rayofsun

23, Software and Systems. - Play games and bond with LLMs

[rayofsun @rayyyyyofsun]

Umm

[Lloyds @LloydsBank]

We’re aware some customers are having issues with our app and online banking. We’re really sorry about this. We’re working hard to fix it and will let you know as soon as we’re back to normal.

[rayofsun @rayyyyyofsun]

AI should never give you ideas,
You give it ideas and it should help you build

[rayofsun @rayyyyyofsun]

"needful" is a bullshit word, and no other way

[Mitchell Hashimoto @mitchellh]

I've got an agent in a loop optimizing a renderer with the goal to minimize frame times (and tests to measure). It got times down from 88ms to 2ms and allocations down from ~150K to 500. Sounds good, right? Wrong. This is exactly why agent psychosis is a big fucking problem.

As an experiment, I rewrote the Ghostty core render state in Go, with access to identically laid out data structures as Ghostty and the exact same validation tests. I made a purposely naive renderer (simple, correct, but slow). 88ms per frame with 150,000 allocations (horrendous, lol)!

I then kickstarted a Ralph loop to bring the frame times down. I told it it can't modify input data structures or the public API or tests (they're correct), but it can do anything else it wants. It got to work.

It has worked for about 4 hours. I've spent around $350 on this experiment so far. The results?

88ms => 1.5ms
150K allocs => ~500 allocs

Incredible right? Nope.

My hand-written renderer I ported has frame times (same benchmark) of ~20us (0.020ms) and 0 allocations in the update path.

This is the problem with psychosis and lacking systems understanding. If you don't understand the system, you're going to accept that this is an incredible result. If you understand the system, you'll see better solutions immediately and can do roughly 75x better on throughput.

The people who blindly trust agent output are in the former camp. They're sheeple, overdrinking from a fountain of mediocrity.

Standard disclaimer: I use AI all the time. I like AI. The point I'm making is to not blindly accept results. Think. Analyze. Learn.

[rayofsun @rayyyyyofsun]

You learn something new about javascript everyday

[antirez @antirez]

@ericc__ch @mitsuhiko In the most nightmare-ish way.

const re = /foo/g; re.test("foobar"); re.test("foobar");
The first is true the second false. Great design.

[rayofsun @rayyyyyofsun]

@orcdev It's annoying to the point, it doesn't give you time to look at your browser tabs in between prompts.

[rayofsun @rayyyyyofsun]

First time seeing a model reciprocating the vibe instead of being excited all the time

[rayofsun @rayyyyyofsun]

so composer-2.5 loves monorepos yeah, and jsut npm workspaces, no turborepo bull shit.

Yeah Im down

[rayofsun @rayyyyyofsun]

The prompt on that left, that exact one I have written for https://t.co/NPX0X2M8c0 roughly 90 days back.

[Micky @Rasmic]

Codex is elite....

[rayofsun @rayyyyyofsun]

Idk one more domain, but the other old one costs almost 40$ to renew.

Should have opted to subdomainise projects

[rayofsun @rayyyyyofsun]

I have got the most atrocious Typescript pun, you'll feel it soon

[rayofsun @rayyyyyofsun]

@CloudflareDev @dok2001 cc: @ritakozlov 🫠

[rayofsun @rayyyyyofsun]

Okay 60sec for a 10K out tokens for 20B model is notorious, and Kimi-K models never work using @CloudflareDev gateways.

pretty SAJ and MAJ

[rayofsun @rayyyyyofsun]

Working with Workers AI bindings, and there are some issues right now. (Not ground breaking at all, but some QOL features)

1. kimi-k's were failing via gateway for HTTP: 408 (??) (img-1)
2.llama-guard-3-8b listed as a Text Generation model but is just a classification model.(img 2-3)

[rayofsun @rayyyyyofsun]

@CloudflareDev cc: @dok2001

[rayofsun @rayyyyyofsun]

Finally got to do the cliche, Vidyarthi Bhavan and Lal Bagh walk.

[rayofsun @rayyyyyofsun]

nvm found a /models url in depths of Documentation, but it is behind auth for some reason.

[rayofsun @rayyyyyofsun]

Working with Workers AI bindings, and there are some issues right now. (Not ground breaking at all, but some QOL features)

1. kimi-k's were failing via gateway for HTTP: 408 (??) (img-1)
2.llama-guard-3-8b listed as a Text Generation model but is just a classification model.(img 2-3)

[rayofsun @rayyyyyofsun]

@ritakozlov sorry to bother you here

[rayofsun @rayyyyyofsun]

2. (Cont) The models listed in dashboard and in Workers AI documentation are different, list of models. It would be helpful to have a /models endpoint like openrouter to fetch all the models, in a gateway may be or available via Workers AI bindings.

3. 10k neurons are included per day via Workers AI plan, and they are not exhausted (img-4) 9.79K/10K shown in Workers AI, but it shows as $0.10 in gateway.
($0.10 is a very less amount to be worried about tbh, but for just a clarification that amount shown is exclusing the free neurons per day via Workers paid or additional amount that day is confusing)