I mean I literally get json objects printed out as error messages… the bugs I see in these tools are bugs we know how to solve and expect to be solved when you’re not sitting on a literal pot of gold

It seems clear that they don’t think a high quality user experience is important for their bottom line (and empirically they seem right) but I don’t think we should pretend they’re operating at some Pareto optimal frontier of code quality and delivery speed.

Generating programs from natural language isn’t one of those problems, so that’s an unreasonable standard in this case. But my experience using apps like Claude and Claude Code is that they work correctly less often than I am used to expecting and it’s not a good look for vibecoding.

you’re not wrong but lets not forget we can do better than we do. we even know the best we can do: as long as no more than 1/3 of the machines involved have a hardware fault we can do the right thing every time for a huge class of problems.

would have a lot more respect for the AI companies if they would just acknowledge their tools are constantly broken!!!!

This is my favorite climate change chart. Japanese monks, aristocrats, and emperors kept meticulous records of cherry blossom festivals for 1,200 years and accidentally built the world's longest climate dataset.

I really do hate how in silicon valley believing surprising or counterintuitive things has become culturally valorized. Actually beliefs that sound wrong are usually wrong, and if you hold some you’d better have good evidence you aren’t just subject to confirmation bias.

Not the ones I work with! Maybe this is why i feel like people are exaggerating about compile times

This is really a bad dynamic in social media. When everyone has a personal brand, everything becomes viewed through the lens of promoting your own work. Like the Zig self hosting incident a few months ago. I’m sure I’m quite hypocritical here as well. Hard to escape.

impressive ability to talk his own book from Boris Cherny. There’s nothing counterintuitive about using an automated process to prevent human error; the only thing counterintuitive is suggesting it’s the same as “going faster” using AI, which is “counterintuitive” because it’s untrue.

You really don’t want to these to be traits, lol at your compile times if every function that takes a slice of anything needs to be monomorphized

Yes this is why Rust chose “git gud.” I wouldn’t choose that if I were designing a language.

My claim is that the cost in this case (an extra 8 bytes at the head of your heap allocation) is trivial for you compared to the cost imposed on code which doesn’t have clear ownership structure (deep copies of data to satisfy single ownership). So insisting on “zero cost” is a bad trade off.

No I mean figuring out how to structure code to only need one owned String. If you can do that you fall into the category where having a ref count would be a cost on you - the count only ever 1.

One answer is “git gud and know your ownership pattern.” Another is “use the bytes crate instead of vec, have fun with library compatibility.” But I think too large a portion of users fall into the worst option: “throw deep clones around because you don’t really know what’s going on”

For example in Rust because strings and vers are just arrays in the heap, reference counting them becomes more complex. If instead they imposed the cost of having a ref count field on everyone (not zero cost) users would deep copy them less often.

Zero cost abstraction can be quite a dumb idea because it sometimes imposes a cost on users in the name of not imposing a cost.

All these agent persona markdown files that people create feel to exactly me like when certain kinds of programmers obsess over creating, sharing and download vim config plugins

love to live a life where ive almost always muted the main character already

Obviously we are way closer to the vision than we were say 18 months ago, but I still find it hard to relate this experience to the belief promoted by Silicon Valley that the singularity is nigh.

I asked to raise the limit and include in the prompt that the cards should be just what I asked. With that change, the app started hanging forever. No further attempts to change it succeeded, after an hour I downloaded a less targeted anki deck from the internet.

Unfortunately, it did not fulfill my need. It went back and forth trying different things and after 5 or 6 attempts it produced something that did do the thing I asked. Unfortunately the maximum deck it could generate was 20 cards, and the cards came out very verbose.

I had cause to try this last night and we are not there yet.

I needed anki flash cards, Claude suggested it could create an applet to generate them from prompts. This is exactly what I can imagine as the viable UX for mass market AI - make the user a bespoke tool to fulfill their need.

only thing i miss from twitter is your takes

A lot of left critique is stupid and ignorant and deserved response but if your default posture is defending tech companies against posters you have fully lost the plot.

Gotta say I’ve never received a less enticing job offer than a cold email from your clawbot. Maybe just not a culture fit

This was originally expected to ship within a year of const generics - excluded from the MVP because it wasn’t quite finished. That was 2020 iirc

thank you for unfollowing, please never respond to me again

This episode had the same form as all of the others: yes people are behaving in an abnormal way because they feel threatened by this technology and they are probably unwell, but the way you guys raise it to a moral issue just demonstrates your myopia and privilege.

I know I’m like 6 months late to thinking about this but it was actually pretty fucked up when a lot of you on here suggested the “clanker” people were as bad as racists. It is perhaps a sin to express vitriol in this way, but it is actually not comparable to white supremacy.