←
Eric Xing681 posts
Eric Xing avatar

Eric Xing

@ericxing

Researcher, educator, entrepreneur, and administrator in computer science, artificial intelligence, and healthcare.

◷ Joined May 2010▧ 46 Photos and videos
Eric Xing avatar
Eric Xing@ericxing·8h

Full transparency and reproducibility promised and delivered, not something you can expect from corporate LLM producers. We stand by LLM360 definition of full openness for people to truly study, understand, and use without fear AI.

Looking for a reference manual for building a frontier LLM? We wrote one 👇 The K2 Horizon technical report is now on arXiv: 134 pages on how we built six models, from architecture to inference. It may be the most detailed LLM training reference you can read today. • Architecture: we bring two new attention designs, MoVA and softplus attention gating, to strengthen reasoning. With them, our 36B model nearly matches a 32B dense one on one-seventh of the active parameters. • Pre-training: we trained four models, from 3.7B to 36B, on the same 22T tokens. Their normalized loss curves collapse onto nearly the same line, so a recipe that works on a small model carries over when scaling up. • Data: 60+ pages on the part most reports skip. How we generated ~10T synthetic tokens, and why we chose the data mix on predicted scores at scale, which reversed the small-run ranking. • Post-training: we document our full production-scale run for the 375B flagship. • Inference: Uno gives an autoregressive model diffusion speed with one small adapter. Up to 2.2× faster decoding, lossless by construction, no draft model. This is the last piece of the openness puzzle. K2 Horizon already shipped with 380 checkpoints, 13 TB of pre-training data, 58 training runs and 17 code repositories. The report explains the decisions behind them. More details: arxiv.org/abs/2611.21003

0:44
Hector Liu avatar
⟳ Eric Xing retweeted
Hector Liu@waterluffy·17h

Probably (or hopefully) the best reference resource for how to train an LLM.

Looking for a reference manual for building a frontier LLM? We wrote one 👇 The K2 Horizon technical report is now on arXiv: 134 pages on how we built six models, from architecture to inference. It may be the most detailed LLM training reference you can read today. • Architecture: we bring two new attention designs, MoVA and softplus attention gating, to strengthen reasoning. With them, our 36B model nearly matches a 32B dense one on one-seventh of the active parameters. • Pre-training: we trained four models, from 3.7B to 36B, on the same 22T tokens. Their normalized loss curves collapse onto nearly the same line, so a recipe that works on a small model carries over when scaling up. • Data: 60+ pages on the part most reports skip. How we generated ~10T synthetic tokens, and why we chose the data mix on predicted scores at scale, which reversed the small-run ranking. • Post-training: we document our full production-scale run for the 375B flagship. • Inference: Uno gives an autoregressive model diffusion speed with one small adapter. Up to 2.2× faster decoding, lossless by construction, no draft model. This is the last piece of the openness puzzle. K2 Horizon already shipped with 380 checkpoints, 13 TB of pre-training data, 58 training runs and 17 code repositories. The report explains the decisions behind them. More details: arxiv.org/abs/2611.21003

0:44
Xuezhe Ma (Max) avatar
⟳ Eric Xing retweeted
Xuezhe Ma (Max)@MaxMa1987·18h

Today, we release our tech report for K2 Horizon!

Looking for a reference manual for building a frontier LLM? We wrote one 👇 The K2 Horizon technical report is now on arXiv: 134 pages on how we built six models, from architecture to inference. It may be the most detailed LLM training reference you can read today. • Architecture: we bring two new attention designs, MoVA and softplus attention gating, to strengthen reasoning. With them, our 36B model nearly matches a 32B dense one on one-seventh of the active parameters. • Pre-training: we trained four models, from 3.7B to 36B, on the same 22T tokens. Their normalized loss curves collapse onto nearly the same line, so a recipe that works on a small model carries over when scaling up. • Data: 60+ pages on the part most reports skip. How we generated ~10T synthetic tokens, and why we chose the data mix on predicted scores at scale, which reversed the small-run ranking. • Post-training: we document our full production-scale run for the 375B flagship. • Inference: Uno gives an autoregressive model diffusion speed with one small adapter. Up to 2.2× faster decoding, lossless by construction, no draft model. This is the last piece of the openness puzzle. K2 Horizon already shipped with 380 checkpoints, 13 TB of pre-training data, 58 training runs and 17 code repositories. The report explains the decisions behind them. More details: arxiv.org/abs/2611.21003

0:44
Vipin Gautam (Viipin I Gautam) avatar
⟳ Eric Xing retweeted
Vipin Gautam (Viipin I Gautam)@viipin8·Nov 10

Longship, whirlpool, camera sweeping around the ship. I don’t know what I expected when I tried this on my early access to PAN 2, but it wasn’t this, and the sea keeps churning as the camera moves

0:16

Introducing PAN 2: an open foundation world model for physical intelligence. A language model predicts the next word. PAN 2 predicts the next world: what happens after each action, grounded in time, physics and space. Towards AI that simulates “what if?” before it acts. More details: ifm.ai/pan/

1:10
Markandey Sharma avatar
⟳ Eric Xing retweeted
Markandey Sharma@TechByMarkandey·Nov 10

The world doesn’t pause because the camera’s busy. I tried PAN 2 early on a hiker on a sea cliff at sunset: the camera circles, and the wind and waves just keep going.

0:16

Introducing PAN 2: an open foundation world model for physical intelligence. A language model predicts the next word. PAN 2 predicts the next world: what happens after each action, grounded in time, physics and space. Towards AI that simulates “what if?” before it acts. More details: ifm.ai/pan/

1:10
Utkarsh Sharma avatar
⟳ Eric Xing retweeted
Utkarsh Sharma@techxutkarsh·Nov 10

Counterfactuals are where world models earn their keep. “Show me a tower falling” is easy. “Would this tower fall if I pulled this block” is the one I care about, and it’s what IFM’s Jenga demo asks.

Introducing PAN 2: an open foundation world model for physical intelligence. A language model predicts the next word. PAN 2 predicts the next world: what happens after each action, grounded in time, physics and space. Towards AI that simulates “what if?” before it acts. More details: ifm.ai/pan/

1:10
Sumit Doriya avatar
⟳ Eric Xing retweeted
Sumit Doriya@sumitdoriya21·Nov 11

PAN 2 is taking AI beyond language by helping it understand and predict the physical world. Instead of predicting the next word, it predicts what happens next through physics, time, and space. A step closer to AI that can imagine possibilities before taking action.

0:16

Introducing PAN 2: an open foundation world model for physical intelligence. A language model predicts the next word. PAN 2 predicts the next world: what happens after each action, grounded in time, physics and space. Towards AI that simulates “what if?” before it acts. More details: ifm.ai/pan/

1:10
Nelly; avatar
⟳ Eric Xing retweeted
Nelly;@nrqa__·Nov 10

A moving camera shouldn’t mean a moving world Obvious to a person, surprisingly hard for a model Here’s what I got from PAN 2 on early access: a flight around a peak as first light hits it, and the mountain keeps its shape

0:20

Introducing PAN 2: an open foundation world model for physical intelligence. A language model predicts the next word. PAN 2 predicts the next world: what happens after each action, grounded in time, physics and space. Towards AI that simulates “what if?” before it acts. More details: ifm.ai/pan/

1:10
Matthew Callahan avatar
⟳ Eric Xing retweeted
Matthew Callahan@MattVanta·Nov 11

External dynamism is the dimension nobody posts about. In IFM’s koi demo a fish swims in from off-screen and the pond ripples. The world outside the frame still exists and still acts on what’s inside it.

0:10

Introducing PAN 2: an open foundation world model for physical intelligence. A language model predicts the next word. PAN 2 predicts the next world: what happens after each action, grounded in time, physics and space. Towards AI that simulates “what if?” before it acts. More details: ifm.ai/pan/

1:10
Institute of Foundation Models avatar
⟳ Eric Xing retweeted
Institute of Foundation Models@IFM_AI·Nov 11

Huge thanks to @vllm_project for day-0 support of PAN 2. ​ ​ Serve it on your own machine today: huggingface.co/IFM/PAN2

🎉 Great work from the PAN 2 team on an open-weight world model built for real-world simulation, grounded in time, physics and space to predict how scenes evolve as actions unfold. vLLM-Omni brings Day-0 support for PAN 2 text-to-video and image-to-video generation, making it available through an open inference stack.

Mingkai Deng avatar
⟳ Eric Xing retweeted
Mingkai Deng@mdeng34·Nov 10

I do a little surfing in my free time. The hardest call is *when* to pop up — too early and the wave rolls over you, too late and it’s already gone So I asked PAN 2, the world model we trained at @IFM_AI, to show me the same wave, three different timings. It shows what wipes me out, and what leads to a clean ride Simulate first, wipe out less #PAN2 #IFM #WorldModel #Simulation

1:18
Eric Xing avatar
Eric Xing@ericxing·Nov 10

Excited to share our latest edition of world model PAN 2 built on a Generative Latent Prediction (#GLP) architecture. A world model needs to operate in an "Aristotelian world" instead of a "Platonic world" where physical intelligence translates to 6 dimensions of capability: obeying the laws of physics, coherence over history, persistence beyond sight, conformity to constraints, invariance to viewpoints, adherence to objective dynamism; and furthermore provides 3 advanced functions: acting as a world simulator that is steerable with actions and plans (long sequence of sections), thereby enabling stimulative reasoning or thought-experiments, and therefore supporting agent to solve complex tasks, and doing so in a fully auditable fashion (hence Aristotelian). A full paper will be posted soon to explain in details how we achieve this. Congratulation to the @IFM_AI PAN team.

Introducing PAN 2: an open foundation world model for physical intelligence. A language model predicts the next word. PAN 2 predicts the next world: what happens after each action, grounded in time, physics and space. Towards AI that simulates “what if?” before it acts. More details: ifm.ai/pan/

1:10
Nicholas Ho avatar
⟳ Eric Xing retweeted
Nicholas Ho@eigenNick·Nov 2

Contextualized gene representations work surprisingly well for predicting multigene perturbation effects. I’m excited to introduce EMBER, a pretrained gene-centric foundation model designed specifically to learn context-specific gene representations. A joint work with @jmuiuc and @ericxing. With innovations in single-cell pretraining and dataset curation, EMBER gene representations can be combined with a lightweight probe for single gene knockouts, genetic interactions (double knockouts), and even drug perturbations. Both a 🧵 and a 🎥, got in front of a camera to share some highlights! (1/n)

4:12