Where I've been and where I'm going

Earlier this year I built http://contextgraph.me -- a 2d infinite canvas that broke LLM conversation turns into nodes and allowed the user to arbitrarily rewire them into a DAG. I built it mostly to compensate for distraction, context poisoning, etc, in frontier models and in even less capable open source models.

At first it felt like a revelation-- all these new patterns-- creating chains of user-authored nodes instead of having long threads ('potential token savings?' I thought), being able to make really fine-grained, precise edits on long pieces of text, accurately, without having read the text. For certain research-specific tasks that involved a lot of iteration and branching of ideas (that I could then visually merge in the canvas), it felt like a natural extension of my thought process.

Nonetheless, manually writing graph topology added a layer of friction to conversations. For actual chat (outside of coding agents), I had mostly been using Claude web anyway (and swearing at it) all through development. By the time I started building features to ingest conversations created in other chat interfaces and designing UI fusions between the canvas and linear chat, models got better at understanding user intent. I stopped swearing during linear chat conversations. My commit velocity on the codebase slowed.

I tried to evolve the product, chasing PKM-like objects, and found that shifting into "just another memory system" or "just another PKM", wasn't what I wanted to do with my life. Instead I decided to double down on studying and then solving the problems that initially inspired me to build Contextgraph as thoroughly as I could. And fundamentally those problems boil down to something like getting AI systems to understand and act on human intent instead of myopically chasing an objective that is, in some sense, 'wrong'.

I discovered that problem has a name. Getting a system to do what its principal actually wants -- not the literal objective, not the proxy, but the intent behind it is a clean statement of what Christiano calls intent alignment. And the more I look at the world as systems that optimize and learn, the more it generalizes past AI: I can't reliably align myself with my own values, which makes the prospect of a superhuman system with the same gap genuinely disturbing.

During the pandemic (I was an undergrad at UC Davis at the time) I spent a stretch genuinely believing we were being funneled, deliberately and universally, into something I can only describe as a matrix-pod reality-- total disempowerment, dressed up as convenience. Getting saner mostly meant realizing none of that was real; the structure I was so afraid of didn't exist. But the disgust at the shape of it never left, and it's part of why I care about this. If we're going to keep moving into more technologically-mediated realities, I'd rather help build ones that are actually good for people than watch even a diluted version of what I feared take hold.

ContextGraph taught me how hard it is to build something that's both novel and useful, and how slow the feedback on whether you're doing the right thing or not can be. There's no version where my current bets are right; I haven't earned that yet. Which means the real problem isn't picking the right thing to work on-- it's choosing what to optimize for when I can't. So I've been pointing myself two things: one is just fun-- letting myself poke at things with no goal, make totally unjustified conceptual leaps and see what happens-- and the other is trying to align my work with something bigger than myself. The fun half I can supply myself. The other half is harder — I've mostly been thrashing at it alone, trying to figure out which of the big things matter and where I'd be useful. So I'm applying to Bluedot to hopefully learn more, while trying to do the best I can on my own. And increasingly, I'm looking for ways to show up and be useful that don't depend on having perfectly developed research taste yet. I hope this this post is where that starts-- figuring it out in the open instead of alone.