Summary
The day split cleanly between a research-and-tooling morning and a cost-and-hardware evening, with the afternoon slot empty. The strongest cross-slot thread is the economics of agentic coding: the morning's coding-tools cluster (Kilo putting Nemotron 3 Ultra and Step 3.7 Flash in free, semantic indexing, Cursor's context-usage report) sets up the evening's seven-post Kilo cluster on billing, where GitHub Copilot's usage-based pricing went live and Uber reportedly burned its full 2026 AI coding budget by April. Underneath sits the day's biggest single story, Anthropic's "When AI builds itself" recursive-self-improvement report, which dominated the morning feed and drew both hype (@ns123abc) and a sharp methodological pushback from Hugging Face's Elie Bakouch on the confounded 8x-code plot. NVIDIA's Computex blitz ran in parallel all morning (Cosmos 3, Nemotron 3 Ultra, Vera Rubin AI-cloud, Parakeet ASR), and the evening's hardware-squeeze cluster from @ns123abc (TSMC High-NA EUV, a memory crisis forecast to 2031, NVIDIA + SK hynix HBM) names the supply-side counterpressure to all that token growth. The single most novel standout is the prototype AI worm that carries and runs its own LLM on machines it compromises, flagged via Schneier. Signal was strong in research and tooling; the rest, mostly @Scobleizer product hype and celebrity noise, is skippable.
Posts
- Anthropic's "When AI builds itself" recursive-self-improvement report (@AnthropicAI · wiki) [morning] (cluster of 3). Claims Claude writes over 90% of Anthropic's code, engineers ship 8x more per quarter, and a preview model beat human next-step decisions 64% of the time. @_sholtodouglas called himself "IO-bandwidth limited"; @ns123abc hyped it; @eliebakouch warned the trailing-average plot hides a model-version step effect.
- NVIDIA Computex blitz: Cosmos 3, Nemotron 3 Ultra, AI-cloud, Parakeet ASR (@nvidia · blog) [morning] (cluster of 4). Cosmos 3 is billed as the first open omnimodal world model for physical AI. Nemotron 3 Ultra (550B/55B-active, 1M context) targets long-running agents; a fine-tuned Parakeet ASR hit 97.7% Bahasa Indonesia accuracy.
- Coding-tools cluster: free open models, semantic indexing, Canvas, Claude Code vs Codex (@kilocode · @cursor_ai) [morning] (cluster of 4). Kilo added Nemotron 3 Ultra and StepFun Step 3.7 Flash for free and brought back semantic codebase indexing. Cursor shipped Canvas Design Mode and a context-usage report; Bakouch charted Claude Code vs Codex feature parity.
- OpenAI "Dreaming" background memory for ChatGPT (@eliebakouch relaying @MTSlive) [morning]. A memory system that synthesizes and updates user context in the background without explicit saves, rolling out to US Plus and Pro. Consumer mirror of the agent-memory research thread.
- Prime Intellect joins NVIDIA's Nemotron Coalition (@eliebakouch relaying @PrimeIntellect · blog) [morning]. Contributing 2,500+ open RL environments, the verifiers framework, Prime Sandbox, and NeMo Gym, arguing the hard part is now post-training, not pretraining. The open-RL substrate today's self-evolving-agents papers depend on.
- Frontier-lab biorisk letter to Congress (@logangraham) [morning]. Anthropic's Logan Graham tied a letter signed by Altman, Amodei, and Hassabis urging stronger biosecurity to his own LLM bioweapon red-teaming since 2022. Feeds the responsible-ai thread.
- Google passive heart-rate monitoring via smartphone camera (@GoogleResearch · blog) [morning]. A front-facing-camera system passively monitors heart rate during everyday phone use, claiming accuracy across all skin tones.
- Elon Musk pitches SpaceXAI satellites as hardware-agnostic orbital compute (@ns123abc) [morning]. Says the satellites will run "whatever GPU or TPU they want." Positioning only, but lands against SemiAnalysis's space-datacenter economics from the 06-04 digest.
- Agentic coding's billing reckoning (@kilocode · blog) [evening] (cluster of 7). Copilot's usage-based billing went live June 1; Uber reportedly exhausted its full 2026 AI coding budget by April on Claude Code and Cursor. Part vendor promo, but the cost math is real and reinforces the case for task-to-model routing.
- Devin Desktop and agent-fleet model lock-in (@kilocode) [evening]. Reacting to Windsurf's Devin Desktop for orchestrating local and cloud agent fleets. Kilo's "can you swap the model under each agent tomorrow" is the right fleet-routing question even as a sales line.
- Hardware supply squeeze (@ns123abc) [evening] (cluster of 4). TSMC confirmed it secured ASML High-NA EUV tools (R&D only); a separate post forecasts a memory crisis to 2031; plus an NVIDIA + SK hynix HBM tie-up. Unsourced, but the memory-scarcity thread matters for inference-capacity planning.
- AI worm with an onboard LLM (@Scobleizer · Schneier) [evening]. Researchers prototyped an internet worm that carries its own LLM and runs it on compromised machines. Schneier calls it the closest realization yet of the 1975 worm concept.
- Hitachi and Intel AI collaboration (@magicsilicon · press) [evening]. Strategic partnership on AI transformation across industries. Light on technical detail; an enterprise-adoption signal.
- Mira Murati on the 2023 OpenAI board crisis (@ns123abc) [evening]. Resurfaces Helen Toner's testimony and Murati's claim OpenAI "would have imploded" without her. Industry gossip, no new substance.
- "Fast is not smart" (@MillionInt) [evening]. Two short takes: speed gets mistaken for intelligence, and "verifiable tasks" often just means "easy tasks." A jab at how RLVR benchmarks pick low-hanging fruit.
- @Scobleizer product-hype stream (@Scobleizer) [evening]. Typeahead, a "tokenmaxxing" npm stunt, robot hands, Reve 2.0, TownAI, wrapped in VC-rejection commentary. Skip.
- @brivael, @mlevchin, @spencerpratt [morning + evening]. Political reposts, personal commentary, a Lord-of-the-Rings disco remix, and a Jimmy Kimmel celebrity item. No AI substance. Skip.