Hacker News
1Sharing AI progress in mathematics
- Links to OpenAI's "math" GitHub repository and its preprints folder, which appears to hold mathematical research papers or preprints.
Grief and meaning for human mathematicians
A researcher who spent 24 years on Barnette's Conjecture mourns its AI proof, noting he'll never meet its "creator." Others share demotivation, an ERC group's goals solved overnight, and debate whether AI erodes training of future mathematicians or augments them.Concentration of power in proprietary models
Many worry these results come from withheld internal models no one can reproduce, with unknown compute and setups. Some cite Tao's group objecting to strip-mining open problems; others argue open-weight models lag under a year and specialized science tools are always restricted.Significance of results and limits of LLMs
The Unique Games proof would settle many approximation hardness results; others note Barnette, Millennium-problem progress, and huge polynomial exponents. Debate whether LLMs truly generalize or exploit broad coverage plus Lean verification and compute, and whether to fear AI doom.
2Claude Haiku 5.5
- Claude Haiku 5.5 launches as the cheapest, fastest small Claude model, costing about 75% less than Haiku 4.5 and suited to high-volume tasks and subagent work. Sonnet 5.5 cache reads are halved, and Max and Team subscribers get new monthly API credits.
Pricing and the 100k token cutoff
Haiku 5.5 is 90% cheaper than 4.5 under 100k tokens but jumps 5x above it. Critics call the cutoff too low for agentic work and note tokenizer inflation; others say it suits subagents, classification and summaries. Comparisons to GPT-6 Luna dominate.Monthly API credits and Agent SDK fears
Max and Team subscribers get $100–$500 in monthly API credits, praised for shipping AI features and using other harnesses. Some suspected a move to bill the Agent SDK ("claude -p") via credits; Anthropic staff clarified subscriptions still cover it.Benchmarks and real-world performance
Pelican SVGs, Pac-Man, data analytics and image-to-HTML tests show big gains over Haiku 4.5 at low cost, roughly on par with or smarter than Luna. Some found it weak on complex UI, slower, or worse than 4.5 on tuned production prompts.
3Tell HN: GitHub refuses to remove cracked copies of my software after a month
- Photopea's developer says people use AI to strip ads from his web app's code and republish it on GitHub, hurting his reputation. GitHub rejected his DMCA takedown, saying it couldn't confirm a violation. He suspects no human reviewed it and asks whether to hire a lawyer.
Takedown notices and legal options
GitHub likely treated the report as a §1201 anti-circumvention claim. Users suggest a §512(c) infringement notice instead, though some argue deobfuscation qualifies under §1201. An IP lawyer advises using specialist firms or accepting losses, and notes that collecting damages abroad is hard.AI cloning and client-side vulnerability
LLMs can reimplement or decompile software, possibly legally via clean-room tactics. Some predict logic will move server-side. Others counter that client-side processing is what makes indie apps like Photopea affordable, and that anti-piracy friction barely slows determined crackers.Legitimacy of code copyright and adware
Some argue copyright on code, especially JS served openly, is absurd and that stripping ads is fair. Others defend copyright as the basis of the GPL and of protection for indie devs. Users debate ethical business models such as Patreon, subscriptions, or a cheap ad-free tier.
4Shipping JPEG XL in Chrome
- Chrome 155 adds JPEG XL image support, offering better compression, lossless options, and HDR. The decoder was rebuilt in Rust for memory safety without sacrificing speed. The move responds to strong developer demand, and Chrome encourages developers to adopt .jxl images and report bugs.
Why Google reversed course
Users debate why Chrome re-added JPEG XL after removing it. Cited reasons include the memory-safe Rust decoder jxl-rs (now faster than libjxl), PDF spec adoption of JXL, and other browsers' support. Some blame ego or internal politics, while others note Google helped create JXL.Value over JPEG and PNG
Skeptics prefer proven JPEG/PNG or say fast connections make savings moot. Others cite real gains: smaller files for metered or slow users, lossless JPEG transcoding, HDR, layers and extra channels. AVIF and WebP comparisons and slow server-side encoding also come up.Compatibility and fragmentation
Commenters worry about app support gaps, citing WebP and HEIC troubles in Telegram, Discord and school LMSs. Some lament the lack of system-wide codec plugins like Amiga DataTypes. Others note one extension can't signal lossy vs lossless and suggest conventions like .ll.jxl.
5Visa, Mastercard, major banks facing new litigation over 'anticompetitive' fees
- A San Diego pizzeria's proposed class action accuses Visa, Mastercard and five banks of conspiring to inflate credit card interchange fees via anticompetitive rules, seeking relief for merchants who accepted the cards since Jan. 25, 2019, a period not covered by a prior $5 billion settlement.
Surcharges and transparent pricing
Many want merchants free to add card fees or reject costly cards, so users of premium rewards cards stop being subsidized. Others want the listed price to be the price paid; NZ allows surcharges, while Australia recently banned them.Public payment networks vs. private middlemen
Some argue payments should be a public utility, citing PIX, UPI, Interac, SEPA/Wero and EU fee caps. Skeptics point to security and compliance costs, government competence, surveillance risks, and the real value of credit intermediation and chargebacks.ACH apps as a card alternative
Gas chains offer per-gallon discounts to pay via app-linked ACH, which costs merchants far less than cards. Commenters fear giving apps bank access, cite fraud at gas stations and aggressive collections, dislike app sprawl, and note rewards cards often beat the discount.
6GPT‑6 and Intelligent UI for everyone
- GPT-6 is coming to all ChatGPT users and introduces Intelligent UI, which answers with interactive visuals, charts and custom tools. It starts answering while still thinking, improves web search answers and safety, and rolls out to paid tiers today and to Free and Go tiers tomorrow.
Cluttered visual answers feel condescending
Many disliked GPT-6's image-heavy, whitespace-filled responses, citing a roast recipe whose photo didn't match the menu and omitted ingredients. Others defended visuals or "ELI5" simplicity, while some suspected the format exists to slip in ads.Confusing model lineup and capacity strain
Users questioned why Chat uses GPT-6 Sol instead of the more capable 6.1 Sol. Some speculated that 6.1 is a heavier model rushed out to counter Anthropic. Others reported capacity outages and frustration with the Work/Chat/Codex split.Disposable, on-demand generated UI
Some welcomed "paper plate" UIs and devices morphing to fit their needs, saying this could threaten Apple's app paradigm. Skeptics worried about latency, overzealous visuals when a quick answer suffices, and users getting lost in UIs that change every time.
7EmbeddingGemma 2: An open, lightweight multimodal embedding model
- EmbeddingGemma 2 is Google DeepMind's new open, on-device embedding model that unifies text, code, images, audio, and video in one space. Built on Gemma 4, it leads its size class, runs efficiently on phones for private offline search and RAG, and is available on Hugging Face and Kaggle.
Open weights and self-hosting value
Users praise the Apache 2.0 license, noting stored vectors make proprietary embedding models risky if deprecated. Some prefer hosted access with an open fallback; others argue embeddings are ideal to self-host cheaply on rented GPUs for batch workloads.Practical multimodal use cases
People discuss semantic image search, legal evidence search across depositions, photos and video, Edge Gallery image search, and a drop-in upgrade for a Magic site. Local laptop throughput numbers were shared, plus the "decision maker" trick of embedding answers.Benchmarks and comparisons
Text benchmarks match EmbeddingGemma v1, and early tests found it no better for text-only retrieval. Commenters ask about comparisons to SigLIP2 and Voyage, note it uses MRL rather than MatFormers, and suggest MuQ-MuLan remains better for music audio.
8Margaret Hamilton has died
- Margaret Hamilton, the computer scientist who led MIT's Apollo flight software team and helped establish software engineering as a discipline, died Sept. 30 at 90. Her priority-driven software saved the Apollo 11 landing; she later founded companies and won the Presidential Medal of Freedom.
Coining "software engineering"
Users note Hamilton coined the term to legitimize software as an engineering discipline. Early on it was mocked as a joke. Others credit her as a forebear of rigorous methodologies like PRIDE and share interviews and oral histories.Tributes and calls for a black bar
Many call her a true legend and urge HN to display a black bar. Commenters share personal encounters with her, the iconic MIT code-stack photo, her LEGO minifig, and praise for her low-profile, non-self-promoting career.Hacker culture and disputed credit
A TX-0 anecdote from Levy's "Hackers" sparks debate over whether the hackers' interference reflected toxic masculinity or mere irresponsibility. One user dismisses her contributions as overstated, drawing pushback over sexist "slept her way up" claims and missing sources.
9Decisions API is in public beta
- The Decisions API, in public beta with gpt-6-luna, evaluates text and images about 10x faster than the Responses API, returning probabilities, choices, or rubric scores to classify, route, and prioritize content. It charges only for input tokens and supports ZDR and HIPAA.
AI as a commodity market and product stickiness
OpenAI copying Jev is seen as evidence of a price race to the bottom. Some say switching models is trivial. Others cite enterprise compliance, existing contracts and integrations as lock-in. Many doubt high-margin investor pitches survive cheap open-weight competition.Early benchmarks show it trails Jev
One user's evals found gpt-6-luna slower, about 3x pricier and less accurate than Jev, with questionable tagging scores. Others call it a rushed flag-planting release that still wins enterprise customers via existing OpenAI relationships, compliance and cloud availability.Zero-shot classification and probability quirks
Users frame it as a zero-shot classifier with no training needed and praise its fast image input. Predicates are independent, so probabilities needn't sum to 1. Tests show choice order skews results, LLM-like biases persist, and the confidence field's purpose is unclear.
10A font recreated from photographs of classic Commodore 64 keycaps
- C64 Keyboard is a free CC0 font recreated from photos of a Central European Commodore 64's keycaps, featuring uppercase letters, numbers, symbols, accented characters, f1–f12 labels, and all 63 PETSCII graphic legends. It's an unofficial reconstruction, not endorsed by Commodore.
Related fonts and typography history
Users shared related type stories: the Gorton article, a font carved into a tree, and Microgramma/Eurostile used on VIC-20, PET and sci-fi props. Some guessed the original C64 font resembles Akzidenz-Grotesk Condensed; the C64C reportedly used ITC Avant Garde.Quality and AI-generation critique
A rude GitHub issue drew criticism, though some agreed with its points: a misproportioned S and M, wrong stems on 3 and 4, and misaligned punctuation like !#$% and a wonky @. Others called it AI slop.Retro aesthetic and practical use
Many praised the clean look and nostalgia, like "Press play on tape." 8BitDo's C64-style keyboard drew sound and quality reviews. Others wanted lowercase or extra glyphs for coding, or a ZX Spectrum version, and noted the original keycaps feel more modern than the C64C's.