CONFERENCES
Conference notes.
Notes from the AI events I attended this year: the ideas that held up, and what they changed about how I build.
2026
RAISE Summit
MACHINA Summit
NVIDIA GTC
SOON
PARIS · 8–9 JULY 2026 · CARROUSEL DU LOUVRE
Two days across three stages — Master, Grace Hopper and Ada Lovelace — covering models, infrastructure, energy, enterprise adoption, and what it actually takes to run any of it in production.
I could not answer the question I was asked
A VC asked me over coffee how I thought AI would progress. I did not have a good answer, and I did not want to repeat what lab CEOs say on podcasts. What I could talk about was the machinery underneath, and that turned out to be the better conversation. A model can be finished while the power, the memory and the business case are not.
Agents broke the metric I had been optimising
Sachin Katti said token use among OpenAI’s own researchers grew close to 50× in a matter of months, once Codex became the default way to work. Time-to-first-token was the right number when a person sat watching a cursor blink. It tells you almost nothing about a run that takes twenty minutes and makes forty tool calls. I came home and changed what I measure.
The bottleneck is memory, and I had been ignoring it
Pat Gelsinger called HBM a lousy memory with a senior SK hynix fellow sitting beside him, and the SK hynix fellow largely agreed. Stack twelve to sixteen layers of DRAM and you get a thermal sandwich where one defect anywhere kills the whole part. I write software, so I had quietly filed memory under someone else’s problem. It is most of why inference costs what it costs.
You can now write code faster than you can read it
Manoj Nair put up Snyk data from 4,800 customers showing vulnerabilities up more than 108% in a year. That number is the symptom. The cause is that generation got faster and review did not. Then he widened it: the agent’s tools, its shell access and its network calls are all surface somebody is supposed to be reviewing.
Open weights stopped being an ideological position
The reasons I heard were dull and practical. Cost. Where the data physically sits. And a real reluctance to hand a proprietary business process to a lab that might compete with you next year. Nobody on stage argued it as a principle. They argued it as procurement.
LeCun thinks we are predicting the wrong thing
His example stuck with me. Predicting every pixel of the next video frame is a poor way to learn that a cup falls off a table, because most of those pixels are noise nobody could predict. Predict in an abstract space instead. I am not qualified to judge whether that road ends at general intelligence, but the narrow version needs no defending: if the physics never appears in your training data, the model will not learn the physics.
WHAT I CHANGED
Five things I do differently now
I came back with a much longer list than this. These are the five that survived contact with my own work.
I stopped reporting time-to-first-token
For anything agentic it measures the wrong half of the job. What moves a decision is whether the task finished, how long it really took, and what I paid for the results I actually kept.
I price the failures into the token cost
An answer that gets regenerated, corrected by hand, or escalated after it caused an incident was never cheap. Cheap tokens, expensive mistakes was the sharpest line of the two days.
I build the eval set before I touch the router
Fifty real requests with known-good answers. Without that, sending cheap work to a small model is just a faster way to be wrong, and the savings report will look excellent while it happens.
I write down what an agent can reach before I widen its permissions
What it can read, what it can write, what it can call, and which actions need a human first. If I cannot produce that list in about a minute, it already has more access than I think it does.
I stopped assuming available means servable
Racks at 130 to 140 kilowatts heading toward a megawatt, and buildings measured in years. The date a model exists and the date you can serve it at your volume are not the same date.
This page is the short version. The full write-up has the sessions worth watching with runtimes, the hardware I saw on the floor, and every reference.