Aman

Prev. @allen_ai

[Aman @not_aman17]

@IanMagnusson Congratulations Ian, such a cool work.

[Aman @not_aman17]

@reflection_ai Huge Congrats!!

[Aman @not_aman17]

@natolambert @trillium_labs Congratulations!!

[Aman @not_aman17]

@industriaalist 👀👀

[Aman @not_aman17]

@latkins @stochasticchasm Congratulations

[Aman @not_aman17]

@saurabh_shah2 Congrats!!

[Aman @not_aman17]

@kylelostat Congrats!!

[Aman @not_aman17]

@soldni Congrats!!

[Aman @not_aman17]

@natolambert @allen_ai Very fortunate and lucky to have worked with you. All the best 🐐

[Aman @not_aman17]

@saurabh_shah2

[Amanda Bertsch @COLM @abertsch72]

New paper! https://t.co/1ETmDt0ZB8

This tackles a puzzle we found during the training of Olmo 3: how could two models with nearly identical short-context performance (and trained on the same data!) behave completely differently after long context extension?

[Aman @not_aman17]

@kylelostat @allen_ai Learned a lot working with you, Kyle. Best wishes for future endeavors.

[Aman @not_aman17]

@soldni @allen_ai Had great time working with you Luca. Best of luck for new adventures!🫡

[Aman @not_aman17]

@HannaHajishirzi Grateful to have learned from and worked with you.

[Aman @not_aman17]

Great work! It’s a nice insight into how window sizes interact with recurrent memory in hybrids!
In my own experiments with SWA and similar hybrid arch’s, I’ve encountered issues where weight matrices in certain layers collapse to low rank, often correlating with entropy spikes. Did you observe anything similar, and if so, do you have analyses of layer wise effective ranks or attention weight distributions?

[Ethan Shen @ethanlshen]

Today, we release SERA-32B, an approach to coding agents that matches Devstral 2 at just $9,000. It is fully open-source and you can train your own model easily - at 26x the efficiency of using RL.

Paper: https://t.co/aeD6T2WW3O

Here’s how 🧵

[Aman @not_aman17]

@ethnlshn Congrats Ethan!

[Aman @not_aman17]

@sundayrobotics Cool design, gives me fall guys vibe.

[Aman @not_aman17]

@tyleraromero Thank you for saving my runs from the NCCL abyss.

[Aman @not_aman17]

Olmo 3 is out.

New fully open 7B & 32B. Check out the tech report for many interesting findings!
https://t.co/84hnTCKsOO

[Ai2 @allen_ai]

Announcing Olmo 3, a leading fully open LM suite built for reasoning, chat, & tool use, and an open model flow—not just the final weights, but the entire training journey.
Best fully open 32B reasoning model & best 32B base model. 🧵