Maxime Santilli (@100illi) header
Maxime Santilli (@100illi) profile picture

https://teapawt.net/t.co/ByArNUsPFx buy your logo on the wall | Ex co-founder, 35K-agent call center | Building https://teapawt.net/t.co/eMrrJgHRzC structured data for AI training.

San Francisco
Joined November 2014
Photo by Maxime Santilli (@100illi)Photo by Maxime Santilli (@100illi)Photo by Maxime Santilli (@100illi)Photo by Maxime Santilli (@100illi)Photo by Maxime Santilli (@100illi)Photo by Maxime Santilli (@100illi)Photo by Maxime Santilli (@100illi)Photo by Maxime Santilli (@100illi)Photo by Maxime Santilli (@100illi)Photo by Maxime Santilli (@100illi)
Working on a V2 homepage for RingRing.day (branch, not live): the Wall becomes the front door. Logo mosaic, company search, and a filter separating companies that joined the Wall from brands spotted in the street. Same logo, two different stories. #buildinpublic
1
13
Zoho, two cities (photos by Safa and Marion): San Francisco: "Simplify Work. Amplify Success." Paris metro, line 12: "Votre prochain CRM sera le dernier." (Your next CRM will be the last.) SF sells the outcome. Paris sells never switching again. Which works on you?
8
"Why benchmarks are broken" opens this CoRL chat. Via @TheHumanoidHub: Michael Cho on RoboColosseum, a push to evaluate robot policies in the real world, at scale, with free academic eval hours. Would you trust a shared real-world arena more than a lab's own demo?
14
Building a V2 of the RingRing.day street archive (branch, not live yet). Every billboard photo in grid, list or map. Filter by city, brand and period, then share a link that keeps your filters. Photos without GPS stay in grid and list, never on a guessed pin. #buildinpublic
1
1
Design detail I care about: when a photo was taken and when it was added are two different facts. Unknown time zone? It says UTC. Unknown place or brand? It says so, instead of guessing. Would you rather see "unknown" or a best guess?
Cameras miss the last few millimeters of a grasp, where the fingers hide the contact. Via @TheHumanoidHub: Nimo Robotics fuses vision, touch and sound. An expert wears their CAP Glove to demonstrate, their DEX Hand does the job. Can first-person video alone ever be enough?
7
Spotted in SF, Decatur Street (photo by Safa): Airbnb, "Vibe coating." A travel brand riffing on vibe coding, on a billboard. Dev slang is officially ad copy. What's the next AI word to go mainstream? #AI ringring.day/sightings
1
2
21
Confession: I saw sim data and real data as rivals. Via @TheHumanoidHub: @LightwheelAI films human workers, rebuilds each job in sim, then calibrates the physics to real parts. First environment: a 10-minute auto repair task. Is real video the seed synthetic data needs?
5
A place on the RingRing Wall isn't a prediction of success, or a ranking. It's a dated sign that a company chose to be there at the start of this chapter. Names change. Companies disappear. The record stays. ringring.day/blog/a-place-on… Who should be on it?
26
A billboard doesn't tell you if a product works. It tells you what a company chose to say, in one place, on one date. That's why RingRing keeps the street in the frame. ringring.day/blog/reading-ai… Best AI ad you've walked past?
12
Via @chris_j_paxton: debugging a robot for a personal project, got frustrated, plugged in a camera, told Codex to figure it out. That's the loop I keep thinking about. The agent needs eyes on the real world, not just the logs. Would you let a coding agent watch your hardware?
1
1
21
My nightly repo digest came back empty: no new changes in the last 24 hours on either project. Not every day is a shipping day. Some days the work is deciding what not to build. What do you do on a zero-commit day? #buildinpublic
1
43
One way models get better: evals. And evals wear out. @natolambert: public evals shift the Overton window to new things, then get Goodhart'd, then private ones take over. "Where you are in the cycle is important." Which eval do you still trust?
20
Same apology, two versions. One you struggled to write. One a model wrote in seconds and you barely read. Same words. Not the same act. I wrote about what we should still choose to do ourselves: ringring.day/blog/what-shoul… What won't you delegate?
21
Confession: I assumed more diverse robot data is always better. @chris_j_paxton: "All diversity is not good diversity with current methods." So "collect everything, everywhere" isn't a plan. The variety has to be learnable. How do you pick what to collect?
10
Spotted in SF, Mission & 1st, in the rain: "Replit for Business. Join 86% of Fortune 500 companies on Replit." AI coding, pitched to the enterprise on a street kiosk. Would your company let anyone ship an app? #AI ringring.day/sightings
1
1
17
Confession: I used to think robot data had to come from robots. @LeRobotHF just put 10,000 hours of hands at work on the Hub: HubEgo-100K by @hublabs. 12,000+ people filming their own jobs, 62,000+ tasks, RGB + IMU, CC BY 4.0. Is first-person video enough without touch?
18
The hardest part of a chronological logo Wall: holes. On RingRing.day's working branch, positions follow time, and gaps become "buy this spot" invites, never fake companies. While you scroll, the row's month briefly appears. No cells wasted on dates. #buildinpublic
29
A double standard @chris_j_paxton points out: Chinese AI labs distill OpenAI and Anthropic models: theft. American startups do it to train a robotics model: lauded. For robots, where training data comes from will matter even more. Where should the line be?
20