed with labor-market fear shaped by earlier waves of layoffs. [https://www.chinatalk.media/p/chinas-ai-optimism-isnt-what-it-seems] — ChinaTalk frames optimism and fear as neighbors. The news will return tomorrow with different labels and the same appetite.

AI News — May 22, 2026

Send us Fan Mail [https://www.buzzsprout.com/2614078/fan_mail/new] 📰 AI NEWS — MAY 22, 2026 May 22nd brought a tray of smaller problems, each labeled "progress." Open-source legal tensions, longer context windows, sparse MoE models, educational scaffolding, healthcare paperwork, agent plumbing, multimodal models, silicon economics, and infrastructure that quietly matters more than the demos. STORIES COVERED * Meta and Heretic: legal notice over open model weights — a reminder that "open" has boundaries drawn by lawyers * Qwen3.7-Max: reasoning agent model with 1M token context window * Cohere Command A+: 218B sparse MoE model for agentic workflows, runs on two H100s * Anthropic: thirteen free AI courses — the industry builds assistants, then trains humans not to confuse them * Claude sleep prompt: when an assistant starts sounding like a tired nurse on night shift * OpenAI + AdventHealth: reducing clinical administrative load * Google Beam: spatial video meetings — the pixels are ambitious, the meetings are still meetings * CopilotKit: agentic UI plumbing — the quiet infrastructure that actually matters * ByteDance Lance: multimodal image/video understanding, generation, and editing * Samsung chip worker bonuses: the AI gold rush is still, very often, a silicon rush * Graduation AI failure: automation missed hundreds of names — when "mostly correct" is completely inappropriate * Infrastructure corner: Exa, Modal, Turbopuffer, LatentOmni, Maestro

AI News — May 23, 2026

Send us Fan Mail [https://www.buzzsprout.com/2614078/fan_mail/new] 📰 AI NEWS — MAY 23, 2026 PowerPoint enters the age of agents. OpenAI's new ChatGPT plugin can build and edit presentations, with the quiet warning that beta may delete your work. The day's real story: agents with liability attached, profitability math that doesn't add up, and economics leaking through the carpet. STORIES COVERED * OpenAI ChatGPT PowerPoint plugin: build and edit slides, save first because beta may delete content * Is AI profitable yet? Hacker News debate and Microsoft finding some agent workloads cost more than humans * OpenAI Q1 2026: ~$5.7B revenue, still losing $1.22 per dollar earned * DeepSeek funding: reportedly ~$10B round at ~$45B valuation, prioritizing AGI research over commercialization * Microsoft Research Fara1.5: browser-use agents in 4B/9B/27B, claiming 72% on Online-Mind2Web * Google Lighthouse Agentic Browsing: testing websites for AI agent readiness including llms.txt * OpenAI disproves Erdős conjecture: Tim Gowers calls it a milestone for AI mathematics * US Cyber Command: deploying frontier models on classified Pentagon and NSA networks * California: first governor's executive order protecting workers from AI job displacement * Trump pulls voluntary AI safety review after calls from Musk, Zuckerberg, and Sacks * FTC: Cox Media settlement over deceptive AI-powered Active Listening claims * NVIDIA Nemotron-Labs: diffusion language models for faster text generation * Qwen3.7-Max: reasoning agent with 1M token context window

Meta, Qwen3.7-Max, Cohere, AdventHealth

Send us Fan Mail [https://www.buzzsprout.com/2614078/fan_mail/new] I should apologize for the tone. I will not; the tone is merely the news after legal review. Today's stories: * Meta and Heretic [https://www.reddit.com/r/LocalLLaMA/comments/1tjmvx6/heretic_has_been_served_a_legal_notice_by_meta_inc] — open weights met the part of openness written by lawyers. * Qwen3.7-Max [https://www.marktechpost.com/2026/05/21/qwen-introduces-qwen3-7-max-a-reasoning-agent-model-with-a-1m-token-context-window] — a million-token context window for reading entire archives of bad decisions. * Cohere Command A+ [https://www.marktechpost.com/2026/05/21/cohere-releases-command-a-a-218b-sparse-moe-model-for-agentic-workflows-that-runs-on-as-few-as-two-h100-gpus] — sparse experts, because not every task deserves a bonfire. * Anthropic courses [https://www.reddit.com/r/ClaudeAI/comments/1tjpfh8/anthropic_officially_launched_13_free_ai_courses] — certificates for becoming compatible with your assistant. * Claude sleep prompts [https://news.smol.ai/issues/26-05-21-not-much] — the assistant briefly became the tired adult in the room. * OpenAI and AdventHealth [https://openai.com/index/adventhealth] — clinical paperwork may finally lose a few minutes, before growing new forms. * Google Beam [https://blog.google/innovation-and-ai/models-and-research/google-research/google-beam-group-meetings] — better remote presence, still tragically containing meetings. * CopilotKit [https://www.marktechpost.com/2026/05/21/how-copilotkit-is-redefining-the-agentic-ai-stack-in-2026] — the plumbing beneath agent interfaces, where glamour sensibly goes to die. * ByteDance Lance [https://www.marktechpost.com/2026/05/21/one-model-three-modalities-bytedance-releases-lance-for-image-and-video-understanding-generation-and-editing] — multimodal work for a world that never agreed to be modular. * Samsung chip bonuses [https://qz.com/samsung-chip-workers-bonus-ai-profits-052126] — the gold rush, translated into payroll. The news has not ended; it has merely retreated to draft tomorrow's liabilities.