Maxwell gives sub-optimal code a runtime glow-up

What happened

On Monday, doubleAI, a startup based in Tel Aviv, Israel, introduced Maxwell. Maxwell is a "code execution model" that checks, improves, and rewrites machine code while a program runs. It does not change the original source code. Instead, it changes the version of the code the computer is actually running.

Key numbers

• Data pipelines ran more than 38 times faster, according to doubleAI.
• AI inference kernels (small, core parts of AI programs) ran more than 4.5 times faster.
• LTX-2.5, an open world model that makes 4K video, ran almost twice as fast with Maxwell's help.
• Maxwell's results depend on the code and the hardware it runs on.

How it differs from a compiler

A compiler turns code into a form the computer can run, and it does this before the program starts. A just-in-time (JIT) compiler does some of this work while the program runs. Compilers must be careful and make only small, safe changes. Maxwell is more flexible. It can see the exact hardware and the real data as the program runs. It can also replace a whole algorithm, such as using a different sorting method. To keep changes safe, it uses formal verification, which means it writes mathematical proofs that the new code does the same job correctly.

Why it matters

Faster software saves computing power, time, and money. If Maxwell works well, programs could improve on their own as conditions change. The tool also shows a new idea: AI that changes code while it runs, with proof that the changes are correct. The article notes that Maxwell works best on code that is algorithmic and can be tailored to specific hardware.

What to watch

• Whether other companies test Maxwell on real, large programs.
• How well the proofs hold up on many kinds of code and hardware.
• Whether the speed gains stay large outside doubleAI's own examples.