What commenters asked our summary
In our own words. Each line links to the comment it is based on.
- Did you train a probe on the hidden state to predict the probability of being wrong?From this comment
- Is this work similar to existing research on learning from internal representations?From this comment
- Has this been benchmarked on coding tasks with fallback to a larger local model?From this comment
- Could internal signals reveal developer judgments like code cleanliness or smoothness?From this comment
- Does the model's quality degrade in other ways after this training?From this comment
What they doubted or pushed back on our summary
- The claim of knowing when wrong is epistemically flawed; models can only know uncertainty, not falsehood.From this comment
- If the confidence score is accurate, handoff is unnecessary; just re-run until the score is high.From this comment
What the maker answered our summary
9 replies from the maker in the thread; the points below are the ones we could tie to a comment.
- The mechanistic study will be compiled into a shareable report once pending caveats are solved.From this comment
- The work is loosely similar to existing research on learning from internal representations.From this comment
- They are currently working on larger models and found the suggestion interesting.From this comment
- Main weights are frozen and only added weights are trained, so quality should not degrade.From this comment
What a builder can take from it: When describing model confidence, avoid overclaiming knowledge of wrongness, and expect scrutiny of whether your signal actually improves outcomes over simple retries. our reading
The title, taken apart our calculation
This post got 191 points; the median for all 1442 posts is 112. In Open source projects (328 posts) the median is 103 points and 38 comments.
When it was posted our calculation
Wednesday at 17:00 UTC. Of 3,019 Show HN posts submitted in that UTC hour over the window (2025-11-20 to 2026-11-20), 109 (3.6%) reached 50 points; on Wednesdays it was 2.6% of 8,118. The average over all hours is 2.9%. One post's time does not explain its result; the full table of hours is on the Show HN hub.
Check your title our calculation
Type the title you plan to post. It is compared with 1,442 Show HN posts that reached 50 or more points: its length, its shape, and how posts of that length and shape did. Nothing is sent anywhere.
Type a title above to see where it sits.
Draft my Show HN post
Paste your site address and get title options and a first comment for your own Show HN post, written from what your site says and checked against the numbers above. Nothing is posted for you.
Free to try and no account needed. FoxPlug is for people who ship several projects and keep stalling on telling anyone: it turns what you ship into posts you approve before anything goes out. See what each plan includes.
More Open source projects teardowns
Points, comments, titles and times are Hacker News's own numbers, read from the post itself, last updated Nov 2026. Summaries of what people said are ours, written in our own words, and each one links to the comment it came from. Percentages and medians marked "our calculation" are computed from the posts counted on the page: they describe what happened, not what will happen to your post. FoxPlug posts nothing for you. Plans and limits are on the pricing page. Questions: support.
After the launch post: a changelog your users read
A launch brings visitors once. FoxPlug turns what you ship afterwards into a public changelog and ready posts, so the next visitor finds something new.
Start your changelog See a SaaS changelog example
Free to start. Pro is a month.
FoxPlug Pro
- A public changelog page and feed for each project
- Visual cards and walkthrough videos made from what you shipped
- Posts for X, LinkedIn and dev.to written from your changes
- Connect GitHub and the tools you build with