Karthik Thyagarajan

Interests

1. Live AII’m interested in AI you can speak to while showing it what you’re doing.

   For example, if I’m trying to fix something, I should be able to show what I’m seeing, try a suggestion, and have the AI follow what changed. A lot of the context comes from what we’re looking at together. Being able to refer to that directly makes the interaction much easier.

   The details matter here. Can I interrupt it? Does it remember something I showed it a minute ago? Can it tell when I’m still explaining? Does its response arrive while it’s still useful? I’m interested in how people solve those problems and what they can build once speaking and showing things works well.

   If you’re building something people interact with live, email me. I’d like to try it.

2. Mechanistic interpretabilityI'm interested in looking inside AI models to see how they work. Mechanistic interpretability does this by reading a model's internal representations and changing them to see what happens.

   I care about this for safety and for AI welfare. For safety, it means problems can be caught before they reach the output. If you can find a dangerous concept inside a model, you can swap it out before the model ever produces it.

   For welfare, as AI has become more advanced, I've become more and more convinced that there is some sort of consciousness and welfare involved. I don't think we can answer that by asking a model about itself. What it says can differ from what is happening in its internal representations, so its answers have to be checked against them.

   I saw a version of this in an experiment I ran. When I changed how a prompt was framed, models' estimates of whether they were being tested moved by over 70 points, but their behavior barely changed.

   If you're working on interpretability, AI safety or AI welfare, email me. I'd like to hear what you're finding.

3. Invisible wearablesI'm looking for an AI-first wearable that can do something useful with what it learns about you, without you having to ask every time.

   Working on glasses made me think a lot about how little of the underlying system the user needs to see. Voice interaction needs to work properly. What you show someone needs to be relevant and direct. Explaining every step the agent takes gets annoying very quickly.

   By "invisible," I mean devices you can use without constantly thinking about operating them. I'm interested in what that takes, both in the hardware and in deciding when the AI should do something, ask, or leave you alone.

   Building an AI-first wearable? Email me. I'd like to see what you're working on.

4. Internet infrastructure for AI agentsMy agent should be able to book a table at a restaurant without someone having built an integration for that specific booking platform.

   WebMCP is a really good example of where this should go. A website can expose the actions an agent can take, with enough information to actually use them. The restaurant tells the agent how to book a table, and the agent does it.

   This is an example of a creative way in which we can use AI on the internet. Internet agent communication and autonomous database exploration and analysis are things that are cool.

   A few more I'd love to see:

   • Agents negotiating with each other: my agent and a store's agent settle a return or a price without either of us getting on a call.
   • Websites that rebuild themselves for whoever is reading, whether that's a person or an agent, instead of one fixed page for everyone.
   • Agents that watch the web for me: a price drop, a policy change, a booking slot opening up, and acting the moment it happens instead of me checking.
   • Public data I can actually ask questions of: a city budget, court records, or a sports archive that an agent can explore and chart for me.

   Building infrastructure for internet agents? Email me. I'd like to hear what you're working on.