yay

why is RL so hard...

PHP2Shell technical breakdown to show you how an attack could actually run code remotely on PHP servers by single HTTP POST request crafted to cause OOM error to exploit the destructor callback. AI adoption is becoming ever more important as it's ability to find security vulnerabilities is currently greater than our ability to put the patch push through. In this case, PHP2Shell was known for 10 weeks until multiple leaks forced them to frantically patch the update earlier than scheduled. What a nightmare.

so this whole PHP shenanigans is really more human error in not being able to keep the sec vulnerability under wraps than about AI mayhem causing havoc. this could have gone very differently if we held it together.

does OpenAI get credit by default for every vulnerabilities that Patch the Planet?

Quick update on Fred (that's his name now), he seems to have reverted to The Thinker pose after 9 hours of training with SAC.. i may revert back to PPO xitter.cf/calebfoundry/status/21…

DeepSeeek just open sourced DeepInfer that only supports NVIDIA Blackwell (data center GPUs not consumer chips like 5090) and Ascend 950 NPUs for now.. This is targeted directly as data center inference efficiency and they report that at the cheapest API cost for running DeepSeek V4.1 Flash, the hourly revenue/GPU was $38 capable of making 155 trillion tokens per day at a cluster level. The downstream effect for efficiency at the infra level is lower cost of intelligence for consumer further driving down cost given higher inference margins. github.com/deepseek-ai/DeepI… This week has been an incredible run for open source AI.

UPDATE: my constipated agent is doing better. xitter.cf/calebfoundry/status/21…

Cooking some RL on sfcompute tonight

RL training to mimic indigestion after Taco Bell

I can't tell if I'm in my own bubble because most people around me seem unaffected by AI.. but I'm literally studying more than I ever have in my life. It's hard to tell if the right approach to AI is to go touch more grass or lock in harder. Curious to hear ppls thoughts.

RL scaling for MiMo-V2.6 explained. Great work @XiaomiMiMo and thank you for sharing your knowledge and experience openly to the world.

This is insane

At this point, should we just assume that if Chinese labs DON'T mention which GPU they used to train their models, they most likely used NVIDIA chips?

quick preview of my next video.. (hopefully tomorrow)

holy crap, RL is a huuuuge rabbit hole..

Today I learned that higher batch size doesn't always mean higher training efficiency. There's a thing called critical batch size which is the point where batch size increase has diminishing return because gradients contain enough information for the optimizer and adding more data (which means higher compute tradeoff) starts to become less worth it.

I like how DeepSeek just continues to build their models on top of each other like a LEGO piece. Still really impressed by what DeepSeek achieved with the V4.1-Flash model.. can't wait to see what V4.1-Pro might look like

I had early access to Gemini 4 Argon and the first thing I tried was 1M output tokens. Here is me translating the entire Pride and Prejudice novel in one shot (~200k output tokens) from English to Korean. For reference War and Peace and The Bible has about (~800k to 1M tokens) so Argon could technically output the novel all at once. Yes, the use case for 1M output token is rare especially since harness nowadays handles context management better, and honestly, this multi-agents can divide and conquer this faster. But maintaining coherence for context that long is pretty darn impressive if you ask me.