Notes from ASPLOS 2026

30 Mar 2026

In March I spent a few days in Pittsburgh at ASPLOS and gave a talk in WACI, the Wild and Crazy Ideas session, where you get ten minutes to pitch something you can’t fully defend yet and two minutes for people to ask you why. Mine was called The Asemantic Code Revolution: Breeding Software for Silicon Physics. More on how the talk went below. It is the rest of the conference that I keep thinking about.

The idea, in one paragraph. Every layer of the software stack exists so that humans can coordinate with each other: the ISA, the ABI, the careful handling of every NaN in IEEE 754, the timing headroom a chip carries because no two dies come out of the fab the same. Each of those layers is paid for in cycles and joules, and published measurements put the total at something like 59% of what a processor does. The wild part is to stop shipping binaries and ship intent instead: formal invariants that a local synthesizer turns into code tuned to the exact die it runs on, with a proof certificate the hardware checks once at load time. The code it produces is unreadable by design. That is the “asemantic” bit. The paper is three pages if you want the rest.

The conversations

The best part of the conference was not a session. It was the conversations in between: in the corridor, over coffee, at lunch. I asked questions after talks, introduced myself to people whose papers I had read, and ended up in a long conversation with a senior leader from my own company whom I would never have met in a meeting. What struck me is how open everyone was. Nobody was guarding their time. Most people at a conference are hoping someone will come and talk to them, and a question about their work is the easiest way to start.

The most useful advice I got came from a longtime ASPLOS attendee: go to the business meeting. Most people skip it, yet it’s where the conference governs itself in the open. The chairs break down submission and acceptance numbers, explain recent changes to the review process, and open the floor to debate what should change next year. Conference policies aren’t handed down from some central authority. They’re decided by volunteers in a room anyone can walk into. I showed up knowing very little about how the sausage gets made and left understanding not just why the review process looks the way it does, but why it’s worth having a voice in shaping it.

What I learned about how conferences work

Almost everything is about AI now. Nearly every paper across the sessions touched LLMs somehow. My first reaction was fatigue. My second was that this is probably right. Better tools for thinking make every other kind of research better, and the people in the room were working out how to use them well.

Review is not as clean as we pretend. Roughly one in ten submissions gets into ASPLOS, and when you talk to enough authors you realize a lot of it comes down to chance. That is not a reason to stop submitting. It is a reason to not read a rejection as a verdict on the idea.

I want more of this. Researching something, building it, writing it up. This paper carved out a small area and I intend to keep working in it.

How the talk went

Well, mostly. I prepared more than I thought I needed to, and it showed. The delivery was steady and I wasn’t nervous, which surprised me.

Two things I would do differently. My slides were too dense. I knew it while making them and did it anyway, because cutting felt like losing. And the delivery started light and got intense. By the middle I was serious and cramming in details, when the whole job of a ten-minute talk is to make people curious enough to read the paper. Next time: fewer words on the slides, one idea per slide, and leave them wanting the paper.

If you were in the room, or you read the paper and think I’m wrong, I’d like to hear it. The idea is wild on purpose. I also think it is roughly where things are heading, and I’ll have more to say about that here.