@andand_blr@SebAaltonen At a certain scale, bad architecture directly causes poorer performance, and it's not as simple as getting both. The model's fixes to increase performance are often way, way more complicated than just getting the architecture right on the first go-around
Anyone telling you to rewrite your app in a rust/C#/etc implementation from a more readable codebase (python/ruby/go) is grifting.
Do you want to spend 10k in tokens and a week fixing resulting bugs, and only be able to use AI to ever edit again?
Then by all means go ahead
@DuaneJRich This is what all the "the code is the documentation" people don't get. Turns out comments that discuss why a design choice was made, shockingly, provide the agent more context to not overwrite that design decision
@SJDauncey Fascinating! I'm so curious to understand how the tokenizer works during test-time inference - can it output super-tokens?
Would be extremely interesting to analyze the average token length across different corpii.
๐บ๐ธ Married + home by 30 years old:
1950: 50%
1960: 52%
1970: 48%
1980: 45%
1990: 43%
2000: 35%
2010: 25%
2025: 12%
This is a devastatingly bleak statistic
@iglesias_gabby What a joke. Trump promises peace by surrendering Ukraine to the aggressors that have been bombing its cities for almost three years.
Russia is a despotic, imperialistic regime. Ukraine is trying to move towards the West, towards the US. In what world do we support Russia?
@TropicalWatcher Aggregating quite quickly. The Eastern flank of the wave is catching up with the vorticity to the West, and the low-level and mid-level circulation is visible.
Wouldn't be surprised if we saw more development prior to the lesser Antilles than some models are showing.