Hacker NewsFriday, October 9

Many commenters shared their own no-backend tools that store content in the URL, including slide stacks, big-text pages, parameter pages, zstd-compressed file sharing, recipe viewers and word searches. Others noted that data URLs and the older bigassmessage.com do similar things.

Some called "The URL is the app" meaningless LLM-speak and pointed to AI-style phrasing on the site. Others defended the author, citing non-native writers and the lack of evidence. The author said the 80-character title limit forced brevity.

A deaf user plans to use it for communicating, and a panel chair praised the countdown timers. Others discussed cheap digital signage via browser kiosks. Suggestions included the Wake Lock API, downloads for printing and PWA support (since added), plus a Firefox text-fitting bug fix.

Users compare Whistle with Parakeet, Whisper Large and Qwen ASR. Some find Parakeet fast but less accurate than Whisper, especially with accents. Many prefer small, embeddable models and clean the output with an LLM. Others report hallucinations like "Thank you." during silence.

One user replaced Amazon's Echo Show software with Whistle constrained to command templates, raising accuracy to near-Qwen levels. Others recall Sphinx4/JSGF and Windows grammar-based recognition. Benchmarks showed custom Whistle beating Vosk, Kaldi and Sphinx.

Commenters note binary size isn't the main problem; impaired speech, accents, filler words and proper names are. Suggestions include dictation-focused models, Gemini Transcribe and Handy. Streaming output is seen as essential, with Nemotron Streaming praised.

Critics call it an affectation of technical austerity: rich colors, Unicode dots and images instead of real ASCII constraints. Others cite proper ASCII pieces like julia-set, accept the loose label for discoverability, and link classic examples to explain purist views.

Many praise the beauty and evident fun, especially split-flap and nature scenes, imagining digital wall art. Users recall BBS animations, share related projects like a C port and play.ertdfgcvb.xyz, and ask how scenes are made: luminance-to-character mapping and color cycling.

A commenter shares Flow, an ASCII animation for Claude Code activity, recalling a wish for per-server art to aid SSH orientation. Another proposes flame decay, random scenes/palettes per session and crabigator integration; the author welcomes PRs. Others share MOTD art.

Many criticized the proofs as noisy and nearly unreadable without AI help. Some cited the Four Color Theorem as precedent for accepting non-human-scale proofs. Others suspect OpenAI could improve clarity cheaply, and some ex-mathematicians said certain papers read fine.

Users asked whether anyone has actually checked the proofs. An arXiv paper argues the Lean formalization of the Navier-Stokes result doesn't match the natural-language proof. Others worry about Lean exploits, and many results lack any Lean proof at all.

A rough estimate of ~$810 per solved open problem (3 hours per attempt at a ~5% success rate) sparked debate. Replies noted the unreleased model is likely pricier, failed attempts add cost, and verification takes human time. Still, it looks ~10,000x cheaper than paying mathematicians, worrying those concerned about funding.

Users lament a ~$1T company with little public disclosure. Explanations include the 2012 JOBS Act relaxing the 500-shareholder rule, post-Enron/SOX costs, and abundant private capital. Some say IPOs now dump on retail after growth; others argue staying private keeps losses away from the public.

Debate over whether OpenAI misled or investors and media wrongly "grossed up" figures to match Anthropic's method, which includes cloud partner sales. Some call it securities-fraud-adjacent or blame vague ARR metrics; others say the FT lacked due diligence and the story is overblown.

Many see secrecy, constant fundraising and about $2.5B monthly revenue as signs of weakness, comparing Altman to SBF and predicting a bubble pop. Others note the strong growth and the difference between a bet and fraud. Ed Zitron's accounting criticisms are seen as vindicated, though his track record is disputed.

Some argue frontier labs' subsidized plans make open models moot. Others report spending only a few dollars daily on DeepSeek via cache-heavy usage or cheap subs like OpenCode Go. OpenRouter routing, cache hits, ZDR and privacy shape real costs.

Huge VRAM needs and soaring memory prices make local hosting hard. Some blame DRAM price fixing; others cite fab lead times. Many note n-gram tables can sit in host RAM, and projects like DwarfStar lower the bar to consumer hardware.

Some say open models trail Opus/Fable by 6–12 months. Many users find Flash good enough for most coding, often using it as implementer or orchestrator alongside stronger planners. Others note it is chatty, weak at explaining, and worse at complex design.