The Register reports that doubleAI, a Tel Aviv startup, has introduced Maxwell, a code execution model (CEM) that evaluates, optimizes, and rewrites binary code during execution. Unlike a just-in-time (JIT) compiler, Maxwell can perform high-level reasoning and swap out entire algorithms, while formally verifying that the replacement code is equivalent. The source code itself is not touched.
According to co-founder Gal Beniamini, Maxwell sees the exact hardware, data shapes, and runtime environment, allowing it to specialize in ways a compiler cannot. It operates continuously, so it can keep improving programs as new bottlenecks appear. The formal proof capability is essential, because rewriting code on the fly without proof would risk introducing errors.
DoubleAI claims Maxwell has made data pipelines more than 38x faster, AI inference kernels more than 4.5x faster, and the open world model LTX-2.5 generate 4K video almost twice as fast. The company also says Maxwell can prove the absence of memory errors, and has found such issues in Python itself. Tracing and profiling overhead is at most 2 percent, dropping to zero when not tracing.
Installation is light-touch: developers install the runtime and run their code, with the Maxwell model hosted by doubleAI, which requires trust. CEO Amnon Shashua described CEMs as "agents on behalf of the hardware," the first AI that works for the machine rather than the programmer.