Viewing profile — dymk

The wire format certainly isn’t strict

I know what they mean, the “rewards” are not free

Credit cards are not free money lmao

Well, no, it might not be important. Maybe you won't use that knowledge about CSS grids and flexbox for another year or two, or maybe ever.

That doesn’t mean “use fewer thinking tokens”. It might use more as it mulls over how to make its response concise.

Projects like SkillOpt[1] are pretty good evidence that the ceiling for skills isn't "describe it in a sentence and let the LLM generate it". There's value in iterating on a skill …

No, it was a viral tweet from last week. They’ve been verbose since GPT-4o.

The point of STE100 is to avoid ambiguity like this, though.

Precommit hook

You can memorize a book too but that doesn’t mean they’re not useful

I’ve noticed much the same thing. Adding rules for ruthless minimalism and no anticipatory features helps, if context hasn’t degraded too far. Still, I tend to ask it to remove cru…

Backyard trampolines without nets are particularly dangerous. A gymnasium or trampoline park with in-ground trampolines are my go-to if someone wants to learn how to do tricks.

This feels like a subtle attempt to start an argument about Star Wars being sci-fi or not.

Front page says 150MB RAM, About page says 90MB. Whole site is full of LLM-isms (em-dashes, stilted 'X. Y. Z. .' constructions). It's so boring. It says "We put one prompt worth of…

He had schizophrenia.

A train system is a much safer bet than praying traffic won't be bad

Then sign in, thats enough signal for them to track you

They certainly have knives, go look at what a tiger is sporting on his paws

What a button look like to you? Blue roundrect sure says button to me.

Then let's not pretend what software engineers have been doing for the last 50 years is engineering, either. But what else would you call it?

Understanding how to write good prompts (prompt engineering) is still very much a relevant skill if you want to effectively use LLMs. Harnesses aren't magic.

The weights are the output artifact, the training corpus and system are the source.

If that’s the case I still don’t get it. 640k was enough in 1981, same as how 150 (or 9k) tok/sec would suffice for 2026 The comment seems like the nerd equivalent of 6 7