apparently you make $16,000 for a semester of AI research in a PhD program. but $2.5M if you clone doordash in next.js and sell it to an AI lab something isn't adding up…
@jxmnop
-
Backpropagation Through Sampling: Implications for AI Society
By
–
society if it was trivial to backpropagate through sampling
-
Training and Testing on Everything: Are We Gaming ML Benchmarks?
By
–
it’s been an interesting ride watching the conventional nomenclature of machine learning gradually lose all meaning. there used to be TRAIN and TEST and everything was simple. now we train on the universe. and we test on the universe, too. are we gaming our benchmarks? are we
-
OpenAI’s Journey from Open Research to Closed Commercial AI
By
–
> be openAI, circa 2017
> do lots of interesting research
> all open source > no one cares that much
> invent chatGPT in 2022
> get too busy to do open research
> everyone gets mad
> wheres the open research
> more like closedAI
> three years pass. now 2025
> finally release an -
Curriculum Learning Makes Comeback Without Random Sampling
By
–
curriculum learning will make a comeback. unfortunately, random sampling is not the way heard it here first
-
Rubric-Writing vs Prompting: New Frontier in Model Optimization
By
–
rubric-writing is much more interesting than prompting look at Kimi K2: > Responses must not begin with compliments directed at the user (e.g., “That’s a beautiful question”). abstract behavioral descriptions baked directly into weights there's almost no research on this btw
-
Recommendation Systems Have Been Problematic Forever
By
–
i guess recommendation systems has been like this since forever
-
Private Industry RL Research Gap: LLM Judges vs Automated Rewards
By
–
for the first time i am aware of, there is an entirely private subfield of AI research every company that actually trains models is doing RL with rubrics and LLM-judged rewards but academic work is stuck on RL with automated rewards (math problems and code). much cleaner for
-
Regularization During Training for AI Models
By
–
couldn't you just regularize for this during training? i bet it'd work fine