500B tokens: AI agents decompile a shooter to 83% byte-exact
A 3-month run with up to 15 Claude Code and Codex agents rebuilt 99% of a game's functions — plus free local Mac dictation and OpenAI's ultrafast tier.
Copy markdown
99% of functions back, 83% byte-exact — on cheap models
A developer aimed fleets of up to 15 Claude Code and Codex agents — mostly Sonnet 5 and Luna — at a commercial FPS for roughly 3 months and an estimated 600–700B tokens, rebuilding 99% of its functions (83% byte-for-byte) into a build that reportedly runs flawlessly. The reusable trick: a hashed, machine-checkable byte-matching oracle with PASS/FAIL CI made correctness objective enough that cheaper models could do most of the work.
BetterWispr: free, local voice dictation for Mac
A Show HN launch puts Wispr Flow-style dictation on macOS for free and fully open source (Apache 2.0), transcribing on-device via Whisper Large v3 Turbo or Parakeet with no cloud fallback. Hold ⌥Space to dictate prompts, commits, and docs privately — and it can even hand text to Claude Code or Codex for cleanup.
OpenAI adds an 'ultrafast' tier to GPT-6.1 Sol
Set service_tier: "ultrafast" on the Responses API and GPT-6.1 Sol narrows the gap between output tokens — noticeably snappier streaming for voice interfaces, agent loops, and anything latency-sensitive. It's open to all API users (subject to rate limits), with global processing and US or EU data residency.