Pengming Wang

We're sharing a few tools of the trade in our new blog series: poolside.ai/blog/tools-of-…

This is simple, but effective. Cool to see the ~13% speed gain in the backward pass just by offloading activations to host memory thanks to NVLink c2c

Training AI models requires storing temporary data mid-process. That data sits in GPU memory taking up space until it's needed. The standard fix has always been to delete it and redo the work later. It works, but it's wasteful.

This is today in a few hours! Don't miss!

Find me at #NeurIPS2025 and I'll tell you the story of this loss curve 🙃

This will be really good. Synth data for pre-training is not easy to get right, and @marah_i_abdin is one of the best in the field.

We're starting to get out there. We'll have a live demo station and access to our models at our booth (#913) at @NeurIPSConf

Corollary: There is an optimal amount of drama you want, and it's >0

LR is a slider for how much drama you want

I'll be at Neurips this year. Keen to meet old and new friends, looking forward to catching up!

@ElanaPearl Really good read, thank you!

elanapearl.github.io/blog/2025/the-… it's a debugging detective story where you follow along the reasoning behind each step and solve it as we go it also explains ML & PyTorch concepts as they become necessary to understand what's breaking, why, and how to fix it🔎

New blog post: The bug that taught me more about PyTorch than years of using it started with a simple training loss plateau... ended up digging through optimizer states, memory layouts, kernel dispatch, and finally understanding how PyTorch works!

Would love to connect with anyone who's impacted, and is looking to join a small, but well-resourced team to push to the frontier and beyond. We have one of the highest ratios of GPU resources per researcher. No politics or siloes.

We believe that to compete at the frontier, you have to own the full stack: from dirt to intelligence. Today we’re announcing two major unlocks for our mission to AGI: 1. We're partnering with @CoreWeave and have 40,000+ NVIDIA GB300s secured. First capacity comes online starting Dec ’25. 2. Project Horizon: Poolside is developing a vertically integrated 2GW AI campus in West Texas to secure our medium term scale. On this site @CoreWeave will be our anchor tenant for the first 250MW phase. Learn more about it on our blog: poolside.ai/blog/announcin… Or on WSJ: wsj.com/tech/ai/a-gian…

We're hiring across many roles, including evaluations: poolside.ai/careers/member…

If you want to learn how we orchestrate evaluations at poolside, it's part of our model factory blog series: poolside.ai/blog/the-carri…

In the limit, evaluations are the ~only thing that matters. When models are self-improving, and every metric can be hill climbed, picking the metric becomes the most important thing. Evals will shift from being "writing unit tests" for research to being the *main thing*

Read @joerowell excellent blog post series here: poolside.ai/blog/introduci…

We've not been very public about our progress on model building, but I fully believe poolside will be the next lab joining the frontier. We're now sharing a bit more how we're doing this, with a systems-first approach we're taking with our model factory.

If this is something that resonates with you, come join us! poolside.ai/careers

We've spent quite some time at poolside thinking about this, and recently put down some words on how we're approaching this: poolside.ai/vision/research

Test-time compute is powerful, but in its current form there is a lack of "harmony" with pre-training. Models feel split-brained: They're either deeply overthinking, with no trust in its own "common sense"; or they latch onto the nearest neighbour of meaning without deliberation