Today we're publishing our first research blog, Understanding a Law Firm through Study. We're sharing a glimpse of a future where agents are trained with native memory:

Thank you to @Nasdaq for supporting Engram on our launch day yesterday! Some have commented that this photo looks AI-generated. It's not. This really happened. Feel free to send this picture to your moms. We're certainly going to.

We are immensely grateful to @buckhouse @Aweiland from @sequoia for the branding, storytelling, and website. They were generous with their time and ideas and we learned a lot from our process with them.

The single biggest impediment in AI isn't absolute intelligence or coding ability. It's deeply understanding large repositories of knowledge that every person and every company has. Our repositories will explode in size, with AI agents writing much of them. This is a two-way street: we need to understand what our AI generates, and AI needs to better understand us. A breakthrough in memory is needed – gradients need to be taken, and our team is the right one to take the big swing.

Part of my speech at the Columbia Medical Center PhD hooding ceremony: “The world we’re graduating into now has machines that review literature, write code, analyze data, and even draft manuscripts that pass peer review. Our scientific profession is shifting beneath our feet. Where does this leave the graduating class? I suggest we focus on three things: deep domain knowledge, scientific taste, and character. First, deep domain knowledge.  AI can answer general questions well, but when you go deep, you find that the AI is wrong often, and sometimes in fundamental ways. The world will soon be flooded with AI-generated content. As PhDs, we become world experts in our domain, among the few able to detect subtle AI hallucinations. We can also be a part of the solution, if we so choose. We can help shape AI to better serve our fields. Second, scientific taste. People talk about the “AI scientist”—a machine that reads papers, spots gaps, and proposes new experiments. This is starting to happen. We shouldn’t compete, but collaborate. As PhDs, our edge lies in intuition: we can guide AI more directly and elegantly toward the discoveries. Such collaboration will also be cheaper and consume less datacenter energy. Remember: we know many things that the internet doesn’t. Unpublished results, failed experiments, and informal conversations over the years. We may better know what to try next and what to avoid. Third, is our character. In a world where every scientist has access to superintelligence, what sets us apart is not what we know, but how we lead. Memorizing the literature and crunching numbers will not be our competitive edge. The scientists who shape the future will be the ones who can guide teams, make good judgments, and act with clarity.”

How can we use small LLMs to shift more AI workloads onto our laptops and phones? In our paper and open-source code, we pair on-device LLMs (@ollama) with frontier LLMs in the cloud (@openai, @together), to solve token-intensive workloads on your 💻 at 17.5% of the cloud cost while maintaining 97.9% of the accuracy. See Gru and the Minions in action below, 🔉on please (h/t @cartesia)!

⚡️Announcing️Lightning Pose⚡️ A pose estimation method that learns from both labeled images and a lot of unlabeled videos, via inductive biases like temporal smoothness, posture feasibility, and multi-view consistency. Preprint: tinyurl.com/litpose

@Natalie_Bid, @DShohamy, and I have some new work suggesting that people apply policy gradient algorithms to experiences retrieved from memory: gershmanlab.com/pubs/Biderma…