
AutoResearch MTS @hiloopai prev @reductoai @moondreamai @so_claros
your KV cache
Joined April 2021
- Tweets7.8K
- Following813
- Followers899
- Likes9K
Vatsa Pandey retweeted
there’s a trillion dollar pdf somewhere in Qwen3-8B you just have to tickle the right neuron to get it out
Tell me Oppenheimer, in making le bleach costume, you did le one piece slander
Yes Albert, yes I did
"eigenspace the esoteric math concept"
Reducto X Pangram ?????
Wow Based
What matters when you can generate anything?
Join us for a Non-Slop hackathon with @NousResearch and @reductoai at the Pangram HQ in NYC November 22
Grab your spot now 👇
there are many many OOMs in cost/efficiency to be gained with swarms, by eoy 2028 I'd bet a full swarm might run cost the same as one session from a frontier model today
the swarm format will probably use 10x less inp tokens and still have stronger results than ever
some of the best r-1 evals came from documents we weren’t sure it could handle. think 10m+ pixel construction blueprints, 32k+ output tokens for a single page, and even the muffin-vs-chihuahua captcha.
In this video, @TheOtherGandhi_ and I discussed why the weird tests are often the (most) useful ones and what they taught us about r-1’s capabilities and limitations! we challenge you to keep sending us your strangest documents.
When you let the agents pass messages to each other and they're slopping each other up and not one of them has asked a clarifying question or even knows the definitions the other one is using
Vatsa Pandey retweeted
Replying to @ar0cket1
pass@k != pass@1 — but increasingly does when you've squeezed all the juice there is to squeeze out of a model in post training (without going back for more midtraining) :)
they think figure summary cant handle it, they're in for a surprise
Replying to @donaldwu_
mine's smart, she knew i'd use reducto so she used gifs on each page even the lawyers didn't know how to parse it
"This lesson is free, you are welcome"
> Benchmarks have become marketing
1. We tested one of our competitors on their own benchmark dataset. Great score.
2. We then added some random noise to the same documents: tiny tilt, small while pixel padding, nothing crazy.
3. Their best model/api option struggled on the *SAME* data points. Score went down, free fall, even on easy data points.
We have receipts.
You can overfit on the benchmark test sets, but please also remember to apply simple data augmentation techniques to your next training run. This lesson is free. You are welcome.
if I do it long enough @code_star and @xeophon are going to pull up like this to get another data reader out of my insanity

























