poolside

The finishing touches. In the final post of our Model Factory series, we dive into post-training at scale—how we run supervised fine-tuning and reinforcement learning workloads, the infrastructure behind them, and the tools we’ve built to support it all, including our dataset viewer Podium and GPU<>GPU weight transfer. Check out the post: https://lnkd.in/eSCGtxgn With this, we’ve shared the full arc of how we train, serve, and evaluate foundation models at poolside—from raw data to post-training. We hope you’ve enjoyed the series, and if you’re interested in building toward AGI, drop us a note.

Inference and evals, built into the core. In part five of our Model Factory series, we introduce Atlas—our inference codebase—and show how we run it across diverse hardware, serve it at scale, and pair it with our evals platform to track model quality throughout training. Check out the post: https://lnkd.in/eje842ds And let us know what you think!

New on the blog: how we built our code execution environment. This post walks through the system behind code execution at poolside—from pulling external repos into our revision serving system, to using AI agents for image building, to serving execution requests at scale. We also share how we efficiently represent multiple repo revisions using a layered approach. Part four in our Model Factory series: https://lnkd.in/e2CQ4CEk As always, curious to get your thoughts in the comments.

New on the blog: a look inside Titan, the Model Factory’s furnace. Titan is our distributed training codebase, purpose-built for large-scale foundation model development. This post gives an overview of Titan, as well as how we use it to conduct architectural experiments and train foundation models at scale. Part three in our Model Factory series: https://lnkd.in/ePfyBFCC Curious to hear what others are building—drop your thoughts in the comments.

Foundation models live or die on data. In part two of our Model Factory series, we unpack how we handle it end to end: ingest, synthetic generation, filtering, deduplication, blending, and streaming. Check it out here: https://lnkd.in/eru-JpgX

Welcome to the team, Phil! Phil joins us as Chief Investment Officer at a pivotal moment as we accelerate toward AGI through software engineering. His focus: forging strategic partnerships and securing the resources needed to push our foundation models to the frontier. The path to human-level AI requires not just breakthrough research, but also the right strategic alliances and capital deployment. Phil brings exactly that expertise to help us scale our vision. https://lnkd.in/eiEG3AxU

We built our Model Factory to solve core problems in foundation model research, namely slow iteration and too much manual work. Our newest blog series breaks down how we approach end-to-end model development at poolside—from raw data to inference. The first post covers why we built the Model Factory and the principles that shaped it. Upcoming posts will dig into: - Data gathering and preprocessing - Pre-training, fine-tuning, code execution, and RLCEF - Inference and evaluation at scale If you’re curious about how we move fast—or want to help push the frontier—start here: https://lnkd.in/eqzniFwU

We’ve released our research vision. The key insight: Reinforcement Learning is the most important scaling axis to AGI. In a world full of noise about AI's future, we have a specific view on how we’ll get to AGI through software. We’re on the path. Read our full research vision here: https://lnkd.in/eWYqu8HH

“You can't fine-tune your way to AGI.” Our co-founder and CTO Eiso Kant recently joined Denis Skrinnikoff, CTO of IREN, for a fireside chat at RAISE, hosted by Gemma Allen. The conversation explored the critical importance of compute stack optimisation for achieving AGI. Thanks to RAISE Summit for having us!

“The AI workplace conversation is full of extremes—but the truth lies in the middle” Our CEO Jason Warner spoke to Bloomberg to cut through the noise around AI's impact on work and society. His key insights: 1️⃣On job transformation vs. job loss: We understand that the pace of change can be disorientating, but believe it will ultimately create more opportunities. Job functions will evolve, but that doesn't mean entire sectors disappear. 2️⃣Why software development is AI's sweet spot: It's both rote and creative—the perfect combination for AI enhancement. Software developers will always exist, but the day-to-day tasks they handle will increasingly be automated. 3️⃣The pace problem: The speed of AI improvement is genuinely disorienting. Take base code building: a core part of many developers' roles that AI will soon own entirely. Our relationship with specific tasks must evolve. The bottom line? We can’t predict the future—but we do believe it’s going to be rife with opportunity as a result of the exciting developments in AI. 📺 Watch Jason's full Bloomberg segment from 1:30 onwards: https://lnkd.in/eFEWQ_nr