Igor

I like computers and bossa nova
prev. ML @ Teachy, AMD, UIUC

[Igor @1xfalso]

me in buenos aires

[Igor @1xfalso]

abg (always be grokkin)

[Igor @1xfalso]

lawd today

[Igor @1xfalso]

quando o sol se põe vem o farol

[Igor @1xfalso]

also real productive activity is also suuuper boring for the most part

[Igor @1xfalso]

notice how the first sentence cannot be said for industrial machinery / manufacturing / agriculture

the global economic system is great at pricing future value, but often sucks at allocating resources to real productive activity, which is not speculative and hard to perform

[Josie Zayner @josiezayner]

@seekergupta If AI were deleted tomorrow. Life would go on for everyone, even sama and dario

If you have the two confused
Life is the thing you do when youre not creating shareholder value

[Igor @1xfalso]

damn I might actually need 350 hours of A100 40GB compute

[Igor @1xfalso]

@photomatt @wpengine @WordPress Thank you for fighting for a better internet. What WP Engine is doing is insane. Stay safe

[Igor @1xfalso]

gonna name my next LLM assistant Jesus see what happens

[Igor @1xfalso]

all lowercase never stress low cortisol agentic coding and personal handwriting

[Igor @1xfalso]

I severely overestimate what I can do in a day and I severely underestimate what I can do in two months

God be blessed, another day comes tomorrow

[Igor @1xfalso]

tl;dr model is decent at reasoning but terrible at agentic coding, as it simply forgets everything important after three turns

[Igor @1xfalso]

am now daily-driving Deepseek V4 Pro as an exercise in faith

[Igor @1xfalso]

there's too much alpha in all lowercase writing now

[Igor @1xfalso]

@_brentbaum can't believe I live in this timeline

[Igor @1xfalso]

@josiezayner That's reasonable. My biggest gripe with modern discourse is how eschatological everything seems to be, and how hard it is to not engage? I.e. even Ant/OAI talk about permanent underclasses and the quasi-end-of-the-world

[Igor @1xfalso]

@henrytdowling Fireworks is faster I think. On the official API I think it's just as fast.
How's it been for you?

[Igor @1xfalso]

am now daily-driving Deepseek V4 Pro as an exercise in faith

[Igor @1xfalso]

@henrytdowling It's definitely dumber than 5.5 but not by much. Its just as good in reasoning but for actual coding it seems to forget more and be lazier, whereas 5.5 just piledrives until completion.

As for faster I think it depends on where you're inferencing

[Igor @1xfalso]

@henrytdowling gpt-5.5 mainly

[Igor @1xfalso]

Your goal is not to prove you're smart. It's to make problems go away