Kyle Lo
@kylelostat
Followers
4K
Following
3K
Media
74
Statuses
775
language model pretraining @allen_ai, co-lead of data research for OLMo w/ @soldni, he/him, https://t.co/5Hm9cx3Urz🧋
Seattle, WA
Joined January 2019
lol so during neurips, we kept the RL run going and the model kept getting better 😂 Olmo 3.1 is a.. 🐡 32B Thinking, still best fully-open model to-date 🐠 32B Instruct, for ppl who hate long yapping, as good as qwen3 we added like 10 more pages to the paper too! thx for
Olmo 3.1 is here. We extended our strongest RL run and scaled our instruct recipe to 32B—releasing Olmo 3.1 Think 32B & Olmo 3.1 Instruct 32B, our most capable models yet. 🧵
1
25
89
new olmo 3.1 artifacts: https://t.co/Xs4t0qCf1y paper (arxiv soon): https://t.co/QKxvMbO9cj demo:
huggingface.co
0
0
4
huge shout out esp to the student researchers @allen_ai @uwnlp who really pushed this through, esp @saumyamalik44 @heinemandavidj who are PhD applicants this cycle 🤩 now to finish this huge paper & get it onto arxiv lolll 🥱
1
0
25
deadline for Olmo internship application is Dec 15 🗓️ chatted with prospectives at #NeurIPS2025, here's some fish & tips 🐟 approach it like your phd apps. reach out to potential mentors. are they even hiring? what are they excited about? 🐠 it's not about num papers published.
5
15
211
Neurips 2025 was such a blast! We snuck a grand piano into the CreativeAI Track to demo Aria, our pretrained chat-style music model:
23
47
398
pretraining has been living on a remote tropical island w a new identity
Overheard at a #NeurIPS2025 social: "rumors of pretraining's death have been greatly exaggerated"
1
3
36
Happening now at poster #5203 Hall C!
@heinemandavidj @allen_ai The main force behind FlexOlmo is @WeijiaShi2 of @uwnlp who drove this project w @sewon__min while interning with us at @allen_ai. The team (+@AkshitaB93 @notkevinfarhat) are expanding this idea to real-world, high-stakes data permissions scenarios https://t.co/kghVY9BVjI
0
5
63
happening now in Hall C #1903!
Signal and Noise is led by @heinemandavidj, talented researcher in our Olmo group at @allen_ai who is applying to PhD programs this cycle! Many of the ideas in his work shaped the design our Olmo 3 evaluation suite! https://t.co/9rMEManlV1
1
5
53
At NeurIPS 🌴! Check out our spotlight paper on the signal and the noise in language model benchmarking! (Wed 4:30 - 7:30 -- #1903) ++ come say hi for more details on all things evaluation for Olmo 3 + beyond. DMs are open!! 🐄
Evaluating language models is tricky, how do we know if our results are real, or due to random chance? We find an answer with two simple metrics: signal, a benchmark’s ability to separate models, and noise, a benchmark’s random variability between training steps 🧵
0
1
14
If you want to hear about why data curation is challenging, impactful, and interesting, come see my talk today at 5pm in the #NeurIPS2025 exhibitor hall!
1
5
24
lotta Ai2 folks in san diego #NeurIPS2025 find us for free tokens!
We're at #NeurIPS2025 with papers, posters, workshops, fireside chats, & talks across the conference. Come learn about our latest research + see live demos!
1
3
26
@heinemandavidj @allen_ai The main force behind FlexOlmo is @WeijiaShi2 of @uwnlp who drove this project w @sewon__min while interning with us at @allen_ai. The team (+@AkshitaB93 @notkevinfarhat) are expanding this idea to real-world, high-stakes data permissions scenarios https://t.co/kghVY9BVjI
Can data owners & LM developers collaborate to build a strong shared model while each retaining data control? Introducing FlexOlmo💪, a mixture-of-experts LM enabling: • Flexible training on your local data without sharing it • Flexible inference to opt in/out your data
0
2
18
Signal and Noise is led by @heinemandavidj, talented researcher in our Olmo group at @allen_ai who is applying to PhD programs this cycle! Many of the ideas in his work shaped the design our Olmo 3 evaluation suite! https://t.co/9rMEManlV1
Evaluating language models is tricky, how do we know if our results are real, or due to random chance? We find an answer with two simple metrics: signal, a benchmark’s ability to separate models, and noise, a benchmark’s random variability between training steps 🧵
1
2
16
I'll be at #NeurIPS2025 from Tues-Sun! Come say hi 👋 if you wanna chat about 🦈 olmo 3 stories, pretraining data & evals 🐟 model specialization 🐡 AI for education 🐠 open intern & full time positions at Ai2 We have two spotlight papers 🥐 Signal and Noise (Wed) shows why
7
11
70
Artificial Analysis' Openness Index is now live 🎉 Huge congrats to @allen_ai 👏 OLMo 3 leads the pack, demonstrating top-tier openness for model availability and transparency across data and methodology. 🔗 https://t.co/4fHVjTi5vD
9
27
179
Re the openreview mess, I sent this to my students and wanted to share:
0
53
359
can confirm touching sand is pretty fun
As a former conference result worrier, my advice is to submit rebuttal and go touch sand. Maybe get some tea. It's not really in your control anymore
1
0
16
an ICLR paper i reviewed responded w a 2500 word rebuttal that was very good 😬 feel real bad for them ☹️
1
3
162
yikes 😬 it can be difficult to separate mean reviews from critical but helpful reviews recommend folks avoid seeking this info. no way to consume it rationally & could harm ur ability to form meaningful professional relationships in future
3
2
40