Today's News
Jarred Sumner via The Pragmatic Engineer
Jarred Sumner used Anthropic's Claude Fable 5 with 64 parallel agents to rewrite Bun's entire codebase, consuming 5.9B input tokens and 690M output tokens across structured phases: 3-hour prep work creating a porting guide, trial runs with adversarial review, parallel file rewrites over 2 days, 12 hours of compiler error fixes, and 3 days of test suite debugging. This demonstrates a concrete, repeatable blueprint for AI-accelerated large-scale code migrations that would have taken a single engineer 1–2 years. The migration was motivated by stability issues inherent to mixing garbage-collected JavaScript with manually-managed Zig memory, resulting in use-after-free crashes and memory leaks that Rust's type system prevents at compile time. Bun now gets 22 million monthly downloads and is embedded in tools like Claude Code and OpenCode.
Bloomberg via Techmeme
Taiwan Semiconductor Manufacturing Co. raised capital expenditure guidance by at least $4 billion and revenue growth projections from 30%-plus to slightly above 40%, reflecting confidence that torrid AI chip and data center demand will extend into 2027 and beyond. The company announced an additional $100B US investment. Now 77% of TSMC's wafer revenue comes from 7nm or smaller chips, cementing its role as the critical constraint in AI hardware scaling. The guidance hike signals that AI demand is real and durable enough to justify massive capacity expansion, though TSMC management has emphasized disciplined, strategic pricing and close coordination with customers rather than opportunistic supply flooding.
Artificial Analysis via AINews
Moonshot AI launched Kimi K3, a 2.8-trillion-parameter model with native vision, 1M-token context window, and Delta Attention architecture enabling 6.3x faster decoding in long contexts. K3 achieves Opus 4.8-class performance (57 on Artificial Analysis Intelligence Index) while priced at $5.40/1M blended tokens versus $9 for Opus, with open weights promised by July 27, 2026. The model ranks #1 in Frontend Code Arena with 76% pairwise win rate, surpassing Claude Fable 5 (63%) and GPT-5.6 Sol (58%), and demonstrates strong long-horizon agentic capabilities: it autonomously optimized GPU kernels on H200s, built MiniTriton (a Triton-competitive compiler from scratch), designed a chip in 48 hours, and completed a two-week astrophysics research project in two hours.
K3 represents a strategic inflection: Chinese labs are closing the capability gap with top US closed models while compressing pricing and commoditizing frontier performance. However, the traditional cost arbitrage of open-weight models is eroding—K3 requires 64+ GPU supernodes to self-host efficiently, and its $5.40 API pricing converges toward frontier closed models rather than the $2 pricing of prior-gen GLM 5.2. This signals that at frontier scale, "open" no longer correlates with "cheap" in model economics.
Americans skeptical of AI's societal impact, complicating data center expansion strategy
Matthew Yglesias
Pew Research shows 40% of Americans predict AI will have negative societal impact versus only 16% positive—a fundamentally different baseline from housing opposition, where broad agreement exists on the principle of construction. This baseline skepticism means data center bans reflect genuine concern about AI progress itself, not just standard NIMBYism, creating a harder political problem for infrastructure buildout than typical zoning disputes.