Robin

[Robin @rspcunningham]

spaceX is extremely undervalued

[Garry Tan @garrytan]

Sanders and AOC introduced a bill to pause ALL AI data center construction. 300+ local bills filed. Half of planned 2026 data centers facing delays or cancellation. Each one brings billions to local economies.

The people who say they want American jobs are trying to block the biggest job creation engine since the interstate highway system.

[Beff (e/acc) @beffjezos]

Not like this... This is not the way.

Not a good precedent to set.

Will lead to govt biasing the models to reinforce their views within the population.

This is why decentralization of models is of.utmost importance.

Freedom of synthetic thought should be protected.

[Robin @rspcunningham]

@sammymanss yo i kinda love this

[Robin @rspcunningham]

@sammymanss fire

[Robin @rspcunningham]

@thatsKAIZEN hell yeah

[Robin @rspcunningham]

@shaunmmaguire yup

[Robin @rspcunningham]

What struck me when revisiting this recently is that this pattern of scene --> physics-based model --> observations powers a lot of modern robotics perception!

In Fourier ptychography, we train 'object' and 'pupil function' tensors; in 3DGS, for example, we train a set of gaussians; in SLAM, we train a 3D map. Makes me excited to be diving into more robotics these days, and seeing so much good work in this space here on X.

Massive shoutout to @sammymanss who worked through this all with me back in the day

[Robin @rspcunningham]

I was chatting with a friend today about my capstone project back in undergrad, and I was reminded of how cool it was, and struck by how relevant some parts are to today's world of robotic perception, so i figured I would walk through it here

The basic premise was about something called Fourier ptychography -- essentially: getting higher-than-should-be-possible-resolution images out of standard microscopes. Here's how it works:

[Robin @rspcunningham]

Given real observations, we backprop to 'train' the parameters of the latent scene. then, we can inference that learned scene through the forward model to get a synthetic observation grounded in real data.

By simply setting the resolution of that synthetic observation to be higher, we get a much better quality image of the sample.

[Robin @rspcunningham]

@yacineMTB yup yup yup

[Robin @rspcunningham]

The autoresearch pattern is really fun to use for a lot of different things!

I’ve been doing this a lot lately; so far one of my favourite tasks has been optimizing a quantum statevector simulator. opus 4.6 was able to write some pretty effective metal kernels and beat other SOTA simulators on my mac

[tobi lutke @tobi]

Autoresearch isn’t just for training models https://t.co/0Gu8m37Onc

[Robin @rspcunningham]

@yacineMTB i remember about 4 months ago thinking you were crazy, then giving myself a day or two to try it out, and realized how absurdly correct you were. never looked back

[will brown @willccbb]

anthropic: using claude code to build a claude wrapper technically counts as distillation and breaks ToS if we feel like it

openai: here’s how you can use codex to do recursive self improvement research

[Robin @rspcunningham]

@fortworthchris @AustinTunnell We also design office layouts around them! for example, in my friend's building, executive offices are all on the 8th floor, because that's how high firetruck ladders can reach

[Robin @rspcunningham]

something interesting: rf devices sold in america have to file with the FCC, and a lot of those filings are public

that means you can pull up design schematics and photos of product internals, which can be super helpful for reverse engineering things

[Robin @rspcunningham]

today is day 0 of getting claude code to help me to hack into the rf control of my chinese drone

usually i vastly prefer codex, but it absolutely refused to help me with this. super safety-conscious claude had zero issues

[Robin @rspcunningham]

@sammymanss if everyone is redacted how does this help :(