Pratyush Maini

@pratyushmaini

Assistant Professor @cornell_tech & Founding Team @datologyai | PhD @mldcmu | BTech @iitdelhi

pratyushmaini.github.io

Joined November 2019

Replies Reposts Media

Pinned
Pratyush Maini
@pratyushmaini
Mar 19
If I had to compress my PhD into one idea, it is this "The data a model sees early in training leaves an imprint on its representations that is very hard to undo later" This thread runs through - Rephrasing the Web - Safety Pretraining - TOFU This is the Finetuner’s Fallacy🧵
00:00
21
Pratyush Maini
@pratyushmaini
Sep 1
If you lead your vertical, you have years of data that exists nowhere on the internet. That data is now your most valuable asset and the way to realize that value is to train your own models on it @datologyai x @thomsonreuters is our first public example of what that looks like
Aaron Levie
@levie
Sep 1
Now that the base open weights AI models are getting far better, and post training infra is becoming more mature and commercialized, there are going to be all new plays for companies that have large amounts of data to have their own models. Licensing data for external model
2
Pratyush Maini
@pratyushmaini
Aug 27
It’s incredible how the surge of money into AI made everyone feel poorer. At the start of the decade, researchers made less, but slept better. Now everyone makes more, compares more, and feels they’re getting less. Somewhere along the way, money acquired satisfaction?
2
Pratyush Maini
@pratyushmaini
Aug 18
The best data researchers are obsessive about looking at data. But doing that at scale takes incredible time and care. This is exactly where swarms of agents can tirelessly sift through petabytes of data across data curation pipelines. People should look at autonomous data
Oliver Li
@oliveraochongli
Aug 18
1/ Finally get to share what I’ve been working on this summer at @datologyai. We built DataSmith 🍊, an automated data researcher. The idea is simple: scale test-time compute across multiple directions, then let agents synthesize their findings. Blog: datologyai.com/blog/datasmith
00:00
3
Pratyush Maini
@pratyushmaini
Aug 18
true story
Haoli Yin
@HaoliYin
Aug 18
General coding agents often lack research taste: they make small tweaks and get stuck in local optima. We built DataSmith to automate data research. It inspects failures, searches wider, and makes better research bets, beating Claude Code by 5.1 pts on avg. 1/n
1