Summary
The evening slot's strongest signal is a cluster of seven posts from @kilocode on the economics of agentic coding: GitHub Copilot's usage-based billing went live June 1, Uber reportedly burned its entire 2026 AI coding budget by April (on Claude Code and Cursor, not Copilot), and the pitch is model-portability as the durable hedge. This is promotional but it names a real shift, and it lands on the same theme as today's hardware cluster from @ns123abc, which carries TSMC buying ASML High-NA EUV tools, a memory crisis forecast running to 2031, and an NVIDIA + SK hynix HBM tie-up. Both clusters point at the same squeeze: compute and memory are getting scarcer and more expensive at exactly the moment agentic workloads are exploding token consumption. The standout single item is Schneier flagging a prototype AI worm that carries its own LLM and runs it on machines it breaks into, the closest thing yet to a 1975-style self-propagating worm. The rest is mostly @Scobleizer's product-hype and VC-rejection stream plus one off-topic celebrity post, all noise.
Posts
- Agentic coding's billing reckoning (cluster of 7) (@kilocode · blog). GitHub Copilot's usage-based billing went live June 1; Uber allegedly exhausted its full 2026 AI coding budget by April on Claude Code and Cursor. Kilo's argument: agentic workflows burn tokens far faster than flat per-seat budgets assumed, and the fix is not letting one vendor own your meter. Pitch is for 500+ model access with bring-your-own-keys, so this is part vendor promo, but the underlying cost math is real and reinforces the case for task-to-model routing.
- Devin Desktop and agent-fleet model lock-in (@kilocode). Reacting to Windsurf shipping Devin Desktop to orchestrate fleets of local and cloud agents from one surface. Kilo's framing, "can you swap the model under each agent tomorrow without rewriting the harness," is the right question for fleet-scale routing even if it doubles as a sales line.
- Hardware supply squeeze (cluster of 4) (@ns123abc). TSMC chairman C.C. Wei confirmed the company has secured ASML High-NA EUV tools (R&D only for now); a separate post forecasts a memory crisis lasting until 2031; and an NVIDIA + SK hynix HBM partnership ("GPU King and HBM God locked-in"). Unsourced tweets, but the memory-scarcity thread matters for anyone planning inference capacity.
- AI worm with an onboard LLM (@Scobleizer · Schneier). 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 original 1975 worm concept. Worth a look for the agent-security angle.
- Hitachi and Intel AI collaboration (@magicsilicon · press). Strategic partnership to push AI transformation across industries. Light on technical detail; an enterprise-adoption signal more than a chip story.
- Mira Murati on the 2023 OpenAI board crisis (@ns123abc). Resurfaces Helen Toner's testimony and Murati's claim that OpenAI "would have imploded" without her actions during Altman's brief ouster. Industry gossip, no new substance.
- "Fast is not smart" (@MillionInt). Two short takes: speed gets mistaken for intelligence, and "verifiable tasks" often just means "easy tasks." A pointed jab at how RLVR benchmarks pick low-hanging fruit.
- Typeahead, tokenmaxxing, robot hands, Reve 2.0, and assorted hype (@Scobleizer). A stream of product plugs (Typeahead inline writing, a "tokenmaxxing" npm stunt, dexterous robot hands, Reve 2.0 4K image model, TownAI) wrapped in personal VC-rejection commentary. Skip.
- Jimmy Kimmel vs Spencer Pratt (@spencerpratt). Off-topic celebrity item, no AI content. Skip.