raghav

LMs from scratch · molecular tokenizers · sophomore cs

[raghav @1ens0r]

hit better val loss than karpathy as karpathy intended :)

im having fun

[raghav @1ens0r]

@superdapp yessirr

[raghav @1ens0r]

phase 2 running with bot tested training runs 😁

I'm probably going to sleep in another 15 minutes and let the agent do its job.

will write a short writeup after the modelling is done for the blog!

gn :)

[raghav @1ens0r]

finally done with about 5 assignments 🫡

now it is time to lock in, and test out the new deepseek v4.1 flash on opencode. ill be modelling a recommender system on data cleaned by my team. we already have beautiful graphs ready from eda.

gonna be fun :) will opensource it soon

[raghav @1ens0r]

some thoughts on deepseek v4.1 flash free on opencode :-

the model is actually nerfed in terms of context window offering only 200k tokens. will not be using this for the entire project as my codebase is pretty huge obv

will cook a writeup once the model's all warm and ready:)

[raghav @1ens0r]

finally done with about 5 assignments 🫡

now it is time to lock in, and test out the new deepseek v4.1 flash on opencode. ill be modelling a recommender system on data cleaned by my team. we already have beautiful graphs ready from eda.

gonna be fun :) will opensource it soon

[raghav @1ens0r]

@martinckeely yes but dont vibe code it like fortnite

[raghav @1ens0r]

@TimSweeneyEpic @wawilow108 @FortniteLinux

[raghav @1ens0r]

@digitalix banger video incoming :)

[raghav @1ens0r]

@0xSero waiting :)

[raghav @1ens0r]

finally done with about 5 assignments 🫡

now it is time to lock in, and test out the new deepseek v4.1 flash on opencode. ill be modelling a recommender system on data cleaned by my team. we already have beautiful graphs ready from eda.

gonna be fun :) will opensource it soon

[raghav @1ens0r]

@miniapeur fun fact : all other tabs are the same exact ones

[raghav @1ens0r]

@cneuralnetwork im convinved now that intelligence is commoditized we'll start to center everything around eq than iq

[raghav @1ens0r]

@Priyansh_31Dec bhaiya delivers always 💯

[raghav @1ens0r]

@opencode hey i dont see it on my opencode :(

[raghav @1ens0r]

@rento714714 always

[raghav @1ens0r]

after second revision :

nobody can ever get a 10/10 even against standard rubrics. so if my document is corrected by another llm, im probably ceiled at 9/10 even if everything is right.

this has gotta be fixed if we are adopting llm based correction methods

[raghav @1ens0r]

here's a cool experiment:

take a question paper with attached rubrics. upload it to chatgpt or any llm and ask it to generate answers to get a 10/10. make a document with the same answers and upload it back. ask it to grade it /10 against the same rubrics.

youll be surprised.

[raghav @1ens0r]

mine scored 8.5 btw

[raghav @1ens0r]

@zhilinjerrywag hi zhilin 😀👋

[raghav @1ens0r]

@opencode oh boy right on time

[raghav @1ens0r]

@tuemerkan ahem this is concerning

[raghav @1ens0r]

@VazeKshitij we just preserve the practice for the dopamine it actually delivers when its compiles instead of a segfault.

theres nothing we can do about it and in the future we probably wont even see the code just like we dont see compiled binaries