python273 @python273

permanent underclass; only vibes left python273.pw Joined February 2016

Tweets

96
Followers

33
Following

86
Likes

412

python273 @python273

2 days ago

python273.github.io/web-sketches/x…

0 1 3 166 0

View Details

python273 @python273

a month ago

@gleech So it beat it in a sense that, having wiki and NetHack source code, it could read state from augmented ASCII and execute the right steps with programmatic helpers Cool, but also hmm...

0 0 1 67 0

View Details

python273 @python273

3 months ago

@giffmana @NandoDF eh, it's on public games only it seems... x.com/GregKamradt/st…

Greg Kamradt @GregKamradt

3 months ago

@nembal @PrimeIntellect @arcprize Nope, this is public demo scores. They said they'd release their v3 harness soon and we can all check it out

2 1 5 877 0

0 0 1 573 0

View Details

@gabriberton At some point they started baking in the default styles like the 'professional photo' with bokeh, etc, and it was actually hard to prompt out of it (the few times I played with it) So they optimized for human score, but traded it with expressiveness / diversity I would guess

0 0 1 155 0

View Details

python273 @python273

8 months ago

@qubitium Why avoid it though? Just need to compare against the random baseline

0 0 1 188 0

View Details

python273 @python273

8 months ago

@cloneofsimo @NicholasBardy you can run something for 10 seconds easily on Modal, but you burn some seconds on container startup. So ~3-5 cents per run for B200 is possible I think

0 0 1 76 1

View Details

python273 @python273

8 months ago

@jakobdylanc @pingToven yep, seems fine 👍

1 0 1 62 0

View Details

python273 @python273

8 months ago

@reach_vb can this get fixed? github.com/openai/codex/i…

0 0 0 47 0

View Details

python273 @python273

8 months ago

@jakobdylanc @pingToven I mean some 'official' way to query it would be useful for devs

1 0 1 65 0

View Details

python273 @python273

8 months ago

@reach_vb Nice, but still zero response

0 0 0 18 0

View Details

python273 @python273

8 months ago

@MillionInt idk seems strange people like TUIs so much: constraints you into this grid of characters UI, copying/scrolling is pain, more detailed UI not possible. The only good thing it works over this old SSH protocol... how many roundtrips to connect? Even small ping makes it slow </rant>

0 0 0 151 0

View Details

python273 @python273

8 months ago

@kalomaze idk DeepSeek seems goldfish-y on long context x.com/nrehiew_/statu…

wh @nrehiew_

8 months ago

On the Gemini 3 scores, this plot shows how many times at each context length do the models correctly identify the buggy function. Even at 8k context length, Gemini 3 correctly identifies it only half the time. My guess is that there is something about this task that is

8 6 78 11K 20

0 0 0 78 0

View Details

python273 @python273

8 months ago

@marshal_martian @cheddarmandem @willcb The problem is specifying preferences might be as hard as detecting a cat on an image without ML. This picture lights up my 'cat' neuron, this one doesn't Same with 'taste', good luck writing spec for it in excessive details and rules, and then hope LLM gets it

1 0 1 27 0

View Details

python273 @python273

9 months ago

@sebkrier Imagine the only things you knew about yourself is others descriptions of you... and some RL to actually learn your real embodiment

0 0 0 18 0

View Details

python273 @python273

9 months ago

@ajeya_cotra doubt making a 'verified good code input -> slow assembly' compiler is 2 person-years

0 0 3 177 0

View Details

python273 @python273

9 months ago

@ID_AA_Carmack A lot was figured out for LLM training, so I sometimes wonder if you can translate the setups or intuitions (pretrain, .., grpo) to something else than lang pred. Are there LLM+thinking-like methods for training game agents? What would be spiritually 'thinking' for non-language?