0:00 / 5:18softwarefeature Bento packs a whole slide editor into one HTML file#
A Show HN post introduced Bento/Slides, an MIT-licensed presentation tool where the slide data, viewer, editor, assets, and runtime live inside a single HTML file. The author says it can be opened in a browser, edited, presented, printed, saved, shared by file transfer, and used offline after the initial load; live collaboration uses an encrypted blind relay, according to the post. It matters because it points to a growing pattern on HN: AI-assisted, self-contained web documents that behave more like portable software than cloud-hosted files.
Discussion: Positive — HN reaction was strongly enthusiastic, with many commenters praising the single-file, offline-first idea and comparing it to TiddlyWiki, Reveal.js, draw.io-style documents, and other self-contained web apps. The main reservations were practical: Firefox performance, live-collab behavior under HN load, future version upgrades, email clients blocking HTML attachments, and a concern that the homepage's 'nothing phones home' message may conflict with observed Cloudflare-related behavior. (single-file web apps feel newly practical with AI-assisted coding, offline-first and no-account tools are attractive for corporate slide workflows, collaboration and performance need more polish)
0:00 / 3:53aideep dive Terrence Tao puts ChatGPT through a high-end math workout#
The link is a shared ChatGPT conversation in which Terrence Tao asks the model to analyze an alternate construction related to a claimed Jacobian Conjecture counterexample. The model walks through explicit polynomial maps, says it verifies a constant nonzero Jacobian, and relates the constructed map to a displayed original formula via linear coordinate changes. The important angle is not independent validation of a mathematical breakthrough, but a public glimpse of an elite mathematician using an LLM as a fast symbolic explainer and algebraic scratchpad.
Discussion: Mixed — HN was fascinated and mostly impressed by the transcript, especially by how much a world-class expert could extract from an LLM with terse, domain-heavy prompting. The caution was that most readers could not verify the math, and several commenters emphasized that this kind of use depends on already having enough expertise to catch errors. (AI as an expert collaborator rather than a one-shot oracle, Domain expertise as the key to useful prompting, Advanced mathematics as unusually dense and inaccessible)
0:00 / 3:54securitydeep dive Passkeys hit a usability backlash#
A widely discussed tweet from Nikita Bier argues that passkeys were designed by security engineers who failed to account for how ordinary consumers think about logging in. The HN thread turned that into a broader debate over whether passkeys reduce phishing and password pain, or simply add a vague new layer whose storage, syncing, sharing, and recovery rules are hard to understand. The practical stakes are high because passkeys are increasingly appearing on mainstream accounts, including shopping, health, and workplace services.
Discussion: Mixed — The discussion leans skeptical, with many HN commenters saying passkeys are confusing across devices, browsers, password managers, and shared accounts. A smaller but vocal group says passkeys work well when credentials are synced through ecosystems like Apple or password managers such as 1Password and Bitwarden. The strongest shared frustration is not with public-key authentication itself, but with unclear UX, recovery flows, and competing vendors trying to own the credential store. (cross-device confusion, account recovery anxiety, password manager and platform lock-in)
0:00 / 1:19ai No, AI labs probably aren’t gaming the pelican test#
Dylan Castillo tested whether AI labs have been optimizing models for Simon Willison’s famous prompt, “Generate an SVG of a pelican riding a bicycle.” He generated 1,008 SVGs across seven frontier models, varied animals and vehicles, scored the outputs with an LLM judge, and found little evidence that pelicans, bicycles, or the exact pelican-bicycle combination perform unusually well. The more plausible possibility, which the experiment cannot prove or disprove, is broader “SVGmaxxing”: labs getting better at SVG generation in general rather than targeting this one meme benchmark.
Discussion: Mixed — HN mostly enjoyed the experiment and the absurdity of taking the pelican-on-a-bicycle prompt seriously, with several commenters praising the methodology. The main debate was whether labs are not pelicanmaxxing specifically but are instead improving or over-optimizing SVG generation more broadly, which some see as a useful capability and others see as another benchmark-gaming risk. (Appreciation for a quantitative test of a running HN joke, Skepticism that the experiment can rule out broader SVG optimization, Debate over whether SVG generation is a meaningful AI capability)
0:00 / 1:59ai When Claude builds it, did you make it?#
Beej Jorgensen Hall’s essay argues that AI-generated code, art, and prose can feel less fulfilling than work made by hand, even when the user initiated and guided the result. He illustrates the point by presenting AI-generated artifacts as if they were his own, then revealing the prompts and contrasting them with a small flash-card app he wrote manually. The piece matters because it frames the AI coding debate less as productivity versus nostalgia, and more as a question of authorship, agency, and the emotional value of craft.
Discussion: Mixed — HN largely engaged with the essay in good faith, but split sharply on whether LLM-assisted work counts as “making.” Many commenters related to the loss of joy and craft; others argued that direction, taste, and intent are enough to take pride in the finished product. (loss of craft and enjoyment, LLMs as tools versus delegated labor, making versus producing or directing)