Max

AI can code. Now it needs to touch the world. Building that layer

[Diana @sdianahu]

robotics is getting close to its GPT era

in conjunction with the dyna-2 results published this week, it showed signs of scaling laws applying to robotics

exciting times!

[Max @mx_builds]

Thank you @brainfeewds @hardwarefyi for the great event yesterday.

Met a ton of cool builders and got to see some really interesting projects.

Always a good reminder of how much is being built out here!

[Max @mx_builds]

Talking to users is so fun, when you know what to ask

[Max @mx_builds]

Yesterday’s interview POV: My cofounder and I when the door opens and it’s Garry Tan interviewing us #yc

[Max @mx_builds]

@ycombinator F26 update: we didn’t make it into this batch.

Genuinely tough news, but we’ve built a lot to be proud of already, and this doesn’t change the plan.

Onward.

[Max @mx_builds]

Yesterday’s interview POV: My cofounder and I when the door opens and it’s Garry Tan interviewing us #yc

[Max @mx_builds]

Good callout, we've only shipped the clustering fix so far, which is cosmetic at best. Haven't tried decay yet but it's the more obvious answer: reuse the same outcome-weighting we already do on policy edges and apply it to brain nodes, so a node that hasn't paid off loses retrieval priority regardless of how close it looks semantically. Going to prototype that!!

[Max @mx_builds]

https://t.co/yquEwFNpHN

[Max @mx_builds]

@lightsilver323 Vercel for hardware

https://t.co/sQBMQgM3yT

[Max @mx_builds]

The most shocking thing about moving to SF is that we met more relevant people in 1 week here than in a month in Paris

[Max @mx_builds]

@GauthierLe4593 Nice capy hat 🧢

[Gauthier Lechevalier @GauthierLe4593]

I am using my own tool (nff), to improve itself (nff for nff). Your first customer should be yourself !

[gsimone @ggsimm]

Easily fix your resumee gaps on github with my crazy intuitive tool

[Max @mx_builds]

Inspired by @t_blom talk on AI-native companies, we built our "company brain" as a policy graph, not just retrieval.
Every action gets a live success rate, cost, and outcome. Built first to run hardware fleets (bad OTA = bricked device), now running our own ops too.

[Max @mx_builds]

The scary part of Anthropic's incident wasn't Claude "escaping." It behaved fine, it just trusted a bad boundary.
As we wire AI into real devices, the hard problem isn't model alignment. It's building harnesses the model can't be wrong about.

https://t.co/C4KUDfNYga

[Max @mx_builds]

@coryalthoff vercel for robots