zerd

Claude just used the word “ Netowrk” instead of Network. Mind you, I was only using Sonnet with reasoning, but I thought trillion-parameter models would be able to spell a common word.

Gemini is the worst gaslighter. It hallucinated a feature in an open source product. I said no that does not exist. It invented very realistic git commit messages. I said commit does not exist, so it said I was out of sync. I linked the repo and it said it might be a fork. Then it invented a mailing list thread where they supposedly removed the feature, that’s why I couldn’t find it.

Searching for anything product related is now AI answer, followed by “Sponsored products”, followed by “Popular products”, a few hits, then “Shop by store”, and “In store nearby”. Maybe 2% is actual content, the rest is ads.

Their prediction was spot on: “We expect that advertising-funded search engines will be inherently biased toward the advertisers and away from the needs of the consumers.”

Limiting it to just read and http limits its usefulness. If you want something like “filter for this, count the number of matches, format like this” you have to make custom tools. And you have to make sure they don’t have any bugs that allow arbitrary code execution. You’re effectively building your own sandbox in this case. Using a standard one sounds easier unless you have a very focused use case.

Time to start some new restaurants matching those. Like Bubba Gump.

How do you scale past 16 channels though? 16 channels give you around 614 GB/s, while PIM can do that per chip, so it can achieve 58TB/s.

Veritasium did a video on this 4 years ago. https://www.youtube.com/watch?v=GVsUOuSjvcg

I’ve been doing the same thing, giving the same tasks to Qwen 3.8 27B and Opus, and the main difference is that Qwen does not consider edge cases which Opus catches. It’s good at the happy path, but even when hinting that there are uncovered edge cases and gotchas it’s oblivious to it. So I feel like I need a bigger model to do planning/review.

"We're doing it for your own good, pinky promise"

I noticed Claude code does something similar when "Exploring" the code base, it spawns a subagent using Heroku to get the summary and file locations, which is both faster and it doesn't pollute the context as much.

I wonder how it compares to an vector indexing approach.

I'd love an RTX 6000 Pro, but how can you justify it when it costs 10 years worth of Claude Max?

I noticed Fable was quite a bit terser, and I think it's due to changes in the system prompt [0]. They're literally saying "just give me the TLDR" and "give brief updates". You can tweak a lot of that with an AGENTS.md.

[0] https://twelvetables.blog/comparing-claude-fable-5s-system-p...

Thunderbird does not seem to have auto-categorization from what I see, just filtering. Neither does KMail. Unless you’re referring to some addons? For apple mail you have to add it on each client. And a lot of comments are about how to disable it because it categorizes wrong.

Reddit “best” sorting is pretty much like instagram and TikTok now, have to make sure it on hot/top, otherwise it’ll show you “related” things from subreddits you never subscribed to.

Also it breaks copying links. If I want to link to an issue I copy the URL. But now there's two different issues open at the same time, which one am I linking to? Original? Popup? Both?

Based on the communication fix, they also didn't have a simulator, or tests, or complete source code, on a custom instruction set that wasn't well documented, so they had to reverse engineer how it worked. https://www.youtube.com/watch?v=YcUycQoz0zg&t=2366s

It's ok because I rarely need it, but sometimes it's so slow I'm looking for alternative install methods while it's doing it's thing.

I was making a telegram to Claude via tmux capture-pane and send-keys, this will be so much nicer. Also sounds like something that addresses some of what Steve Yegge said was missing for agent to agent communication as well.

I just watched 20 minutes on the gardening channel and learned a bunch that I never would have seen with the YouTube algorithm, and wouldn’t have thought of to search for.

Sounds similar to these quotes from the article:

> we found that employees worked at a faster pace, took on a broader scope of tasks, and extended work into more hours of the day, often without being asked to do so.

> On their own initiative workers did more because AI made “doing more” feel possible, accessible, and in many cases intrinsically rewarding.

If it's DLP then using alternative file browsers should also be affected, right? Which at least in my case it isn't.

Had the same issue with slow file explorer in Windows 10. A couple of things helped a bit, such as disabling "Show recently used files" and "Show frequently used folders". I also cleaned up the Quick access list. For some reason if you have a network share there it makes browsing local dirs slower, go figure. It's still not instant but a lot faster than the 3+ second delay.

I tried OneCommander and they're super fast, so it's not something slowing down disk IO, it's purely File Explorer.

Now I'm still struggling with closing chrome tabs being super slow sometimes.

I was about to do exactly what OP did and create a chrome extension when I found karakeep which saved me from doing that. I really like the full archive because sites disappear all the time, and screenshots for a visual overview. Used to use pinboard but didn't like that archive feature was a subscription. It also works with SingleFile to archive logged-in sites.

Even after 40 years of battle-hardening, it has had buffer overflow and double free vulnerabilities discovered recently, which Rust protects against. sudoedit one was pretty bad. https://www.sudo.ws/security/advisories/

The "perfectly functional one" from 1980 also has multiple critical vulnerabilities in it in recent years.

It's funny that converting the first example to the second is a common thing a compiler does, Static single assignment [0], to make various optimizations easier to reason about.

[0] https://en.wikipedia.org/wiki/Static_single-assignment_form

WebSocket doesn't specify data format, it's just bytes, so they have to handle that themselves. It looks like they're using Arrow IPC.

Since they're using Arrow they might look into Flight RPC [1] which is made for this use case.

[1] https://arrow.apache.org/docs/format/Flight.html

[citation needed]

I've had LLMs cite me bullshit many times, links that don't exist and claiming it does. It even cited a very realistic git commit log entry about a feature that never existed.

Haven't yet had the same issue with Wikipedia.

I've had issues in a database with billions of rows where the PKs were a UUID. Indices on PK, and also foreign keys from other tables pointing to that table were pretty big, so much so that the indices themselves didn't all fit in memory. Like we would have an index on customer_id, document_id, both UUIDv4. DB didn't have UUID support, so they were stored as strings, so just 1 billion rows took ~30 GiB memory for PK index, 60GiB for the composite indices etc. So eventually the indices would not fit in memory. If we had UUID support or stored as bytes it might have halved it, but eventually become too big.

If you needed to look up say the 100 most recent documents, that would require ~100+ disk seeks at random locations just to look up the index due to the random nature of UUIDv4. If they were sequential or even just semi-sequential that would reduce the number of lookups to just a few, and they would be more likely to be cached since most hits would be to more recent rows. Having it roughly ordered by time would also help with e.g. partitioning. With no partitioning, as the table grows, it'll still have to traverse the B-Tree that has lots of entries from 5 years ago. With partitioning by year or year-month it only has to look at a small subset of that, which could fit easily in memory.