Krisztian Kovacs

Full-stack ML engineer.

[Krisztian Kovacs @kk1694]

@boristane why store the graph? Can't agents just discover the necessary services on demand? maybe a bit more token-hungry, but you don't need to deal with keeping this graph up to date.

[Krisztian Kovacs @kk1694]

great article.

Saying AI will take most job is almost equivalent to saying that most people wont' be able to provide any value to anyone.

Also, besides the entertainment, there are many places where we have actual preferences for humans. Regardless of intelligence.

We pay a premium for human teachers, human nurses. We take a sales call with a person, not a robot. We go on dates to places with waiters/bartenders/barristas rather than to a microwave + coffee machine

[Krisztian Kovacs @kk1694]

@yongfook if it's really good; why not just make it and release it (the movie)

Instead of selling shovels, why not sell the final product?

[Krisztian Kovacs @kk1694]

@RichardHanania that's a great response.

[Krisztian Kovacs @kk1694]

@thdxr isn't this mostly a hardware issue? AWS (or anyone else) isn't be able to rack the GPUs in time as demand increases.

[Krisztian Kovacs @kk1694]

@RichardHanania By this logic feudalism really protected the serfs

[Krisztian Kovacs @kk1694]

But models are not trained to survive / replicate / seize resources!

They are trained to assist humans in economically meaningful ways. Do a task and then disappear.

You could say that survive / replicate / seize resources will naturally emerge (instrumental convergence) - but to me this sounds like a really contrived argument.

When I ask codex to fix a bug, it will try (and usually succeed) in fixing a bug.

It doesn't go "to fix the bug most effectively, I first have to take over the universe".

[Krisztian Kovacs @kk1694]

@theo also: if these chinese labs use proxy services, how do you identify which chinese lab they belong to?

They say IP address correlation (huh, I thought they used proxies?), metadata (???), infrastructure indicators (what does this even mean)?

[Krisztian Kovacs @kk1694]

yeah, was thinking the same. Also, so many other wholes:

"Illicitly distilled models lack necessary safeguards, creating significant national security risks"

How could this be? If the Anthropic API rejects a request for safety concerns, and you distill on that, don't you have the same guardrails? How would distilling on rejected request make you less safe?

[Krisztian Kovacs @kk1694]

@yongfook it's BS now, but it used to be important.

1 year ago adding magic incantations really did improve output. (the whole idea of reasoning models came from the observation that saying "think step by step" gave better results).

current models are smart enough not to need it though.

[Krisztian Kovacs @kk1694]

also: colocated tests and docs!

The LLM reads the function definitions, and right next to it, the examples on how to use it. No need to do separate file reads.

Great for writing as well: you change a function and only need to update one file, not file + test file + docs separately.

[Krisztian Kovacs @kk1694]

@esrtweet Not quite correct. Pretraining indeed caps at the best human engineers. RL does not.

[Krisztian Kovacs @kk1694]

google: start as a paper, build the tech, try to sell / license it, nobody buys, okay- let's build a product

openai: start as a non-profit, try various forms of deep learning, this language model (gpt2) seems interesting, okay let's scale up, sell gpt3 on the API, (a year later) why did nobody build a chat app on top of it, let's try that

[Krisztian Kovacs @kk1694]

@theo Qtile (window manager) has a layout called tree tab. Kind of like using tmux but for all kinds of windows, not just terminals. Haven’t seen it in any other window manager though.

https://t.co/yUEoewqbzX

[Krisztian Kovacs @kk1694]

@thdxr @kentcdodds Sometimes I find chaining up instructions is better than longer prompts.

Examples:
- add tests
- make the component mobile friendly
- (previously) remove comments
- refactor that

[Krisztian Kovacs @kk1694]

@yacineMTB spend is not a good metric - imagine someone releases a super cheap super smart model. Most people will choose it, yet spend little

[Krisztian Kovacs @kk1694]

@theo finally signed up

[Krisztian Kovacs @kk1694]

@aboodman this looks really cool. Might be a stupid question, but what are some of the main differences relative to convex?

[Krisztian Kovacs @kk1694]

I’m very curious how much money we are talking about here.

My guess is that it will be tiny- not worth the trouble for most sites.

I wouldn’t even be surprised if it turns out to be net negative: if you ask ChatGPT/Google to pay; they’ll just use other sites to formulate their answers. (And have reference links to other sites).

Only works if you have some truly important/unique/valuable data. But in that case you’re probably not on the open internet anyway.

[Krisztian Kovacs @kk1694]

Also add cost of capital.

Say you borrow money at 10% (or can invest your spare cash at 10% opportunity cost). That’s 2.5k / year or $8/day. Add in electricity and you’re pretty close!

(Obviously, most providers raise equity, not debt; but that should increase cap cost. Plus this assumes H100s live forever)