a life update: i started a new company called Pragmatik (p7k) Labs (语用科技) in shanghai, focusing on the research of next-generation agents across digital and physical worlds. thanks to Gaorong Ventures and HSG (红杉中国 & 高榕创投) for co-leading this round, and to Tencent (腾讯) and Shanghai Engine Fund (上海未来产业基金) for the support. @pragmatik_labs ·

1/ Still looking for a minimalist, high-performance framework for agentic RL research? Meet Molt — an agentic-first, PyTorch-native reinforcement learning framework with roughly 9K lines of RL code for 700B models. ⭐

i recently discussed with some friends about the strike of model effects and costs. while frontier models are smashing a lot of things it is always expensive to use them. me personally always think about using as many tokens as possible with my subscriptions and coding plans, and i think if i turn the usages to tokens that might be a much bigger amount of money (i remember a media previously said if u use codex crazily u can use over 8000 dollars a month with 200 dollar subscription. not sure about the number). how about enterprises that really use tokens and count expense on the unit of token? some interviewees told me that they tried their design their system by tasks, like easier or to say more formatted tasks go to seed, qwen, dpsk, harder tasks like agentic tasks go to fable or sol. glm 5.2 is nice for coding but still after claude and gpt, recent grok is a bit similar. this just reminds me that if we design the system well on the user side, we can have more productivity based on sufficient usage of ai today without too much worry about expense. maybe there is still much room for this.

i still remember the discussion of the case of us visa application and i thought that must be mission impossible for ai and just said u guys must be crazy... now it seems that i am the dumb hahah! anyway, time for frontier models to fight again! bon courage!

ai making ai better recursively is an impressive idea but sometimes depressing as human researchers seem to become less and less significant in this process. i do believe that there is great potential in this direction while human supervision will become more principled and higher-level. then imagination and vision matters more than ever.

Today, we’re releasing Continual Learning Bench 1.0: the first, realistic benchmark for measuring how AI systems can improve in online settings. Benchmarks today assume models are stateless. Each example is independent, and once a system finishes a task, it moves on as if nothing happened. But deployed AI systems should learn from experience. We tested 10+ frontier systems against novel, expert-validated tasks and find there’s still plenty of headroom for learning. (1/n)

i do like this passage, and here are some thoughts: 1. critical thinking is essential in the era of agents. i still remember that many years ago when i studied the lesson of critical thinking, i learned that keeping debating with yourself by listing out reasons can really deepen your thinking. today, critical thinking becomes humans debating with agents, so that they can think more deeply together and analyze problems in a more comprehensive way. 2. designing a healthy and well-structured organization and system is essential for creation and building. with systematic support and efficient tooling, humans can work exponentially more effectively together with agents. that gives people more time to take care of their physical and mental health, while also exploring new opportunities. 3. new era often favors newbies, because they have less past experience and therefore less fear of current difficulties. what oldbies should really think about is which parts of their experience are actually worth leveraging. from my perspective, we should think more carefully about which experiences are truly aligned with first principles. but anyway, ai first is super, super exciting!

we need agent evals that are really consistent with real world usages. otherwise people are optimizing foundation models for the wrong direction. the problem of targeting is even bigger than benchmaxxing.

🧵 1/ Our agent Terminator-1 scored ~100% on 8 major AI agent benchmarks, e.g., SWE-bench Verified & Pro, Terminal-Bench, beating Claude Mythos. It solved 0 tasks. Benchmarks are the field's shared language for measuring AI progress. Our new work shows that language is broken. Here’s how.