Anthropic quietly rewrites Claude Code on Bun's Rust runtime
π§ LAUNCH
Moonshot AI suspends new subscriptions due to Kimi K3 demand. The demand for Kimi K3 is so intense that Moonshot AI has had to halt new signups entirely β a rare capacity-driven freeze that validates K3's disruptive pricing and performance claims from earlier this week. When a model launch breaks your billing system, you've either massively underprovisioned or massively overdelivered. Watch for reopening announcements if you missed the window. (197 likes | 67 RTs) Read more β
π§ TOOL
Anthropic quietly rewrites Claude Code on Bun's Rust runtime.
Simon Willison's teardown reveals that Anthropic swapped Claude Code's underlying runtime to Bun β the Zig-and-Rust-based JavaScript runtime β a significant infrastructure move away from Node.js. The performance implications for CLI responsiveness are real: Bun's startup time and I/O throughput are measurably faster, and for a tool that developers invoke hundreds of times a day, those milliseconds compound. This signals where high-performance developer tooling is headed β native runtimes over interpreted ones. (370 likes | 508 RTs) Read more β
OpenAI cuts Codex context window by 100K tokens.
A quiet PR in the Codex repo drops the model's context from 372K to 272K tokens β a rare regression that reveals the cost-performance trade-offs happening behind the scenes at OpenAI. If you've built Codex-powered workflows that rely on stuffing large codebases into a single prompt, audit your pipelines now. The cut suggests that maintaining massive context windows at scale isn't economically sustainable, even for OpenAI. (292 likes | 143 RTs) Read more β
π¬ RESEARCH
Study: AI advice makes people 3x less accurate but 2x more confident.
A new study quantifies what many builders have suspected β AI assistance doesn't just fail to help in some cases, it actively suppresses critical thinking. Participants given AI advice were three times less accurate on tasks but reported twice the confidence in their answers. The mechanism is disturbingly simple: when an AI gives you a plausible-sounding answer, you stop thinking critically about the problem. Anyone shipping user-facing AI features should be redesigning their trust signals yesterday. (213 likes | 100 RTs) Read more β
π‘ INSIGHT
Simon Willison: AI mania is degrading institutional decision-making. Willison's latest essay argues the damage isn't just individual β organizations are making worse decisions at scale because AI hype short-circuits the slow, unglamorous work of actually understanding problems. Coming from one of the most technically credible voices in AI, this isn't anti-AI doomerism β it's a practitioner sounding an alarm about the gap between what AI can do and what institutions assume it can do. Read more β
Jensen Huang's Japan tour locks in NVIDIA's Asia infrastructure play. Huang's Tokyo visit yielded deals spanning data centers, automotive AI, and Japan's entire tech ecosystem β a coordinated blitz that makes NVIDIA's Asia strategy look less like business development and more like industrial policy. For anyone building on NVIDIA infrastructure, the geopolitical diversification of chip supply chains just got a concrete timeline. Read more β
ποΈ BUILD
Qwen3.6-27B Fable Fusion: A community-built 27B multimodal GGUF model fusing Qwen3.6 with multi-token prediction is trending on HuggingFace with 16.7K downloads. The open-model community keeps pushing creative architecture mashups that actually run on consumer hardware β grab it and try it locally with llama.cpp. (103 likes | 16.7K downloads) Read more β
π MODEL LITERACY
Context Window Economics: OpenAI cutting Codex from 372K to 272K tokens reveals an uncomfortable truth β context windows have real cost-per-token trade-offs, and bigger isn't always sustainable. Every token in a context window consumes GPU memory during inference, and costs scale superlinearly as windows grow (attention is quadratic, even with optimizations). When a provider shrinks a context window, they're telling you the economics didn't work at the old size. As a developer, understand that the advertised maximum is a ceiling, not a promise β and design your systems to work well within smaller windows rather than depending on the biggest number on the spec sheet.
β‘ QUICK LINKS
- Apple vs. OpenAI: Apple's lawsuit could derail OpenAI's hardware ambitions and IPO timeline. Link
- Dave Eggers at OpenAI: Told 200 staffers that ChatGPT is "silencing an entire generation" β and Sam Altman invited him to say it. Link
- Neil Rimer on AI wealth: Index Ventures co-founder says the historic AI wealth concentration will face redistribution pressure β voluntary or not. Link
π― PICK OF THE DAY
The confidence trap is real, and we're shipping it in every product. The study showing AI advice makes people 3x less accurate but 2x more confident isn't a UX footnote β it's an indictment of every product that ships AI suggestions without friction. Think about what this means at scale: millions of users making worse decisions while feeling better about them. The mechanism is human, not technical β we're wired to conserve cognitive effort, and a confident-sounding AI answer is the perfect excuse to stop thinking. This is the first rigorous evidence that "AI-assisted" can mean "AI-impaired" when the interface design assumes more information equals better decisions. If you're building AI features, the takeaway is clear: add friction before confidence. Force users to commit to an answer before showing the AI's suggestion. Show uncertainty ranges. Make disagreeing with the AI feel as easy as accepting it. The products that get this right will be the ones still trusted in five years. Read more β
Until next time βοΈ