A new artificial intelligence model can rewrite software while it's running, delivering performance gains of more than 38 times faster for data pipelines and over 4.5 times faster for AI inference tasks, according to DoubleAI, a Tel Aviv-based startup that launched the technology Monday. The company's Maxwell system—a code execution model—evaluates, optimizes, and replaces binary code during program execution without touching the original source code. Named after a thought experiment by physicist James Clerk Maxwell, the model preprocesses algorithms and machine instructions to squeeze maximum performance from whatever hardware is available.
Maxwell improved the speed of LTX-2.5, an open world model that creates 4K video, by nearly double through optimizations that reduced data transfer between GPUs and streamlined kernel selection to eliminate unnecessary token padding. The system's tracing and profiling create a maximum overhead of just 2 percent when active, falling to zero when not in use, while the performance improvements it generates far exceed that cost. DoubleAI also applies Maxwell to the runtime environment itself, producing gains larger than 2 percent across all operations, ensuring programs never run slower.
Gal Beniamini, doubleAI co-founder and PhD holder in theoretical computer science, explained that Maxwell differs fundamentally from just-in-time compilers because "a compiler can't do high-level reasoning. It can't, for example, change the algorithm. Maxwell can and does." Unlike JIT compilers, which make deterministic changes that must be correct by construction, Maxwell uses frontier AI's expressive capabilities combined with formal verification to prove its optimizations are sound, expanding the scope of allowed optimizations far beyond any traditional algorithm or compiler. Beniamini noted that Maxwell works with information unavailable before runtime—the exact hardware, actual data shapes and types, environmental state, and real bottlenecks—allowing it to specialize in ways compilers never can.
The technology works best on algorithmic code that can be customized for hardware in non-trivial ways, but even high-level Python code can see substantial improvements due to what Beniamini calls "available overhang," with PyTorch graphs remaining far from optimal even under Torch's own compiler. Because Maxwell operates continuously rather than in a single pass, it keeps refining programs as new performance bottlenecks emerge during execution. The formal verification isn't optional—it's essential to prevent the model from introducing errors on the fly, and Beniamini said Maxwell can even eliminate memory errors by proving their non-existence as part of the formalization process, noting the team has found memory safety issues in Python itself this way. Developers install the runtime and execute their code normally, with the Maxwell model hosted remotely by doubleAI, though this arrangement requires trusting the company with code visibility. Co-founder and CEO Amnon Shashua, who also founded Mobileye and teaches computer science at Hebrew University of Jerusalem, said Maxwell represents "agents on behalf of the hardware as opposed to code agents that are on behalf of the human user," calling it "the first AI that works for the machine, not the programmer." The shift from compile-time to runtime optimization could reshape how organizations think about performance tuning, particularly as the boundary between writing code and executing it continues to blur.