0:00 / 2:31aideep dive Terry Tao’s ChatGPT math session becomes an AI-collaboration case study#
A shared ChatGPT transcript labeled as Terrence Tao’s conversation walks through algebra around a claimed Jacobian Conjecture counterexample. In the transcript, ChatGPT reports checking explicit polynomial identities, deriving a map G from C^3 to C^3 with constant Jacobian determinant -1, and relating it to the displayed “original counterexample” through linear coordinate changes and output rescalings. The HN story is less a verified mathematical announcement than a window into how a top expert might use an LLM as an interactive symbolic assistant.
Discussion: Mixed — HN was strongly fascinated by the transcript as a glimpse of expert-level AI use, especially Tao’s short, precise prompts and the way the model was steered through dense algebra. The enthusiasm was tempered by repeated cautions that LLMs can make subtle mistakes and that this kind of workflow only works when the human has enough domain expertise to verify the output. (expert prompting beats one-shot prompting, AI as symbolic-algebra or research assistant, need for independent verification)
0:00 / 2:42aideep dive No smoking gun for AI labs gaming the pelican-on-a-bike test#
Dylan Castillo tested whether AI labs are secretly optimizing for Simon Willison’s famous prompt, “Generate an SVG of a pelican riding a bicycle.” He generated 1,008 SVGs across seven frontier models using an 8-animal by 6-vehicle grid, scored them with an LLM judge, and found little evidence that pelicans, bicycles, or the exact pelican-on-bicycle combination were unusually strong. The more plausible takeaway is that some labs may be improving SVG generation generally, but this experiment cannot prove who is doing that or whether it is benchmark gaming versus useful capability work.
Discussion: Mixed — The thread is mostly amused and appreciative of the experiment, especially because Simon Willison himself endorsed the methodology. Commenters broadly accept that narrow “pelicanmaxxing” was not shown, but many pivot to the bigger question of whether labs are improving or over-optimizing SVG generation in general, and a few argue the right-facing result or small sample still leaves room for suspicion. (approval of the quantitative methodology, humor around silly AI benchmarks, debate over benchmark gaming versus useful capability gains)
0:00 / 1:23ai 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)
0:00 / 1:32biotech Creatine is proven for muscle, not clearly for brains#
Dynomight’s essay argues that creatine is a normal nutrient rather than a hormone or steroid-like drug, and says the evidence for strength and short-burst exercise benefits is unusually strong for a supplement. It also pushes back on common worries about testosterone and baldness, tracing them mainly to a suspect 2009 rugby-player study. On the headline question — whether creatine makes you smarter — the article is far more cautious: the cognitive evidence appears mixed and any effect may be small, so the honest answer is closer to “maybe a little” than a proven nootropic.
Discussion: Mixed — HN was broadly comfortable with creatine’s physical-performance reputation, but much more skeptical about claims that it makes people smarter. The thread split between calls for stricter evidence, personal anecdotes about sleep deprivation or energy, and side debates over hydration, supplement-industry trust, and biological plausibility. (skepticism toward nootropic and supplement marketing, anecdotal reports of improved focus under sleep deprivation, stronger confidence in muscle and sprint-performance benefits than cognitive benefits)
0:00 / 3:14softwarefeature 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 / 1:31security Fake recruiter hid malware in a take-home coding test#
A developer says a LinkedIn recruiter sent a take-home Python/FastAPI project that looked legitimate, but the hidden `.git/hooks/pre-commit` script silently fetched and ran OS-specific payloads from a raw IP address. The follow-on Linux payload downloaded more scripts, installed Node-related tooling, and ran an obfuscated parser; the package dependencies looked aimed at token or crypto-related theft rather than ordinary project setup. The author also notes similar campaigns using `.vscode` configuration to trigger commands when a folder is opened, and found the attackers had apparently copied an innocent public GitHub repo and added malicious hidden files.
Discussion: Mixed — HN liked the write-up and treated it as a useful real-world security teardown, but the mood was uneasy about fake recruiting campaigns becoming more common. Commenters focused on practical defenses, the danger of trusting coding assignments, LinkedIn verification gaps, VS Code workspace trust, and frustration that AI safety filters can block defensive malware analysis. (take-home assignments as an attack vector, Git hooks and editor workspace trust, fake recruiter verification and LinkedIn identity checks)
0:00 / 2:27generaldeep dive Tech columnist John C. Dvorak has died#
The No Agenda Show account announced that longtime tech journalist and podcast host John C. Dvorak has died, saying he was 74 and that more information would follow. Dvorak was remembered on HN as a defining columnist of the PC-magazine era and later as a podcast personality whose sharp, often contrarian style made him both influential and divisive.
Discussion: Mixed — The thread is overwhelmingly elegiac and nostalgic, with many readers recalling Dvorak’s PC Magazine columns, TechTV/TWiT appearances, Cranky Geeks, and No Agenda. The affection is tempered by acknowledgment that he was a curmudgeonly and often controversial figure, with several commenters noting disagreements, failed predictions, or later falling-outs while still crediting his influence. (Nostalgia for 1980s and 1990s PC media, Respect for Dvorak’s distinctive voice and bold takes, Memories of PC Magazine, TWiT, Cranky Geeks, and No Agenda)
0:00 / 2:02general Reddit puts Old Reddit behind a login, and HN sees a data moat#
Reddit has begun requiring users to log in to access old.reddit.com, saying Old Reddit’s logged-out experience is a major source of abusive scraping and automated traffic. The linked post argues that the “safety” explanation is really about making Reddit’s user-generated text harder to scrape, contrasting Old Reddit’s plain HTML with New Reddit’s JavaScript-heavy design. The stakes are bigger than one interface: Reddit remains a major searchable archive of human-written advice at the same time that human text has become valuable training material for LLMs.
Discussion: Negative — HN’s reaction is strongly hostile to Reddit’s change, with many commenters reading “safety” as a cover story for protecting data from scrapers and preserving licensing value for AI deals. A minority acknowledges that abusive scraping can create real operational load, but even those comments generally frame Reddit’s choices as self-inflicted or user-hostile. (Old Reddit login requirement frustrates read-only users, Suspicion that anti-scraping means protecting AI-training data value, Decline in Reddit discussion quality, bots, and deleted content)