I think the real question you're asking is: why isn't AI coded software replacing well-known incumbent software (office, adobe, etc...)

Because I see AI coded software everywhere: HN has been complaining about the recent influx of Show HNs due to AI. Take a specific product category: there are about two new dictation apps on Show HN every week.

Here are my theories:

- adobe/office are simply too complex for AI to replicate (as of now). Theo Browne has been testing all the frontier models with a relatively simple fish game. The results have been subpar, at best. I don't think anyone would prefer the AI-generated games over the original. Even if the AI-generated version was free.

- Most AI-coded software won't be distributed like classic software. Instead of a bag of hundreds of features (adobe/office), AI will be used to create the specific 5% of the features a single user really needs (home-cooked software). The software may even be per project or even per task. Why spend 20X to 100X the tokens/time when you only need a small subset of the features?

- AI may not be able to replace human "taste" or design. There is a lot of design that went into adobe and office. I don't think AI can fully understand all the "why's" of these designs. We can already kind of see this with open source: Gimp and LibreOffice are alternatives to adobe/office. Yet most people prefer the paid options.

- Even if AI solved coding, there are other barriers like marketing.

Some examples:

- I built my perfect window management system with the help of AI. It suits my very specific needs very well, but it's nowhere near ready for distribution. So this is AI-generated software no one else will ever probably see.

- Going with the home-cooked meal analogy: there is a dish served at all Chinese restaurants in Korea: jjajangmyeon. However you can't find it in China because that dish is only found at home, not at restaurants in China.

I think that they use a lot of LLM and they start with a good baseline so they still win

My first impression is it's very similar to the stock iOS weather app Apple provides by default (feature/layout-wise). I couldn't see any major differences based on the description here or in the app store. That's what kept me from even trying it.

So my advice is to highlight the major benefit over the iOS weather app that ships with every iPhone.

The only major difference I could see is the data source. Did you intentionally not use Apple WeatherKit? (If so, what was the rationale?)

My HN reader has QoL features for reading such posts: https://hn.leftium.com/i/49630931?q=%22aspect+ratio%22

- Comments have 3 LOD (L:full, M:single line, S:collapsed)

- Optimal default LODs

- Simple to expand groups of comments (direct replies, comment thread ("spine"), full sub-tree)

- Highlight specific users like OP, or other interesting users

- Search for text within thread (automatically expands LOD)

- Highlights new comments (since last visit or set time)

It matches the way I want to read HN comments, anyways.

> Compression is prediction

(According to recent article on HN[1]. So LLMs _are_ compression!)

[1]: https://hn.leftium.com/i/49263497

Tip: you can also add widgets to your desktop

1. Click menu bar time/clock

2. Click Edit Widgets...

3. Search for "Codexbar"

Apple hardware uses a unified architecture. That means the CPU and RAM are integrated together: it is very difficult/impossible to strip the RAM from these systems.

This unified architecture makes Apple hardware very good for AI work, where latency between RAM and CPU is very important

Even the SSDs (which are normally more strippable) use a proprietary hardware form factor.

I think the article mentioned the real draw:

> Apple noticeably promoted the ability to link multiple Mac Studios together into a single, more capable system for running large frontier AI models, a feature aimed at business and developer customers rather than everyday consumers.

(Couldn't read the source cited, might have more info: The Information)

I have been considering adding labels/filters based on sentiment analysis to my HN reader. I wonder if your classifier-based method would be more effective.

I added filters that blur posts from new HN accounts based on age/karma thresholds (as suggested by another HN user). However, instead collapsing, I just blurred them. This allows quickly peeking via hover. My reader also highlights posts with higher points/comments, and those orange highlights survive the blurring.

This is what blurring looks like: https://hn.leftium.com/newest?min_karma=500&min_age_years=5

That’s very nice. Feel free to integrate if you want :) Just request https://classify.stylometry.net/violations/49458161 or whatever item ID (as long as it was posted later than a few weeks ago) and it will return the scores for that post/comment and all its children.

wow thanks! I integrated like this: https://hn.leftium.com/i/49449507?violation_threshold=0.5

The url param temporarily sets the threshold, you can permanently set it via the settings.

My initial assessment is 0.5 setting is too aggressive; the posts that I generally don't care to read seem to violate multiple rules at once.

It's a shame we can't see comment scores so orange highlights could shine through the blurring on "good" comments.

Wow that was quick :) Meh, I feel like comment points have become somewhat untethered to comment quality (at least assuming threads are roughly sorted by comment score). It’s a good heuristic usually but it’s wrong often enough for it to be difficult to rely on.

I want to improve the classifier scores too. There’s some parameters in the loss function I could play around with more when training which basically optimize for more false negatives or false positives. Problems with no ground truth are very frustrating lol (though I suppose dang/tomhow’s manually written warnings could be treated as labels instead of using labels from GPT, but I suspect there’s not enough of them).

My HN client has made HN easier to keep up with, even before AI: https://hn.leftium.com

I call it "Seeking Orange" because interesting posts are highlighted. Features include:

- Point/comment counts become increasingly more orange with higher counts (50+, 100+)

- Orange accent bar on left indicates new items since last visit

- Support for several different feeds including: HckrNews (chronological front page appearance time), Classic Stories, Best Stories, Active Stories

- Thresholds to de-emphasize posts from users with low karma/recent creation dates: https://hn.leftium.com/newest?min_karma=500&min_age_years=5

Nice

When this was posted to HN, a lot of people could relate: https://sive.rs/slow

For me, I'm not sure if "slow" is the right word. I think "delayed" or "sporadic" is more accurate.

Thoughts for me are like when a word is on the tip of your tongue, but you can't think of it right away. Often you'll think of the word later: you weren't actively thinking about it, but the thought was simmering in your mind and just randomly pops up.

Wow Sivers a decade ago! Thanks for bringing it up.

You can add instructions in the CLAUDE.md/AGENTS.md file to modify output behavior like:

    # Response Style

    Keep routine answers concise. Lead with the result, include material caveats and the next action, and omit introductions, repetition, generic reassurance, and optional background. Expand only when the task requires explanation.

While writing a document I got good mileage out of doing something like this for every paragraph:

    Let's analyze this paragraph:
     - List the main points.
     - Can any points be removed? (Do any points need to be added?)
     - Reword each point. Be as succinct and clear as possible.

I'm not sure how many extra tokens it would consume to do this for every output. But it may be better than just asking for "TLDR pls"

(I've actually never used Claude Code. I've never had problems with verbosity in harnesses like OpenCode or T3-code. And recently I've been using Codex, not Claude models.)

no, you can't. Opus 5 decides by itself that it is better to ignore those instructions.

I've noticed Opus 5 repeatedly ignored my Claude.md in several ways. Particularly when it comes to verbosity of comments in code.