0:00 / 2:36aideep dive ByteDance’s Seedance 2.5 aims AI video at longer, more controlled scenes#
ByteDance’s Seed team has launched Seedance 2.5, a new video creation model that it says can generate 30-second audio-video clips in one pass, extend them across multiple rounds, and use up to 30 images, 10 video clips, and 10 audio clips as references. The release emphasizes longer-form storytelling, more stable editing with timestamp-level control, and use cases beyond media, including education, manufacturing simulations, synthetic robot-training video, and autonomous-driving scenarios. The company says Seedance 2.5 is rolling out on Jimeng AI and Doubao Pro, with API access planned through BytePlus ModelArk, while also acknowledging remaining issues around complex physical motion and multi-subject scene stability.
Discussion: Mixed — HN was impressed by the apparent quality jump, especially the longer clips and fewer obvious artifacts, but the discussion quickly split between creative excitement and unease about deepfakes, slop, cost, and whether these systems are useful beyond flashy demos. Several commenters compared Seedance with emerging open-weight or local models and questioned whether ByteDance’s emphasis on action-heavy video fits different creator markets. (AI video quality is improving fast, concerns about deepfakes and synthetic content saturation, cost and inference economics)
0:00 / 3:14aideep dive MIT Sloan: AI money advice is useful, but prompt quality matters#
MIT Sloan summarized a new paper testing financial advice from large language models by having 1,000 adults write prompts and then simulating what would happen if people followed the chatbot recommendations over a lifetime. The study found the advice often aligned with mainstream financial principles — saving during working years, investing in diversified stock funds, and reducing risk with age — but performed worse around job-loss shocks, active rebalancing, and nuanced personalization. More structured prompts improved results, while prompt differences and model behavior produced meaningful gaps by gender, financial literacy, and prior AI experience.
Discussion: Mixed — HN was cautiously interested but far from dazzled. Many commenters argued that basic personal-finance advice is simple enough for an LLM to reproduce and may beat poor or conflicted human advice, while others emphasized that real financial planning involves trade-offs, taxes, behavior, privacy, and market assumptions that chatbots can mishandle. (Basic advice is mostly boring and standardized: save, avoid debt, buy diversified low-cost funds, Skepticism that LLMs can handle nuanced trade-offs, shocks, taxes, and long time horizons, Concern that financial outcomes depend on prompt quality and user literacy)
0:00 / 3:13softwaredeep dive Diátaxis puts technical docs into four boxes#
Diátaxis is a framework for technical documentation that divides user needs into four forms: tutorials, how-to guides, technical reference, and explanation. The site argues that organizing docs around those needs helps teams decide what to write, how to write it, and where it belongs. The HN revival matters because documentation is now being reshaped both by human information architecture work and by teams asking AI tools to produce usable first drafts.
Discussion: Positive — HN was broadly favorable, with several commenters saying Diátaxis helped them restructure real documentation or gave teams a shared vocabulary for writing. The main pushback was that the framework can be treated too rigidly, that its own site feels verbose to some readers, and that reorganizations can add extra clicks or fail to solve doc freshness. (Strong praise from people who have used it on real docs projects, Interest in using Diátaxis as a prompt or structure for LLM-generated first drafts, Cautions not to treat the four categories as gospel)
0:00 / 1:44security Lean fixes a kernel bug that let a bogus proof prove false#
Lean fixed kernel soundness bug #14576 after an AI-assisted, sorry-free “disproof” of the Collatz conjecture turned out to exploit a flaw in how the kernel handled nested inductive types. The issue could make Lean accept a proof of False, but the post says it was reachable only via metaprogramming or direct kernel input, not through the normal frontend, and it was fixed about an hour after a reduced report was filed. The post also explains why an older version of the independent nanoda checker missed the example, and lists follow-up hardening work and additional kernel fixes.
Discussion: Mixed — HN was concerned but not panicked: commenters largely treated this as a serious, expected class of implementation bug in proof checkers, not a collapse of formal verification. The discussion focused on conditional trust, independent checkers, whether AI will find more soundness bugs, and comparisons with simpler systems like Metamath. (formal verification is powerful but not absolute, independent kernels and keeping them updated, AI as both exploit finder and source of suspicious proofs)
0:00 / 1:43hardware Wafer says AMD’s MI355X wins on Kimi K3 inference cost, not raw speed#
Wafer published a benchmark claiming AMD’s 8-GPU MI355X node runs the 2.8T-parameter Kimi K3 model with better performance per dollar than Nvidia’s B300, even though B300 is faster in raw throughput. In Wafer’s test, MI355X reached 952 aggregate tokens per second and 118 single-stream decode tokens per second, versus B300’s 1,568 aggregate and 172 single-stream, but the claimed GPU-hour pricing makes MI355X come out ahead on tokens per dollar. The post also says Wafer fixed ROCm-side issues in speculative decoding and prefill, including a missing top-k renormalization path and an attention-head shape mismatch handled by padding 12 heads to 16.
Discussion: Mixed — The thread was skeptical more than celebratory. Commenters questioned calling weight-released models “open source,” challenged Wafer’s pricing assumptions and benchmark fairness, and several criticized the article’s writing and presentation as sloppy; a smaller set engaged with the technical details around memory capacity, multi-node B200 penalties, and ROCm support. (open source versus open weights terminology, benchmark fairness and pricing methodology, AMD MI355X memory capacity as a practical advantage)
0:00 / 2:40generalfeature Google’s RSS retreat still stings#↻ from 2023
This 2023 Open RSS essay resurfaced on HN today, arguing that Google helped weaken RSS adoption by repeatedly adding RSS support to products and then removing or reducing it. The article points to Chrome’s vanished RSS UI, Google’s handling of FeedBurner, the 2013 shutdown of Google Reader, changes to Google Alerts and Google News, and uncertainty around later ChromeRSS experiments. The broader point is less that RSS is dead—it is still widely used, especially for podcasts and feeds—and more that big-platform product decisions can erode confidence in open web standards.
Discussion: Mixed — The HN mood is strongly pro-RSS and broadly critical of Google, with many commenters treating the shutdown of Google Reader as a symbolic break from the open web. But the thread also debates whether RSS was ever mass-market friendly, whether ads and social platforms made its decline inevitable, and whether users should simply rely on independent readers instead of Google products. (nostalgia for the early open web, anger over Google Reader and other discontinued Google products, RSS as an antidote to algorithmic feeds and walled gardens)
0:00 / 1:51general Kenji’s 30-minute pressure-cooker pho resurfaces on HN#
A 2015 Serious Eats recipe by Kenji López-Alt resurfaced on Hacker News: pho ga, Vietnamese chicken noodle soup, made in about 30 minutes with a pressure cooker. The technique chars onion and ginger, then pressure-cooks chicken drumsticks with cilantro, spices, fish sauce, sugar, and water for 20 minutes, aiming for tender chicken and a clear, aromatic broth without long simmering or constant skimming. The HN angle was less news than applied kitchen engineering: using pressure, temperature, and reduced agitation to compress a traditionally slow broth into a weeknight process.
Discussion: Mixed — The thread was mostly warm toward Kenji López-Alt and Serious Eats, with several commenters saying this is an old favorite they make regularly. The pushback centered on whether a recipe belongs on Hacker News, plus practical caveats about adapting older pressure-cooker recipes to Instant Pots and other electric models. (Admiration for Kenji’s food-science approach, Pressure cookers as a legitimate cooking shortcut, Noting the recipe appears to date from 2015 despite updated pages)
A 15-year-old posted a GitHub project for a cycloidal gearbox and a Python generator script meant to create the geometry for Fusion 360. The repo describes three iterations: a hand-cranked 1:9 test gearbox, a failed compact NEMA 17-footprint version limited by 3D-printing tolerances, and a larger working NEMA 17 version in PLA with a reported 1:9 ratio, 90 mm outer diameter, 1.3 N·m torque, and 66% efficiency. The project matters less as a product than as a strong maker-engineering demonstration: parametric design, iteration, measurement, and documenting failure modes.
Discussion: Positive — The discussion was overwhelmingly encouraging, with many commenters telling the builder to drop the “wannabe” label and keep going. Side threads debated whether “engineer” is a protected or informal title, whether minors should post their age online, and whether AI-assisted projects change the bar for young makers. (strong encouragement for a young builder, debate over the meaning and legal status of “engineer”, basic explanations of cycloidal gearboxes and backlash)