←
Predictable Machines52 posts
Predictable Machines avatar

Predictable Machines

@AIPredictable

We create products that ensure information and automation are trustworthy, accountable, and reliable. Building @predictablecode

◷ Joined February 2025▧ 22 Photos and videos
Gorka Moreno Asín avatar
⟳ Predictable Machines retweeted
Gorka Moreno Asín@gorkma·Oct 25

Empieza a haber mucho hype con la verificación formal. Esta semana arrancamos con lo que han montado nuestros amigos de @AIPredictable @raulraja y @Jorge__Galindo. ¡Muchas ganas de contar lo que somos capaces de hacer con ello!

形式検証は普通に使われるようになるよ 形式検証ではある性質が数学的に正しいことを確実に証明してくれるから、たとえば、「秘密鍵が出力に漏れない」と証明された場合、その経路での情報漏洩攻撃は不可能になる 特にセキュリティ周りは形式検証で穴を塞ぐことが重要になるだろう 証明した性質に違反する、すべての攻撃を防ぐことができるからだ たとえば、メモリの安全性が証明できれば、その種のメモリ破壊攻撃は全て排除できる 以前までは学習・実装コストが非常に高かったが、エージェントのおかげで、それはほとんどなくなった 逆に、攻撃者側は圧倒的に優位になっている 狙われてしまえば、ほぼ勝ち目はないので、開発者は祈るしかない 形式検証がその状態を変えるはずだ pstack も既に形式検証の仕組みを取り込んでいる /crate-formal-verification-skill・/maintain-formal-verification-skill ・/statechart だ これらは、エージェントが必要な箇所に形式検証を行う Skillを作ってくれるし、状態遷移図であらゆるバグの手順を人間に示してくれる セキュリティに関する部分、ユーザーにとってバグがクリティカルな問題になる部分の手順を全部洗い出して可視化し、それを検証し、修正まで人間の介入なく行えるので、本当に形式検証のコストは非常に低い ここまできて、pstack 使わないなんて選択肢あります?!

Jorge Galindo avatar
⟳ Predictable Machines retweeted
Jorge Galindo@Jorge__Galindo·Aug 12

Maybe it's because I'm obsessed with the topic, but it feels like everyone in AI is suddenly talking about verification 🤓. Not all verification is the same, though. Most of what labs ship today is probabilistic: agents cross-checking each other, one model grading another's work. It scales to fuzzy, human problems, and honestly, it's the only thing that does. But you get confidence levels, never guarantees. At @AIPredictable, we're betting on deterministic verification. Mathematical methods that give you a binary answer: does the code do what the spec says? Yes or no. We generate specs, translate them into Lean, and formally verify every possible theorem about expected behavior. Not just the happy path. It doesn't apply to everything. Ask it a philosophical question, and it's useless. But for code in banking, pharma, or healthcare, where "probably correct" doesn't cut it, the game changes. So, honest question: is probabilistic verification enough for what you're building, or do you need the math?

Predictable Machines avatar
Predictable Machines@AIPredictable·Jul 23

Full story in @JYCabello's new article, production horror story included (a Friday deploy, a database index, and a two-week holiday): predictablemachines.com/blog… Worst bug a green CI ever hid from you?

Predictable Machines avatar
Predictable Machines@AIPredictable·Jul 23

That's why we went deterministic. No AI judging your code: LLMs are trained on decades of code full of the same implicit assumptions we've learned not to see. We translate your logic into a Lean model and let the math emit the verdict. You keep writing in your language.

Predictable Machines avatar
Predictable Machines@AIPredictable·Jul 23

Your compiler trusts you. It shouldn't. C# compiles a recursive loop that will slowly shove your database off a cliff. No warnings, no errors. Lean refuses to compile the exact same logic until you prove it terminates. 🧵

Post attachment
Predictable Machines avatar
Predictable Machines@AIPredictable·May 13

For every deviation between code and spec, three options: → Fix the code (scoped prompt, can't break other contracts) → Update the spec (if your intent was different) → Remove the spec (if it no longer applies)

Predictable Machines avatar
Predictable Machines@AIPredictable·May 13

This is the moment that matters 🤘 Syncs each requirement to the code that satisfies it, generates theorems, and proves them. Every theorem that doesn't hold comes back with the exact input that breaks it.

Predictable Machines avatar
Predictable Machines@AIPredictable·May 13

A local viewer to review every inferred contract before anything else runs. If a requirement looks wrong or too strict, you correct it here first. The formal check will then run against the contract you actually want.

Predictable Machines avatar
Predictable Machines@AIPredictable·May 13

Step 1: @PredictableCode reads the codebase and infers the behavioral contracts that the code itself implies it should satisfy. No specs written by hand. No annotations. Just the code and documentation, and what it formally promises to do.

Jorge Galindo avatar
⟳ Predictable Machines retweeted
Jorge Galindo@Jorge__Galindo·May 7

The last article of the #AICodePainPoints series is done! (For now, because every improvement in this field brings new issues with it.) This time, it's about what happens when a developer clicks Merge on AI-generated code reviewed by another AI. What are they actually approving? And in regulated industries, who carries the weight when something breaks? The accountability didn't disappear. It just concentrated in one click.

AI writes the code. AI reviews the code. A developer clicks "merge" in under 10 minutes. In that single click, all the legal and regulatory liability concentrates in one human, for code they didn't write and couldn't fully read.

Post attachment