Reality Strikes Back: Why the Praised AI Scores 0% in a New Benchmark 🔨

Just recently, I read Jack Clark's essay about the imminent singularity and SWE-Bench being torn to shreds, featuring impressive numbers: top AIs hit 93.9% on SWE-Bench, almost autonomously closing real GitHub issues. It seemed like it was time to delete the IDE and go learn to be a barista.

But then folks from Stanford and Harvard released a new benchmark — ProgramBench. And it publicly humiliated all existing frontier models.

Spoiler: the result for GPT-5.4, Claude Opus 4.7, and Gemini 3.1 Pro is exactly 0%.

What's the gist?
The agent is given a compiled binary (from something like jq to monsters like FFmpeg or SQLite) and documentation.
The task is to write a codebase from scratch that 100% reproduces the behavior of the original.

The real kicker is how the authors had to set up the sandbox. Initially, when network access was open, the neural networks, instead of "writing code," simply parsed --help, found the needed repository on GitHub, and did a git clone. When the network was cut off, agents tried to download sources via package managers or just wrote bash wrappers around the original binary.

To stop this clownery, containers were completely isolated, and binaries were given permissions 111 (execute only). No reading. Pure black-box reverse engineering: feed inputs, watch outputs, write an implementation in any Turing-complete language.

And then an astonishing thing came to light. It's one thing to feed a neural network a ready-made repository with a well-structured architecture, where it only needs to write a 50-line patch. It's quite another to make it design a system.

As soon as AI loses its "crutches" in the form of ready-made abstractions invented by meat engineers, it falls apart. It turns out that models can't handle system design, decomposition, and interface construction. They physically cannot hold an architecture from scratch in their "head" if not given rigid constraints. And this is despite models burning up to $5000 per run without hitting context limits.

In short, meat bags still have a fight left. 💪