post AGI classics to read during the weekend: - Scenarios for the transition to AGI @akorinek - The Intelligence Curse by @luke_drago_ - AI Enabled Coups by @TomDavidsonX - The Ambition Singularity by @danfaggella - Gradual Disempowerment by @jankulveit - How long before Super-intelligence (1997) - The British Industrial Revolution in Global Perspective (2009) - Situational Awareness The Decade Ahead @leopoldasch - AI Monotheism vs AI Polytheism by @BerenMillidge

Today we are open-sourcing 🥕Karotte, our framework for building RL environments. We've used it for the past year to build MLE RL environments for frontier labs, and it's been hardened through 1M+ evaluation runs and red-teaming.

We’re on a mission to build a billion AI models, not trained by us, but by you. Introducing @bryel_labs we help your company build its in-house AI lab and give you the software to make model development and continual improvement simpler. We are backed by @ycombinator and have trained models at Apple, Framer, and F500s. We’ve enabled companies to train: • Classification models that are 20x faster • Search models that cost 5x less. • Long horizon agent models that match the frontier at 5-8x lower cost. Are you interested in partnering with us? Reach out. bryel.ai/

We have seen a growing and concerning trend of startups supplying data to China. Today we are taking a public stance against this practice At fleet, we believe this represents one of the largest national security threats today and must be stopped. We must win the AI race for America and shape AI with the same democratic values that have shaped our nation Our full statement here: fleetai.com/supporting-ameri…

6-12 months ago the large labs started buying bio data this spend is starting to ramp very aggressively. the labs are somewhat indiscriminate buyers with enormous budgets. the therapeutics market has long had buyers of data (large and small pharma) but they've historically been the exact opposite of huge-budget and indiscriminate this shift will distort the bio startup market in a bunch of ways. there will be a lot of short-termist behavior to try to get in front of this capital firehouse. but unlike the AI data labelers (mercor, et al) and the robotics labelers (mecka, et al), the bio startups serving up this data have a chance to emerge from this period with businesses that don't rely on selling data the best bio startups will use these sudden resources the labs are dumping on them to supercharge their novel assay and work towards independently powerful models and eventually directly produce therapies bio is the final frontier and its a very exciting time ahead for bio startups IMO

We investigated a seemingly simple question: how do you grade if AI generated parts or assemblies are correct? Here's how we improved on this with our verifier, FrontierCAD.

Backprop has been the only credit assignment algorithm capable of training large neural nets. Introducing Dust, the first zeroth-order method to pretrain transformers to approach and sometimes even *exceed* backprop with large amounts of computation (bitter lesson!) - Dust is on the order of 1,000 to 10,000x more compute efficient than EGGROLL, the state-of-the-art ES method, for training transformers. - We chose pretraining transformers because it's arguably the hardest task possible for zeroth-order optimization. - Search in high dimensions is really misunderstood: zeroth-order usually gets better with overparameterization, not worse, i.e. at a fixed population size, larger models (up to 120x) reach a lower loss than smaller ones. - Backprop-like gradients emerge from a large population of activation perturbations, and the alignment holds up at every scale we tested, up to 1B tokens, which is critical for scaling. The core idea is a *virtual population*, which helps scale to large populations in transformers by perturbing activations in parallel. w/ @bishmdl76, @cs_serdar, @akshayvegesna

Thought Jev is fast? We evaluated Jev on cuaspeedrun.com for computer use agents and found: 1. 1.5x slower than Opus-5.5-xHigh 2. 10x lower accuracy 3. 5x more steps That puts it in the GPT-4 league of performance and Astra-xHigh on time What and Why?? 🤯 🧵

Building AI inside your company? These are your people. Two @bryel_labs events at SF @Techweek_ next week for founders, CTOs, and engineering leaders: Tuesday evening: Join me and @vinitra_s from @ScholeAI to discuss what it takes to level up your intelligence within your company. (Co-hosted by @thehousefund) 🗓️: partiful.com/e/wmdrrEPxrjo20… Thursday afternoon: Curious about training your own models? Building your intelligence stack? Come swap notes with other pioneers making the same decisions. 🗓️: partiful.com/e/BgX5jT0yxM3fa…

Today, I am excited to announce Animation Bench. Our first benchmark to evaluate frontier models on web animation reconstruction. Frontend generation is a common commercial use case for coding agents. In addition to static layout it contains transitions and interactions triggered by dynamic behaviours. We evaluated four frontier models on 48 tasks from real websites across three axes: 1. Visual Similarity 2. Motion Consistency 3. Layout Correctness. We find that motion consistency is often incomplete or missing.

Introducing Synthetic Hospital: an open, fully synthetic longitudinal EHR benchmark with verifiable ground truth! 1,268 patients, 5,602 encounters, zero PHI. Physicians could not reliably distinguish its charts from real ones. 📄 arxiv.org/abs/2609.30027 💻 github.com/sparkcpark/synthe… ✍️ sparkcpark.github.io/posts/f…

In 12 weeks, we built a research facility that is run entirely by AI. AI designs, executes, and observes experiments end-to-end across biology, chemistry, and materials science. We’re introducing SciUniverse: a benchmark that measures AI’s ability to do real-world scientific research.

There’s a good chance your open source model is costing more than the frontier. Cheap tokens ≠ cheap tasks. Here, we introduce Intelligence Density, and Density Aware Training, our post-training technique to achieve less wasted compute, better learning, all with no knobs to tune. Enabled by default in every Trajectory model.

More people need to know about this lab. qlabs.sh/

My new report with @luke__emberson.

Reinforcement learning is having a moment. For the past few months, we’ve been working with @tryterac to put together something special: Convergence — an event for the people building the future of RL. Speakers from @GoogleDeepMind, @databricks, @mercor, @arena, @joinHandshake, @turingcom, @deeptuneai, @contra, @Stanford, @Princeton, @UCBerkeley, and more. Our goal: make this THE event for the RL community. Oct 29 in SF. Come join us → terac.com/events/convergence…

Trajectory is at @HackTheNorth, come say hi!! We’ve got some cool merch in store for whoever can come up with the wackiest benchmark ideas

Introducing WeirdML v3, a fully agentic benchmark featuring 11 complex hand-made tasks. Models must explore and understand unfamiliar data, develop ML and data analysis pipelines and produce results despite limited data, unspecified goals and/or very limited feedback. 1/8