跳到正文
Hacker News:AI 热帖· nicolas-siplis·· 2026-09-18AI 评分68

Bend 语言发布:用法则与证明拦截 AI 写码错误,单核接近 C 且可跑 GPU

Bend - A language that blocks AI mistakes via proof, on CPU and GPU

AI 导读

Bend 语言正式推出,定位为在 AGI 时代让人类精确约束 AI 写代码的方式,通过 LAWS.bend 声明法则并要求 AI 提交证明,违反法则的代码在数学上无法被合并。

正文 · 原文

In the post-AGI economy, humans will eventually stop writing and reading code, but we still need an ambiguity-free way to tell the AIs building the world around us what we want done.

With laws, our intents can be much more precise than natural language. With proofs, we can verify that the AI implemented our prompts correctly. And a fast compiler runs it at speed.

That's Bend - and nothing else.

1.Bend runs FAST.

Bend compiles to native code. On one core, it runs nearly as fast as C. The same binary also runs on sixteen cores, or on the GPU, running up to a hundred times faster than one core.

媒体内容 · 前往原文查看
Apple M4 Max · lower is better

2.Bend compiles FAST.

Bend's type checker is a proof checker, as in Lean and Rocq. Those can take minutes on a mid-sized codebase. Bend takes a second at most, so an AI agent can check after every change.

媒体内容 · 前往原文查看
Apple M4 Max · lower is better

3.Bend is PARALLEL.

No threads, no locks, no kernels to write. Split the work in two, and Bend spreads the calls over every core it can find, then joins them back. Now watch pow2 run on 4,096 GPU cores:

媒体内容 · 前往原文查看
pow2.bend running on the GPU

4.Bend BLOCKS mistakes - with proof

How can you trust code you never read? By demanding a proof. LAWS.bend is where you declare laws. From then on, no AI can ship one line that breaks them, ever. Watch it guard a game:

媒体内容 · 前往原文查看

Law: winning is impossible

So far, it works!

New feature:

“Claude, make the board wrap around”

媒体内容 · 前往原文查看

Without LAWS.bend:

Laws broken. AI mistake: merged.
媒体内容 · 前往原文查看

With LAWS.bend:

Laws intact. AI mistake: blocked!

Without LAWS.bend, the bug went live. With LAWS.bend, the AI had to retry until it built a wall and proved the law holds. Merging a bug is mathematically impossible: it is a theorem.

LAWS.bend

# LAW: no move sequence leads to victory.
law you_cant_win:
  for moves: List<Move>            # any sequence of moves
  board = replay(start(), moves)   # replayed from the start
  is_won(board) == False{}         # never leads to victory

PROOF.bend

# PROOF: you_cant_win holds.
def Laws.you_cant_win(moves):
  # ... written by the AI

LAWS.bend is AGENTS.md backed by proof. “Make no mistakes” is now type-checked.

Skeptical? Try breaking the game.

5.Get started.

5.1.Install

curl -fsSL https://bend-lang.com/install.sh | sh

5.2.Tell your agent to use Bend

Add this to your AGENTS.md:

When using Bend:
- run `bend guide` to learn it
- use `LAWS.bend` to keep important rules
- run `bend PROOF.bend` before committing
- parallelize the code whenever possible

Then, just say: "use Bend"!

5.3.Enjoy bug-free, fast vibe-coded apps!

Hints: ask it to write laws for whatever should never break, and to parallelize everything you want running fast. Bend is young: if anything goes wrong, ask it to open an issue. Bend works best on the back-end, on Linux and on macOS. Enjoy! <3

6.References.

Guide: GUIDE.md is the whole language; bend guide prints it. Paper: BendTT, an affine dependent type theory, Bend's core. Paper: BendRT, a parallel runtime for CPUs and GPUs, the VM.

Bend is still evolving. Expect bugs, and please report them.

来源:Hacker News:AI 热帖 · bend-lang.com