0:00 / 3:05aifeature Qwen’s new flagship model goes after long-running coding agents#
Alibaba’s Qwen team announced Qwen3.8-Max, a 2.4-trillion-parameter mixture-of-experts flagship with 95 billion active parameters, available now via QwenCloud API and slated for open-weight release next week. The post emphasizes long-horizon agentic coding: Qwen says the model ran a 16-day autonomous coding project, reproduced and improved on a research paper over about five days, and competed in a multimodal intent-recognition contest. The significance is less a single benchmark score than the claim that a Max-class Qwen model will be open-weight, increasing pressure on closed AI labs and giving developers another high-end option for coding-agent workflows.
Discussion: Mixed — HN was broadly excited about Qwen’s momentum and especially the promise of open weights, with many commenters comparing prior Qwen local models favorably to paid alternatives. But the thread also turned quickly to labor anxiety, doubts about AI-company moats and valuations, and some hands-on skepticism about whether Qwen’s tools match Claude-level reliability in practice. (Excitement about open-weight frontier-style models, Strong interest in local 27B-class Qwen models, Concern about programming contract work being displaced)
0:00 / 2:42aideep dive OpenAI says Astra found ten new math and theoretical CS results#
OpenAI says an internal version of Astra, its next major model, generated arguments for ten long-open problems across mathematics and theoretical computer science, including sphere packing, non-sofic groups, circuit complexity, quantum games, lattice cryptography, and Ramsey theory. The company says humans prepared the manuscripts with the same model, the model formalized each argument in Lean, and the total token cost to find the solutions would be roughly $2,000 at Sol API rates. The claims matter because, if independently validated by the relevant communities, they would be a notable step from AI-assisted problem solving toward AI-generated research results.
Discussion: Mixed — HN was impressed by the ambition and possible importance of the results, but the discussion was heavily tempered by demands for expert verification and concern that OpenAI’s framing is marketing-first. Commenters debated whether this reflects exponential AI progress, a brute-force search aided by formal verification, or something whose real significance will only be clear after mathematicians digest the proofs. (Excitement about AI contributing to open math problems, Skepticism about PR framing and need for independent verification, Debate over whether progress is exponential, sigmoid, or benchmark-driven)
0:00 / 1:12ai LLMs may amplify expertise, not replace it#
Sean Goedecke argues that the most important skill in using LLMs is not generic prompt craft, but expertise in the domain being prompted about. He uses Terence Tao’s ChatGPT discussion of a counterexample to the Jacobian Conjecture as an example: Tao’s value came from knowing which parts to question, where to steer, and what looked mathematically odd. The broader claim is that as models improve, human expertise still matters because the bottleneck is often knowing what to ask for and how to evaluate the answer.
Discussion: Mixed — HN broadly engaged with the thesis, with many agreeing that LLMs are strongest when guided by real domain knowledge, but the thread split over whether AI accelerates learning or lets people skip the learning that creates expertise. Several commenters emphasized verification, codebase familiarity, and the danger of subtle hallucinations; others argued that for routine tasks, getting useful output may matter more than becoming an expert. (Expertise as leverage rather than prompt tricks, Concern that delegation can erode the path to expertise, LLMs as accelerators for boring or unfamiliar tasks)
0:00 / 1:54ai Open-weights MiniMax H3 lands in ComfyUI for local video with sound#
MiniMax released H3, its third-generation video model, with open weights, and ComfyUI added native support on launch day. The model can generate up to 15-second videos at up to 2K from text, images, video, or audio, with stereo audio generated in the same pass rather than added afterward. Comfy says it cut the smallest variant’s memory footprint from 123.6 GB full precision to 42.5 GB using pruning, a lookup-table replacement for modulation weights, int8 convrot quantization, custom kernels, and dynamic VRAM offloading, making local runs possible on consumer GPUs such as an RTX 3060.
Discussion: Mixed — HN was broadly excited that an open-weights video model with native audio is usable in ComfyUI on consumer or prosumer hardware, but the thread quickly turned practical: speed, VRAM, visual artifacts, and whether the memory-saving trick is really lossless. Several commenters posted early local timings and reactions, with praise for impressive demos alongside reports of jank or poor results at higher resolutions. (enthusiasm for open weights and local inference, practical benchmarking on different GPUs, questions about the LUT-based memory reduction)
0:00 / 1:41science Dunning-Kruger gets a statistical reality check#↻ from 2020
This 2020 McGill explainer resurfaced on HN today, revisiting whether the famous Dunning-Kruger effect is really a cognitive bias or partly a statistical artefact. The article says the effect is often misused as “dumb people don’t know they’re dumb,” while Dunning himself framed it as a lesson in humility about our own weak areas. It highlights critiques from Ed Nuhfer and Patrick and Simone McKnight arguing that similar-looking graphs can emerge from random data and unreliable self-assessment measurements, so the classic pattern may not prove a special flaw in the brain.
Discussion: Mixed — HN was split between readers who felt the colloquial version of Dunning-Kruger is obviously recognizable in real life and readers who welcomed the article as another example of shaky popular psychology. Several commenters drilled into the statistics, with some accepting the random-data critique and others arguing that the simulation or interpretation did not actually disprove the original effect. A strong secondary theme was skepticism toward psychology research more broadly, sometimes escalating into broad dismissals of the field. (colloquial meaning versus scientific definition, random-data and measurement-error critique, replication crisis in psychology)
0:00 / 3:06softwaredeep dive AI coding agents revive the case for open-source devtools#
The essay argues that AI coding agents change the economics of customizing developer tools: instead of waiting for plugin hooks or configuration options, users can patch open-source tools directly and have agents help keep those forks rebased on upstream. The author illustrates this with Shelley, an open-source agent, and a personal diff-filtering tool called meat.dev, then contrasts that with closed-source tools like Claude Code where users are limited to whatever customization hooks the vendor exposes. The practical point is not just ideological open source, but whether devtools can be personalized deeply enough in an agent-driven workflow.
Discussion: Mixed — HN broadly agreed with the pro-open-source instinct, and several commenters said LLMs really do make it easier to inspect or modify codebases. But the thread pushed back hard on the essay’s more radical claim that agents can replace config files, plugin systems, and human-managed maintenance; many saw automated nightly rebases as fragile, costly, or unrealistic. (Open source as practical modifiability, not just ideology, LLMs lowering the cost of understanding and patching code, Skepticism about maintaining private forks and AI-generated changes)
0:00 / 1:41software One developer’s antidote to AI coding amnesia: type it yourself#
The author argues that using coding assistants to generate entire features can create “cognitive debt”: code lands in a project without the developer really understanding it. Their workaround is intentionally slow: ask the LLM for code in chat, then manually retype and adapt every edit, using the process to inspect APIs, catch hallucinations, refactor, and preserve a mental map of the codebase. The piece frames this as a tradeoff—maybe 2x faster instead of 10x, but with far more ownership and comprehension.
Discussion: Mixed — HN was highly engaged but skeptical. Many commenters agreed with the underlying fear of losing codebase understanding to LLMs, but a large share rejected manual retyping as inefficient or joyless; others described middle-ground workflows such as writing code first, using LLMs for review, or asking them to generate tests and explanations. (Concern about cognitive debt and skill atrophy from AI-generated code, Disagreement over whether retyping builds understanding or just rote memory, Preference for LLMs as reviewers, tutors, or pair programmers rather than autonomous coders)
JFrog researchers investigated a batch of newly published SQLite CVEs from a fresh GitHub advisory repo and concluded the reports were overwhelmingly fabricated or invalid. The alleged bugs cited non-existent functions, impossible line numbers, unrelated code, invalid SQL proofs of concept, and patches that did not exist; one CVE was initially scored Critical by Red Hat and later downgraded. The broader concern is that weak CVE intake and enrichment workflows can let plausible-looking, possibly AI-generated advisories propagate into NVD, GHSA, scanners, and enterprise ticket queues before anyone reproduces the bug.
Discussion: Negative — HN was alarmed and frustrated: commenters largely saw this as evidence that the CVE pipeline is too trusting and too easy to pollute, with LLMs amplifying an already noisy system. The discussion also split into a familiar debate over whether the root problem is AI itself, irresponsible users, or weak validation and incentives around vulnerability reporting. (LLM-generated security noise, CVE and CVSS trust problems, need for human verification and reproduction)