been trying to teach a machine to watch parkour since march. still not sure it can actually understand it i do parkour and i like messing around with ml, so the idea: film a trick on my phone, have it tell me what i did. which trick, how many flips, how many twists it turned into a whole stack. one model for my joints frame by frame, one that rebuilds my body in 3d, some physics to split the rotations, llms to deal with trick names thought 3d would be the easy part. it's where it broke first. those models learned from people walking and sitting, never anyone flipping. on a trick with two and a half turns, one of them counted almost four then names. ~2,700 trick names and even traceurs don't agree, "hyper" means 4 different things depending on who you ask. ended up writing my own code that describes a trick by physics only: axis, direction, turns the thing that bugs me: it can count the turns in a flip, but what i see in a good run is someone reading a wall and picking a line. not sure that's ever measurable super early. most fun side project i have right now

set up a personal agent again last tuesday. the next day anthropic announced duet i'd run clawdbot (now openclaw) from january until april, when anthropic cut third-party tools off claude subscriptions. never replaced it this time it's hermes from nous research, on a vps, through my claude sub. it has my calendar, email and backlog. when i work on a repo it writes the prompt for the coding agent, which spins up its own subagents, so i only ever talk to one thing last week it also made me an audio episode on a crypto concept i was shaky on, for the drive to the office meta shipped muse in early september, openai dots three weeks later, now duet. when the three biggest labs ship the same product within six weeks it's not a dev toy anymore i'm keeping mine though. it's my box, my tools, and i can swap the model whenever a better one drops

how can you learn this stuff if you’re not practicing frequently like this guy has an ai lab to run and is still able to outperform the entire humanity + ai agent i can’t believe it’s just natively in his brain and he was born like that

on oct 5 one of our ambassadors skipped the @citemefr dashboard and handed their project to an ai agent it gave up. its verdict: "to go beyond the score, you need the web interface" https://t.co/28QEECAgWT

doomscrolling research papers while shitting what a generational run

citeme passed $10K MRR this week. same week we added Siri AI as the tenth engine you can audit against. Apple shipped citations in Siri with iOS 27, so there is one more surface where a brand either appears in the answer or does not, and no analytics that tells you which. it is the fastest and cheapest engine in our list. 7.7s average, ten source annotations per run, about a quarter of a cent per query. ten engines now. still the same question: what does AI search actually say about you.

I’m 20, and I’ve spent far more evenings in front of a terminal than in a bar. I’m not saying this to complain, let alone to play the martyr. I missed out on things—outings, silly late-night antics with friends, hours spent scrolling without guilt. The classic student experience, the one everyone tells you not to miss. I only caught a small part of it. A year ago, I understood nothing about the subject I now work on. I’d nod in meetings and re-read everything every evening. It lasted for months. Now I understand. Not everything—but enough to hold a real conversation, and most importantly, enough to know when I’m wrong. There was no sudden breakthrough. Just a year of attention. What I want to tell people my age is that the starting line is much lower than it looks. We look at AI and think it will eventually do everything better than we can. Maybe that’s true. But today, a single student can build things that once required entire teams and years of work. And at this age, you have nothing to lose—no reputation to protect, no one’s rent depending on your decisions. The worst that can happen is spending six months on something and then changing your mind. Six months, at 20, costs almost nothing. Pick something and stick with it long enough to get past the point where everything feels confusing. It takes longer than you think—but it does happen.

It seems so crazy that these Chinese models are so cost efficient and ahead of the current state of the art research. Even though they don't have the biggest research budgets like the top frontier AI American labs, seeing these Chinese researchers come up with new systems to constantly improve their smaller models to compete with the frontier labs that have tens of trillions of dollars is amazing. And I can't even imagine if these Chinese labs had access to all the compute and funds that the Western labs have. It will be totally insane. I can't wait for these models to make better use of the LLMs by maybe building better harnesses and tools and software around it.

grateful to be around those wonderful builders in geneva this week

I will be attending the GDC in Geneva next week with the @verana_io team, it's a truly amazing opportunity to be apart of in this new technological revolution and be working on the open trust layer with a great team. i'm excited to attend the conferences and the talks about those different subjects. Let's meet there !