Ai2

Ai2 Climate Modeling’s director Chris Bretherton wrote a great blog post summarizing our group’s recent work on AI-based climate model emulators. Reading the whole post is worthwhile (https://lnkd.in/gPPV-79Z) but I will also summarize here. The post describes two preprints recently on arxiv: 📃 ACE2: Accurately learning subseasonal to decadal atmospheric variability and forced responses (https://lnkd.in/gWSxdwHH) • Introduces the ACE2-ERA5 model—the Ai2 Climate Emulator version 2 trained on the ERA5 reanalysis dataset • This is the first fully AI-based autoregressive atmospheric model to show accurate historical (past 80-year) trends in atmospheric temperature • At the same time as getting long-term trends right, ACE2-ERA5 also generates realistic features like tropical cyclones, stratospheric sudden warmings and the Madden Julian Oscillation—all essential atmospheric phenomena for extreme weather and sub-seasonal to seasonal prediction 📃 ACE2-SOM: Coupling an ML atmospheric emulator to a slab ocean and learning the sensitivity of climate to changed CO2 (https://lnkd.in/gBsi7CKj) • Here we use a simple thermodynamic (“slab”) ocean model to do climate change simulations where carbon dioxide concentrations are increased (up to quadrupled) from present-day conditions • Trained on a physics-based model (SHiELD-SOM) forced with a range of CO2 concentrations, we show that ACE2-SOM is able to accurately capture the expected response of surface temperature to increasing CO2 (see image) • ACE2-SOM also accurately predicts the changes in frequency of extreme precipitation up to the top ten-thousandth percentile There is still work to be done: for example, while the equilibrium climate predictions are accurate, ACE2-SOM does not correctly predict the initial response to a large abrupt increase in CO2. It’s exciting to push AI+weather to climate time scales. The speed of these models really shines—being able to do 1000-year simulations (over a million 6-hour steps!) in less than a day still blows me away. Excited to see what 2025 brings for AI and climate modeling. These models are publicly released and simulations are fast and easy to run, even on your MacBook. See our ACE codebase: https://lnkd.in/g88qk2fp where you can find documentation and links to the models described above. Tagging the team: Spencer Clark, Brian Henn, Anna Kwa, Jeremy McGibbon, W. Andre Perkins, Elynn Wu, James Duncan, Lucas Harris (NOAA GFDL) and Chris Bretherton.

✍️Reminder: we've got open positions on our OLMo team for predoctoral candidates! Apply to be a Predoctoral Young Investigator by January 15.

Everyone wants open-source language models but no one wants to lift these heavy ass weights. We just released our paper "2 OLMo 2 Furious" with extensive language modeling details across the entire stack for OLMo 2. Can't stop us in 2025. Paper: https://lnkd.in/gEkqViFv Blog: https://lnkd.in/gGv6ykhm Demo: https://lnkd.in/g7p5mwfK Collection: https://lnkd.in/ghpWJju3 

Our senior director Hannaneh Hajishirzi has been trying to solve puzzles since she was a child. Don't take our word for it — Hanna's mom is one of the featured interviews in this awesome profile by GeekWire!

Meet OLMo 2, the best fully open language model to date, including a family of 7B and 13B models trained up to 5T tokens. OLMo 2 outperforms other fully open models and competes with open-weight models like Llama 3.1 8B — As always, we released our data, code, recipes and more 🎁 Read about the OLMo 2 recipe in our blog: https://lnkd.in/gPMeBczs

Meet Tülu 3 -- a set of state-of-the-art instruct models with fully open data, evaluation, code, and training algorithms, designed to serve as a comprehensive guide for modern post-training techniques. We invented new methods for fine-tuning language models with RL and built upon best practices in the community to scale synthetic instruction and preference data. We’re excited to bring the Tulu 3 recipe to more models shortly 🥳 Try it out: https://lnkd.in/gk3Q49a5

Much of the #AI spotlight is on Meta, Google, Amazon, OpenAI, Anthropic, and others in "Big AI." But there's another one looking to make waves and has been studying artificial intelligence for over a decade: Ai2. Created by the late Microsoft co-founder Paul Allen, AI2 wants to develop AI for the common good. It's arguably one of the most prolific AI research institutions, having published over 1,000 papers and has won many awards, minus a Turing Award, for its work. Ali Farhadi, a University of Washington professor and a long-time #AI researcher, is now leading that charge. An early employee, he helped build the firm's computer vision team, eventually developing technology that became a startup that Apple acquired. After a few years away, Farhadi returned for an opportunity he couldn't pass up: Taking control of AI2. Under his stewardship, the company's focus has narrowed, targeting areas where its work will have the most significant impact. One of them is promoting openness in AI. "AI is born and raised in the open," Farhadi tells me in an interview conducted at AI2's Seattle, Washington office. Today's issue of The AI Economy examines the work AI2 is doing, the importance of openness in AI, why Farhadi claims we're in an "evaluation crisis," the future of the technology, and what he thinks about artificial general intelligence (AGI) Plus, remember this week's curated news roundup! ---- 🔔 Enjoy what you read? Subscribe now to The AI Economy for more. https://lnkd.in/g6QERP6u 🚨 Follow Ken Yeung for more #tech and #business news. ♻️ Like and share! #artificialintelligence #enterprise #business #innovation #LLMs #technology #opensource

Open access to data. Open models. Open code. Open standards. Our approach to openness is all-encompassing, so we can build a future where AI is accessible to all. Learn more: https://lnkd.in/guNydGzQ

AI has limitless potential, which must be developed safely and responsibly. Our research team is dedicated to upholding our 5️⃣ principles to safeguard AI development and explore its potential.

We're hiring! Come help us redefine how researchers and engineers use state-of-the-art GPU clusters. #Hiring #SoftwareEngineer