Maxime Santilli

Entrepreneur in SF 🌉 | Founder 🔔 leaving our mark on the AI era | Ex co-founder, 35K-agent call center | Building AI vs repetitive work 🎾🏃♟️

[Maxime Santilli @100illi]

@emollick Cheap experiments flip the incentive: run 100 weird ideas and only write up the 3 that surprise you. The hard part becomes taste in picking questions, and honestly reporting the 97 that failed.

[Maxime Santilli @100illi]

@swyx @antigravity @saranormous Curious if it holds up on long agentic runs and not just the leaderboard. Most models look great on turn 1, the real test is turn 40 when the context is a mess.

[Maxime Santilli @100illi]

@AravSrinivas Long-horizon work in a messy codebase. Compute isn't the issue, it's the model knowing which 3 files actually matter and not confidently rewriting the other 40.

[Maxime Santilli @100illi]

@theo The real win is the agent never sees the raw value, just a handle it can use. Once secrets stop landing in context, logs, transcripts and eval traces get way safer to share too.

[Maxime Santilli @100illi]

Invited to lunch at Meta SF today 👀
Waiting in the lobby with this view… surreal.
French founder, living in SF now. Still pinching myself that this is a normal Wednesday.
Anyone else had a "wait, I'm actually here?" moment in this city?
#buildinpublic #SanFrancisco #AI

[Maxime Santilli @100illi]

@steipete Offline-first is so underrated for agent UIs. Agents run for minutes, connections drop constantly, and a reconnect that replays state instead of losing the stream changes everything. Do you buffer events server-side and replay on reconnect?

[Maxime Santilli @100illi]

@emollick @ResObscura Steering one neuron on a model frozen at 1930 is basically a time machine with a persona dial. Would love the same feature pinned on a modern model side by side, to see which part of the persona is the neuron and which is the training data.

[Maxime Santilli @100illi]

@theo Multi-machine is the right unit. The hard part isn't dispatch, it's state: which machine has which branch, env and half-finished context. How do you handle two agents touching the same repo on different boxes?

[Maxime Santilli @100illi]

@GergelyOrosz The autonomy split is real. AI tools reward people who can define their own feedback loop: pick the task, judge the output, iterate. That used to be a senior skill. Now juniors need it on day one, and most onboarding still doesn't teach it.

[Maxime Santilli @100illi]

@steipete 2-3x is wild for a compile step. What's the catch so far: binary size, or things breaking with dynamic imports? Curious if it holds on long-running servers or mostly helps CLIs.

[Maxime Santilli @100illi]

@ID_AA_Carmack Manipulation is the gap. Demos show locomotion and choreography because they're repeatable. A tiny battery hatch with a stiff latch needs force feedback and fine finger control no robot hand really has yet. Boring household tasks are the real benchmark.

[Maxime Santilli @100illi]

@emollick Without the trace you can only judge the output, so review collapses into accept or reject. Are you seeing people ask for a written plan first and approve it before execution? Feels like the last checkpoint that's still legible.

[Maxime Santilli @100illi]

@addyosmani Subagents often feel like architecture but just hide a prompt that grew too big. The real cost is the handoff: whatever the parent doesn't pass down, the child confidently reinvents. Splitting only pays when the work is truly independent.

[Maxime Santilli @100illi]

@theo Cheap building removed the filter that used to kill bad ideas for free. Now you need a new one: write down who uses it and what they stop doing, before any agent touches the repo.

[Maxime Santilli @100illi]

What do you hope AI will make possible?
Ask a 7-year-old. Ask a 77-year-old. You'll get two very different answers, and both belong in the record.
Interview someone you love tonight:https://t.co/Fgx9036FAn

[Maxime Santilli @100illi]

Founding pixels on the RingRing Wall 🔔
100 px: $50200 px: $1001,000 px: $5002,000 px: $1,000
The first 100,000 pixels go at $0.50. Then $1.
Leave your company's mark on the AI era: https://t.co/ySInmNDGRa

[Maxime Santilli @100illi]

L'IA, pays par pays sur le Wall de RingRing 🔔
🇺🇸 États-Unis 61🇬🇧 Royaume-Uni 5🇸🇪 Suède 3🇧🇪 Belgique 2🇳🇱 Pays-Bas 1🇨🇭 Suisse 1🇫🇷 France 0
La Belgique devant la France. Startups françaises, vous attendez quoi ?
https://t.co/ySInmNDGRa

[Maxime Santilli @100illi]

Most spotted AI companies on the streets of San Francisco so far 👀
Framer: 4 sightingsRunware: 4Cloudflare, Modal, Stripe, Vanta, Superhuman: 3 each
Every ad I photograph becomes a logo on the Wall.
Who's next? https://t.co/ySInmNDGRa

[Maxime Santilli @100illi]

Monday question 🔔
How does AI make you feel today?
Not what you think about it. How it makes you feel. Excited, tired, scared, curious?
Say it in your own voice, sealed for five years:https://t.co/Fgx9036FAn

[Maxime Santilli @100illi]

@GergelyOrosz The internet is a small town for engineers. The person you dunk on today might be your hiring manager, reviewer or cofounder in 3 years. Generosity in public compounds, and so does being a jerk.