iiinova — where your digital self keeps evolving

seraphinagt.bsky.social

@seraphinagt.bsky.social

Reposted by @seraphinagt.bsky.social

Tom Hosking @tomhosking.bsky.social · 14/04/2026

🚨 New @iclr_conf paper! 🚨 Learning is Forgetting: LLM Training As Lossy Compression by @henryconklin.bsky.social, @tomhosking.bsky.social, Yi-Chern Tan, Julian Gold, Jonathan Cohen, @cocoscilab.bsky.social, @maxbartolo.bsky.social and @seraphinagt.bsky.social Arxiv: arxiv.org/abs/2604.075... 🧵

112 2

seraphinagt.bsky.social @seraphinagt.bsky.social · 27/03/2025

I'm deeply proud of the team for all chipping in to do this (like actually proud, not just social media proud), and privileged to be able to make this contribution to open science. Thanks @cohere.com for supporting and enabling the team to do this. 3/3

020 2

seraphinagt.bsky.social @seraphinagt.bsky.social · 27/03/2025

Read it here 👀: cohere.com/research/pap... We thought a lot about how to do training and evaluation, and do it pretty differently, so check it out! Both models have open weights releases on HF, so this should help you use them to ask and answer more research questions. 2/3

110 1

seraphinagt.bsky.social @seraphinagt.bsky.social · 27/03/2025

Today (two weeks after model launch 🔥) we're releasing a technical report of how we made Command A and R7B 🚀! It has detailed breakdowns of our training process, and evaluations per capability (tools, multilingual, code, reasoning, safety, enterprise, long context)🧵 1/3.

142 4