A developer names the coworker habit of pasting Claude's answer instead of actually answering you.
What the article says
- The writer keeps getting Slack messages and code review comments that are just pasted AI output, unread and unedited by the person sending them.
- He calls this being a meat proxy, a human relaying a chatbot's words without adding anything of their own.
- AI text is extra work to read. It is wordy, sometimes confidently wrong, and packed with jargon he has to look up word by word.
- His fix is not to stop using AI, but to read what it says, check it, and rewrite the answer in your own words before sending it.
- He points to code review as the worst case, where nobody involved actually looked at the code being shipped.
What HN is saying
- Commenters overwhelmingly recognize the pattern and are worn down by it, especially when it comes from senior engineers who should know better.
- Several say the fix is social pressure. Calling it out publicly, or just asking someone straight up whether a message came from a bot, made it stop fast.
- The sharpest disagreement is whether this is new. Some say it is the same laziness people showed pasting Google or Stack Overflow links long before AI existed.
- One person argues context matters. A trusted colleague passing along a long AI writeup they actually vetted is genuinely useful, not lazy.
- A few worry this habit is quietly making people worse at thinking, echoing an older complaint about search engines and memory.
Alibaba just released its strongest coding model yet, and it's going open weight next week.
What the article says
- The article page itself came back empty, so this is inferred from the title and comments.
- Alibaba released Qwen3.8-Max, the most capable model yet in its Qwen family, aimed at coding and general work.
- It is the first Qwen-Max class model to go open weight, with the weights following next week.
- Pricing lands around two dollars in and six dollars out per million tokens, undercutting most rivals.
- A smaller 27 billion parameter version is also coming, built for people who want to run something similar at home.
What HN is saying
- Commenters are more excited about the smaller 27 billion parameter open weight version landing next week than about the flagship model itself.
- Several people say Qwen's local models already replaced their Claude subscriptions for daily coding.
- One tester had it convert a design into a working app. Claude finished cleanly in sixteen minutes while Qwen took two hours and kept timing out, though its visual sense looked decent.
- A running argument breaks out over whether frontier labs like OpenAI and Anthropic have any real moat left now that capable Chinese open models show up every few weeks.
- Simon Willison points out the confusing timeline. This release looks a lot like the preview model that came out two weeks earlier under a different name.
A GitHub account filed dozens of fake critical SQLite bugs and NVD rubber stamped them as real.
What the article says
- JFrog researchers dug into a batch of SQLite vulnerability reports that NVD and CISA had rated critical.
- The reports cited functions and line numbers that did not exist in the SQLite versions they targeted.
- None of the supposed exploits actually crashed the software when tested, and none appeared on SQLite's own advisory page.
- Out of fifty five reports from the same account, only one turned out to describe a real bug.
- The root cause traces back to funding cuts that gutted NIST's ability to manually verify submissions before they get published.
What HN is saying
- Commenters largely agreed this is a trust and verification problem, not just a technical one, since the vulnerability database now assumes good faith from submitters.
- Several pointed out that big projects like curl and the Linux kernel are becoming their own reporting authorities specifically to filter out this kind of AI generated noise.
- One sharp worry was that security teams under mandates to patch every listed vulnerability will burn real hours chasing fabricated ones.
- A few joked that even the article debunking the fake bugs looks AI written itself, which raises an uncomfortable question about who checks the checkers.
- Someone flagged that this could be weaponized deliberately, flooding scanners with junk to bury or distract from genuine exploits.
Terence Tao gets startlingly good math help from ChatGPT. The reason turns out to matter for the rest of us too.
What the article says
- The core claim is that expertise, not clever prompting tricks, is what actually gets you better results from an LLM.
- Terence Tao's chat with ChatGPT about a tricky math counterexample shows this. His messages are short, he pushes back gently instead of arguing, and he barely takes the model's suggestions, following his own instincts instead.
- The author says that trick only works because Tao actually understands the math. Copying his phrasing without the knowledge behind it won't get you the same result.
- The same pattern holds in coding. Knowing a codebase well lets you say something looks too complicated or ask whether it already does something, and that steering is what makes the output good.
- The upshot is that as models get stronger, the bottleneck stays human. The hard part is telling the model exactly what you want, and that takes real knowledge, not just skill at asking.
What HN is saying
- A widely shared story backs this up. A non technical friend tried to build a simple web app but had no vocabulary for what she wanted, and the conversation just spiraled into feature brainstorming instead of code.
- Several people say naming your expertise upfront changes the model's behavior immediately, making it terser and more technical rather than defaulting to a beginner explanation.
- The sharpest pushback is that this could be motivated reasoning. One commenter admits it would be personally convenient to believe expertise still matters, which is exactly why they're suspicious of the theory.
- Others complicate the picture with a car analogy. AI evens out results for simple, everyday tasks the way cars let anyone get from A to B, but at the very top level, small differences in skill still produce very different outcomes.
- A few worry less about who benefits now and more about the future. If people lean on LLMs from day one, will anyone actually build up the hands on domain knowledge the whole argument depends on.
AI agents just made it cheap to fork and personalize any tool you use, if the source is open.
What the article says
- A developer argues that coding agents have flipped the economics of open source. It used to take too much time to read and change someone else's code, but now an agent can do it in minutes.
- He describes setting up a nightly job where an agent pulls upstream changes, rebases his personal tweaks on top, checks everything still works, and swaps in the new version automatically.
- As a real example, he had an agent bolt a whole new feature into his own editor with a single prompt, something that would have been painful to build through a normal plugin system.
- His conclusion is that this kills the case for heavy configuration menus and plugin systems. If anyone can ask an agent to change the source directly, you barely need built in customization anymore.
- He notes the catch. Closed source tools, including Claude Code itself, cannot be personalized this way since there is no source for the agent to work with.
What HN is saying
- Simon Willison agrees the friction is gone. He now regularly has an agent clone a repo just to explain how it works, something he would never have bothered with before.
- The biggest pushback is that maintaining a personal fork is still real work. Commenters say merge conflicts are often about clashing design choices, not just code, and an agent cannot resolve those for you.
- Several people are unnerved by the nightly autoupdate idea specifically, imagining waking up to a broken tool because the agent quietly rewrote something overnight without anyone reviewing it.
- A former devtools CEO points out the business problem this creates. If customers can vibe code a good enough clone of your product from your own open source code, it gets even harder to charge for it.
- Some read the whole premise as overstated, noting that most engineers already leaned on config files and plugins just fine and never needed to touch the underlying source at all.