I recently left @DbrxMosaicAI to join @flourishailabs. Why did I decide to make the jump? I believe that some of the most interesting problems in science might now be within reach. With all the wild advances in AI, there’s no better time to pursue the big questions. Flourish is doing just that. We are a mix of AI researchers, engineers, and neuroscientists tackling one of the most ambitious problems in the field: human-level intelligence and learning with human-level efficiency. I have immense respect for the founders @TRReardon , Robert Williams, @jovo, @_vishar and for the team - many of whom built CTRL-labs from the ground up. I really loved my time at @databricks and MosaicML before that. The caliber of the people and the research is sky high, and I am incredibly proud of the work we did over the past 5 years. In particular I’d like to thank my mentors @jefrankle , @mcarbin and @bemikelive, as well as the cracked @DbrxMosaicAI team. If you’re interested in working at the intersection of AI research and human intelligence, don’t hesitate to reach out.

Actually it's fine guys! I figured out a way, see below. Claude Fable 5 is a great model afterall, and I also finally appreciate the difference between CLAUDE.md and AGENTS.md. It's all good.

Curious if people have evidence of Mythos doing poorly on frontier LLM research in the coming days

At @flourishailabs, we are tackling one of the hardest problems in AI: human-level intelligence and learning with human-level efficiency. We are inspired by biology, seek guidance where appropriate, but aren't dogmatic about it. If this appeals to you and you'd like to embrace the age of research, consider joining us. We are hiring in NYC and SF.

We’re releasing Nemotron-Labs-Diffusion - the first Tri-mode LM family (3B/8B/14B) that switches between 1⃣Autoregressive, 2⃣Diffusion, and 3⃣Self-Speculation decoding by simply changing the attention pattern/mask. One model Three decoding modes. No extra draft models. No architecture changes. Just significantly better efficiency across different concurrency levels. Up to 4× higher real throughput for a single user. 🤗 HF Collection: huggingface.co/collections/n…, open license 🛜 Project page: research.nvidia.com/publicat… 📰 Tech report: bit.ly/Nemotron-Labs-Diffusi… Details below 👇

Just an AI that can fix my slouchy chungus life

I have a personal update to share. I have taken a new role as head of research at @arcee_ai . I have been constantly so impressed with the talent density, determination, and drive of the Arcee team and I am delighted to join forces to help shape and deliver their vision for open source frontier models. The fastest progress in AI happens in the open. When models are accessible, iteration compounds, and entirely new categories of products become possible. The decision to leave @datologyai was a difficult one. I still believe in their vision and incredibly talented group of people that have pushing the frontier of what is possible in data curation. I was personally motivated by being back directly involved in releasing and deploying models that people and companies use everyday to solve problems. This is what I love to do. BTW we are hiring. If you want to be part of a cracked small team making fantastic open weight American models dm me!

🧠 the Digital Brain Project is now live: $5M total · up to $500k per selected team Let's open-source the modeling of the human brain brain activity! ➡️Apply on: digitalbrainproject.org/

🚨 We're very happy to introduce TRIBE v2: a foundation model of the brain's responses to sight, sound & language. 📄 Paper: ai.meta.com/research/publica… ▶️ Demo: aidemos.atmeta.com/tribev2/ 💻 Code: github.com/facebookresearch/… 🤗 Model: huggingface.co/facebook/trib…

I implemented this in JAX: `uv pip install flashoptim-jax`! Read below for the journey over 4 implementations to get speedups over the torch baseline while matching the ~2x memory improvements.

BrowseComp-Plus, perhaps the hardest popular deep research task, is now solved at nearly 90%... ... and all it took was a 150M model ✨ Thrilled to announce that Reason-ModernColBERT did it again and outperform all models (including models 54× bigger) on all metrics

the next @airstreet nyc ai meetup is on 5/14 if you're building and researching in ai, this is the place for you to meet peers, learn and share best practices, find co-founders and friends you can rsvp at luma(dot)com/nycai

The newest model in the Mamba series is finally here 🐍 Hybrid models have become increasingly popular, raising the importance of designing the next generation of linear models. We've introduced several SSM-centric ideas to significantly increase Mamba-2's modeling capabilities without compromising on speed. The resulting Mamba-3 model has noticeable performance gains over the most popular previous linear models (such as Mamba-2 and Gated DeltaNet) at all sizes. This is the first Mamba that was student led: all credit to @aakash_lahoti @kevinyli_ @_berlinchen @caitWW9, and of course @tri_dao!

Quotient is joining @databricks! @freddiev4 and myself started @QuotientAI in 2023 with a dream and a hunch that the next era of AI will be unlocked by systems that can measure, evaluate, and improve AI in the real world. Over the past couple of years, we partnered with companies ranging from AI-first startups to Fortune 500 enterprises. They used Quotient to monitor AI systems in production, ensure their agents met strict enterprise policies, and reinforcement fine-tune agents to consistently achieve double-digit performance gains. We built it all at the bleeding edge, with a small team that started in Boston, Massachusetts. Joining Databricks gives us the scale and resources to take this much, much further. Very grateful to our team, investors, customers, and everyone who supported us along the way. We couldn’t be more excited for what’s next. Onwards!

The OfficeQA Pro paper is live! Excited to share this benchmark for evaluating frontier agents over enterprise-style documents and workflows. Amazing work with the team to create a benchmark that's both meaningful and challenging for leading agents (and even humans)!

Most AI benchmarks test reasoning in isolation. Real enterprise tasks require grounded reasoning: 1️⃣ Find the right documents 2️⃣ Extract the right values 3️⃣ Perform analyses OfficeQA Pro evaluates this end-to-end. Frontier agents still score <50%. 🧵Paper & details below!

Going to do a more technical deep dive on our enterprise knowledge agents and how we train them with RL. Overall we found that simple, yet principled off-policy RL works at scale for complex agentic tasks with hundreds of steps of tool use and context management. Here are the key takeaways from our 80 page technical report. (1) RL does not just sharpen base model's distribution. We see test-time scaling improves consistently over the iterations of the RL training. Skills learned during RL transfers to unseen prompts and agent learns to solve prompts where base model has zero accuracy under pass@16. (2) Multi-task RL generalizes really well Simple mixing of training data from multiple tasks works well and allows multi-task RL scale beyond your in-distribution training tasks. We found that multi-task RL just works better than multi-expert distillation. (3) End-to-end RL for tools and context management works best. We skipped mid-training, and directly trained everything end-to-end using RL at scale (2m tokens per gradient computation). Models learned to use vector database tools and context compression at the same time.