Update after 2 months, I think Anthropic succesfully killed OpenClaw Back then I said: "The real power user for OpenClaw has become my gf who I put it in a group chat with me, her and OpenClaw, she's mostly stopped using ChatGPT etc and now only uses AI over Telegram with OpenClaw, it's much more user friendly for her in Telegram iOS than ChatGPT's own app" Since I switched it to OpenAI API my gf has since slowly stopped using it and switched back to the Claude app and now uses Projects she says So from my n=1, Anthropic successfully killed OpenClaw

Yesterday at the speedrun café in SF was a movie. Today we open doors at 9:00am PT. It's not too late to register: partiful.com/e/T0y2y4MGMBVya…

The deadline to apply for Alliance’s ALL17 cohort is tonight at 11:59 PM PT. If you’re building a startup, read on to learn why the best builders choose Alliance ↓

Meta's infrastructure. India's best builders. 48 hours. This is India's biggest agentic RL hackathon. OpenEnv is the new open standard to train AI agents, used by PyTorch, AI at Meta, and Hugging Face In April, India's best builders get to build on it and have their work reviewed by Meta and HF engineering teams. Meta PyTorch OpenEnv Hackathon × Scaler School of Technology. The best environments get evaluated for inclusion in the OpenEnv global ecosystem. - Real contribution. Not a portfolio piece. - $30,000 prize pool. - 48 hours. - Bangalore.

Reasoning + Reinforcement Learning may plateau less than expected which means more gains for longer for AI

Hot take from investing in robotics companies: simulation is overrated. Billions poured into sim-to-real. The gap hasn't meaningfully closed in 2-3 years. Locomotion tolerates 10% failure rates — manipulation doesn't. The industry is quietly converging on: bulk imitation data → fine-tuning with RL → less simulation, not more.

Introducing Lab. More companies should be able to own their model <> product optimization loop. If every company had access to frontier model capabilities and the post-training infrastructure that is currently locked behind the walls of the big labs, we would see far more progress. We would see many more ChatGPT or Claude Code level product experience breakthroughs across every area of economically valuable knowledge work. This is what we've built Lab for. It lowers the barrier to running the full post-training lifecycle, from large scale agentic RL to inference and evaluation, without having to manage the underlying infrastructure. Starting today, you can use the same infrastructure we use to train our own open frontier models, in a simple and accessible way. Hosted Training lets you run RL experiments in your own environments across many small and large open-source models. We are launching with support for agentic RL with multi tenant LoRA, built on top of our prime-rl training library. Support for SFT and dedicated deployments for full fine tuning are coming soon. Over time, Lab will serve as our platform to productionize all our research initiatives and make them accessible to anyone. If you are an AI Company, and you seek to have an AI Lab, you now have one.

advice to rl env / data startup founders, from anonymous rl env startup founders: > 1) There's a traditional trilemma of quality, speed, cost. Ignore costs, labs do not care (within reason). You must deliver quality and speed to be in the conversation at all costs. > 2) Focus on depth with one modality rather than contaminating focus amongst many. You are trying to land an extremely narrow wedge and then you can broaden. Do not be an environments company. Be a rust-tool-use environments company to start and broaden over time. This allows you to demonstrate your value to research early and grow wallet over time. Trying to boil the ocean will come at the cost of landing your first few deals. > 3) Relationships with researchers >>> relationships with anyone else in the orgs. Ultimately, they are the decision makers. You need to figure out what problems they are running into and how to solve them early. Get out of "partnerships hell" as fast as possible and build the relationship with a research lead. another founder emphasized: > Labs care about long horizon and data quality at scale. Data types are fast changing as well. and a third founder noted: > - labs have a short attention span, so you get one shot to look meaningfully different > - labs buy on trust, and there is a lot of noise. you need a credible intro to break in > - and finally, it is critical that your env cannot be gamed. if verifiers are not robust, the model (esp frontier ones) will take shortcuts and a lab researcher (focused on coding) added: > Build messy, real world environments that require lots of moving pieces. With agents getting better, I can see them being useful for creating simple environments from GitHub PRs, but you still need human taste to design complex environments. Build proprietary environments that are not publicly available. rl env/data startup founders, what else do you want to know?

this is a great question, I think it's a bit of both (you need the best expert to come up with the (very) complex tasks initially (and the rubrics), but also design the environment for them (which many aren't built yet), which is more of an engineering/infra/workflow problem...

We open sourced our tiny agent-sdk🧙the easiest way to build Claude Code like agents (with any provider) All the value is in the RL'd model, not your 10,000 lines of abstractions🦓 pip install bu-agent-sdk or 🔗↓ or put it to Claude Code and 1 shot your own in any language.

how come manus has better tool calling than anyone else? is this even true? is this an actual moat or will it be closed very soon? can someone shed light on where & how current sota llms like gpt or claude fail & manus succeeds? although i suspect anyone with this knowledge may not freely share it here. please dm me if you don’t feel comfy. i am very curious to learn.

"how can flash beat pro??" -> the answer is RL! flash is not just a distilled pro. we've had lots of exciting research progress on agentic RL which made its way into flash but was too late for pro. can't wait to finally bring them to pro👀

NeurIPS 25 was less of a conference and more holy pilgrimage. The sacred scriptures didn't dissapoint, ranging from automating environment design to search over scaffolds, long-context architectures and making continual learning/multi-turn RL robust. Read on for our takeaways! 1. Manual RL environment design will soon become a bottleneck — it’s costly, strictly human-interpretable, and limits the range of behaviors models can learn. New work like “Self Challenging Language Model Agents” by @YifeiZhou02 showed how generating environments with code verification rewards and example test cases could massively scale and robustify post-training. 2. Multi-agent scaffolds are everywhere, but we don’t yet understand how to evolve them safely or why they commonly fail. @jrosseruk and @j_foerst's AgentBreeder discovers new frameworks that maximize capabilities while minimizing jailbreaking/other undesired behavior. In “Why Multi-Agent Systems Fail”, @mertcemri, @melissapan and Shuyi Yang created a taxonomy of common multi-agent failure modes: specification issues, inter-agent misalignment and quality control failures, which is incredibly useful for moving us towards a “science” behind agent design. 3. The transformer reveals both her limitations and her complexity as we scale context lengths to millions of tokens. Time for a change! @jaseweston explained how "Multi-Token Attention" by @OlgaNLP improves long-context reasoning by comparing phrases rather than single tokens, avoiding the information bottleneck of compressing many cues into one query vector. Meanwhile, @harveyiyun's AbsenceBench fascinatingly exposes a surprising failure mode: transformers struggle to detect what’s missing from text, even when the answer is present elsewhere in context. 4. Continual learning is emerging as the next big thing, and no I didn't just watch Ilya's podcast... @AdamZweiger and @jyo_pari Self-Adapting Language Models use reinforcement learning to meta-learn how to perform test-time training: models generate their own synthetic data, configs, and update strategies, then learn which adaptation approach works best. Meanwhile, @benjamintherien in "Continual Pretraining of MoEs" showed that with the right data replay and learning-rate schedules, MOEs can update to new domains while retaining old capabilities. 5. Finally, reasoning models are powerful answer machines but still lack real interaction: they don’t ask clarifying questions or repair mistakes mid-conversation. Enter multi-turn RL. @natashajaques walked through @marwaabdulhai's “Simulating Human Personas” which reduces persona inconsistency by >55%, producing more stable synthetic partners for training long dialogues. Meanwhile, @JunhongShen1 and @jackbai_jkb's “Thinking vs Doing” shows that allowing agents to increase actions like page visits, clicks, and retries leads to agents that explore, correct themselves, and re-plan mid-episode.

It’s clear that Reinforcement Learning has profoundly impacted AI research this year, but outside of the labs, RL use cases remain surprisingly underdeveloped. In this new post, I set out to dig deeper into this discrepancy. 🧵👇 greylock.com/greymatter/rl-i…

i'll be at @NeurIPSConf starting thursday, and am hiring for a very high impact role which bridges research and applications. DM if you'd like to chat! Applied Research is the newest team at @primeintellect. we are doing real-world RL and building research infra for everybody.

A slur against Indians on an official Polymarket account

I've been having this strong, strong instinct lately that RL is starting to look like what used to be the 'symbolic systems' approach to AI vs what transformers were supposed to be: stochastic predictors of future state. In the late 1950s, Newell/Simon designed a "general problem solver" theory - simply put, they saw the world as a complex web of 'if-this, then-that' state machines. The whole premise was built on the idea that intelligence = reasoning over explicit symbols and rules. Symbolic systems excelled at constrained logic (e.g., chess, theorem proving) but collapsed in messy, real-world situations. They couldn’t handle ambiguity, noise, or missing data because everything had to be explicitly represented. It was also unscalable.. just impossible to code “everything you need to know about the world” by hand. And finally, they weren't grounded on perception/experience. Smbols had no inherent meaning, e.g. “if fever ==> cough then pneumonia,” but the system doesn’t understand what fever is. It couldn't connect it to the fact that fever is a 'disease', for example. Transformers/GPTs were built to be stochastic. They work backwards from how the world operates. I guess, in a way they 'observe' the world and intuit patterns and, in doing so, transformers can emulate symbolic reasoning: they can “learn” to do logic-like operations because large-scale pattern learning gives rise to emergent reasoning abilities, ironically. It's a monkey, but a monkey that sees the universe because it's trained on infinite data. RL today even though feels more grounded (“do what works”), struggles to define the reward and credit assignment (i.e. which actions led to success) similar to how symbolic systems struggled to codify the rules. As tasks get more complex, the credit assignment problem attributing outcomes back to earlier decisions is going to become exponentially hard IMO. Large-scale RL systems like AlphaGo require massive compute, carefully engineered environments, and lots of human shaping. RL may scale better than rule-based systems, but it still doesn’t naturally model real-world causality, or long-horizon planning, etc, and these were the the very things symbolic AI was meant to capture and couldn't. I wonder what's the way out of this or if I am mostly just imagining this, but it does feel like we are going to hit a cul-de-sac on RL esp in physical environments. I'd be keen to hear from folks doing some deep work on RL and if the SOTA work here is pointed in a different direction. Also as I've been reflecting more and more on a tweet @karpathy had put out a few months ago on RL.. and feels like the field needs a bit of a re-think?

> Train models end-to-end with RL in your own environment/application > RL facilitates building specialized models > RL is an infra challenge While all the big labs are doing everything they can to convince companies to build on their closed APIs/models, despite telling everyone they'd steamroll them, I believe Cursor is showcasing a far more beautiful vision of the future here. Enabled by open models and training infrastructure, we can create a free market of startups and companies that outcompete the big labs in every vertical and capture the upside themselves.