What has been most striking about the past year in AI?

The progress of LLM-based agents! When Jonathan Zittrain, Jordi Weinstock, and I taught Agentic AI and the Law in the spring of 2025, we used OpenAI Operator to order cookies for everyone on the first day. It went okay (I had to intervene a few times, especially when it tried to deliver to Portugal), but it was an interesting demo to help illustrate what might be coming.

As I write this now, in January 2026, I think we’ve passed the “interesting demo” stage. I spent a lot more time with Claude Code over the holiday break (as it seems many others did too, judging by my social media feed), and the capability and progress have been genuinely amazing to me. Using it on a variety of BKC and side projects (including embodying Claude in a Reachy Mini, highly recommended!), I was struck by how much less I needed to dig into the codebase to fix stuff or re-direct the model than compared to just a handful of months ago. The bitter lesson continues, it seems 🤷‍♂️.

What (AI-related) worry keeps you up at night?

Eh, do I have to just pick one? Related to my previous answer, I often think about the where, when, and how we should adopt LLM-based systems in our daily lives (and what happens next if we do or do not). Concretely, I pretty strongly believe that as of today, tools like Claude Code, OpenAI Codex, etc. are better at writing and shipping code and are better at designing and running experiments than most junior engineers and researchers I've worked with in my career. That's not to say these systems don't require significant oversight, workflows mindful of their limitations, and guidance tailored to the specific system’s eccentricities, but I'd say all of that about junior engineers and researchers too!

So as we plan out what our org looks like to pursue the opportunities and address the challenges in front of us, it seems like both extremes "let's pretend we still live in a pre-LLM world, continue with business as usual” and “never hire another junior engineer or researcher again” are the wrong answers. Plus, at an individual level, I struggle to see how, for most white-collar work, those who don’t deeply embrace these tools will be able to meaningfully keep up with those who do. But I hope that's just my limited imagination!

With that in mind, what kind of work do you see as important for us to be doing in 2026?

If we're about to live in a world where each of us are increasingly enmeshed with these systems in our professional and personal lives, I think it’s crucial to build them in a way that empowers the user and looks out for their best interest, especially with the ever-widening capability asymmetries between the system and its user. For example, we have a project on measurement and fine-grained control of the agency of LLM-based agents (that builds on some prior work from BKC community members) and another project to detect changes in the loyalty of a LLM (i.e., is the model following my instruction or the model provider’s trained in instruction?).

Zooming out, another important component of work we’re doing at BKC is supporting dialogue and understanding across the various AI communities, as exemplified by the AI Triad course JZ, Jordi, and I are about to start teaching in a few weeks. It’s been interesting to prepare for as I think through shifting that initial tech demo from autonomously ordering cookies, which now seems a bit boring, to autonomously taking the LSAT.

I’d like to end by saying how grateful I am to be a part of BKC. To be part of a research center that takes seriously the values of being a big tent in a time when AI is so divisive, that deeply believes to understand a space you have to build stuff, and that is focused on pursuing the interdisciplinary questions that let us see around the next corner, is an exceptionally special experience.