Sasha

Making software that turns chaos into systems.
🧠 : programming language for AI
📊 : affiliate ops
👁 : AppStore audits

[Sasha @evidype]

@Ronald_vanLoon I wonder how many insects and bird droppings it's going to collect on the way to the client 🤔

[Sasha @evidype]

This is a beautiful idea, but it is not how America actually works for many people.
I have three friends who tried to get into the United States. All three are exceptional software developers. Two of them worked in applied neuroscience. These are exactly the kind of people any country should be desperate to attract.
They were refused multiple times. So apparently the "blank slate" is not for everyone. The promise exists, but... the gate is narrow.

The other thing I find puzzling is the patriotism itself. To me, the brightest minds are usually above nations and tribes. They are not thinking "my country is better than yours." They are thinking about humanity, civilization, intelligence, and the next centuries. Seeing national exceptionalism from someone working at the frontier of AI, like you, François, is upsetting. I would expect exactly such people to be far beyond flag-based thinking. But we all humans, I understand. Sometimes we fall for "primitive-but-feels-good" things.

Please don't mistake this for envy. It isn't. Every country has its upsides and downsides, at the end of the day, people in every country carry roughly the same amount of joy and misery 🤷

[Sasha @evidype]

@ID_AA_Carmack Code is still the odd generated medium: image/audio/video models emit the final artifact, while code models emit sources. Inconsistency detected!

[Sasha @evidype]

@andimarafioti @cerebras Great, now the model can interrupt me before I've even finished being wrong

[Sasha @evidype]

@ID_AA_Carmack For ui work, this lesson never dies: total work matters but the shape and order of the work matter more, once caches, batches, and browsers enter the chat

[Sasha @evidype]

@ben_sage It can refactor only if you create a specific programming language for AI. And you have to make AI refactor on every iteration. It cannot refactor if the programming language is made for human, at least in my experience.

[Sasha @evidype]

@cheaf25master .com only. Build on https://t.co/6ZzV4j4gqQ and people will still search for https://t.co/6Yno0G4hH7. If you don’t own the .com, you don’t own the brand.

[Sasha @evidype]

@benja_maker Hello Benjamin, let's connect! Currently creating a TS dialect to make fast AI coding feel less like adopting a raccoon with npm access!

[Sasha @evidype]

@fchollet This is why good debugging feels less like fixing, but more like cartography

[Sasha @evidype]

Love the premise: not "use your phone better," but "put it down and come back with proof you lived"

I put Venture through First Impré. Search preview found something useful: the mood survives the App Store blur, but the core promise gets quiet fast

Screenshot of 1 section below, I'll unlock the full report + a free year if you like

[Sasha @evidype]

I used Evidype to build First Impré.

It analyzes an iOS App Store listing before the user ever installs the app.
Live here: https://t.co/bAuCf3hr5j
I picked this project because it stresses the parts of AI coding that usually get messy:

image inputs, generated images, cascaded AI calls with tool use, structured analysis, async UI, report rendering, and a lot of small user behaviors.

Evidype's compiler EvidyTS runs behavioral tests with a hidden Chromium browser, so the system checks what a human can actually do in the app.

[Sasha @evidype]

AI progress is easy to miss month by month.

Then you compare a few years and it hits you.

@midjourney progress

[Sasha @evidype]

In 2003 I built a programming language dialect for humans: https://t.co/RBJcBZlpIq

The goal was simple:

make programmers write less.

Less boilerplate. Less repetition. More convenience.
Twenty-three years later I built another programming language.
It does the opposite.
It forces the author to write more.
Much more.

Humans would hate it.

AI doesn't.

▻ AI never gets tired of being explicit. ▻ AI never complains about verbosity. ▻ AI can happily write every type, every constraint, every detail humans usually omit.

So I spent the last six months building a compiler and language around a strange idea:

What if programming languages should be optimized for whoever writes most of the code?
In 2003 that was the human.
In 2026 it's increasingly the AI agent.

That's the idea behind Evidype.

[Sasha @evidype]

@SergioPaniego Nice to see this live inside TRL rather than requiring a separate serving stack just to make training throughput less tragic

[Sasha @evidype]

@jbhuang0604 The recurring lesson seems to be: remove human labels, then spend the next decade discovering all the implicit contracts you accidentally created

[Sasha @evidype]

@midjourney Finally, doctors who previously specialized in giving everyone six fingers will move into hands-on care!

[Sasha @evidype]

@ID_AA_Carmack @AlexiGlad @ninaddaithankar We finally removed the inductive bias and replaced it with several million hours of whatever the camera happened to look at! But seriously, elegant idea: let time provide the supervision, then see how much architecture can be demoted from theory to habit

[Sasha @evidype]

@le20hfrancetele I don't speak French, but I watched this three times hoping my brain would spontaneously develop a JEPA and infer the French. Apparently three epochs is underfitting

[Sasha @evidype]

@liquidai Tiny model, serious latency, and ColBERT in the mix. The "bigger model fixes everything" era continues to look negotiable.

[Sasha @evidype]

@fchollet Finally, medical imaging for people who think MRIs are held back mainly by a lack of vibes