Everything needs to be overlapped. Everything needs to be parallelized. Remove transport overhead where possible. TOTAL FACTORIO MINDSET. Everything that can run SHOULD RUN.
A finance friend brought this up at lunch the other day and I couldn’t help it… it’s such a simplistic take it’s ridiculous. Growth-adjusted multiples are Finance 101. A multiple without a growth rate next to it is half a sentence.
Walmart is growing at ~5%/yr. OpenAI ~3x. Anthropic ~10x. Now put the multiple over the growth rate and it inverts: Anthropic screens around 0.02 on a revenue-PEG vs Walmart’s ~0.26, roughly an order of magnitude lower.
Want to make the bear case? Argue margins or FCF. A naked EV/Rev multiple isn’t it.
Very happy to see Laguna-M.1 on this discussion!
Worked on both sides of inference and training on this one, and we'll keep pushing the RL infra to fly higher every time 🚀
New blog! Is frontier asynchronous RL solved?
The blog covers Async RL theory and infrastructure, surveying 8 open-weight frontier labs for the algorithmic techniques and systems fixes to handle train-inference mismatch. Also answered: why do current methods still fail at high policy lag? Which methods scale with horizon and compute?
New blog! Is frontier asynchronous RL solved?
The blog covers Async RL theory and infrastructure, surveying 8 open-weight frontier labs for the algorithmic techniques and systems fixes to handle train-inference mismatch. Also answered: why do current methods still fail at high policy lag? Which methods scale with horizon and compute?
The @poolsideai Hackathon is over but I'm continuing to teach Laguna XS.2 to be a better scientist 🧑🔬. Our Protein-Ligand Design Gym now has 1000+ examples with harder protein/ligand interaction problems.
Glad to see more teams raising awareness on SWE-style benchmarks reliability ! We discussed this same problem on our blog post earlier last month as part of our reward hacking investigations 🔍🕵️
https://t.co/5p6iiiM8YW
Today we’re releasing DeepSWE, a new standard for agentic coding benchmarks.
On public leaderboards, top models often look relatively close in capability. DeepSWE shows where they actually diverge, reflecting the realistic experience of developers in their day-to-day work.
@AlessandroDuico@Badiaserra@poolsideai@eloquake Design is not finished, the job is never fully done. The important bit is to learn, improve, and create usable outputs along the way so others can join the party 💪 🥳
In the era of #ArtificialIntelligence, when human dignity is threatened by new forms of dehumanization, ours is the pressing duty to remain profoundly human. We must lovingly safeguard the grandeur of humanity bestowed upon us and revealed in its fullness in Christ, the splendor of which no machine can ever replace. #MagnificaHumanitas
https://t.co/6i9MWs6LJl
The level of the @poolsideai hackathon in London was higher than the average in SF.
We tried distilling Laguna XS.2 into a dense model. A ~11x parameter reduction from 33B to 3B.
https://t.co/9bwmzRr2kc
Many thanks to the organizers @poolsideai@eloquake@Badiaserra
🚀Great to see @RedHat_AI and @poolsideai team up to make Laguna XS.2 faster and cheaper to serve in vLLM. A DFlash speculator built with Speculators drafts 8 tokens per forward pass for 2-3x faster decoding at no quality loss, and LLM Compressor adds FP8 / NVFP4 / INT4 checkpoints so you can match your hardware budget.
🔗 https://t.co/3mguMKtwTt