doubleAI's Maxwell rewrites code while it runs, claims 38x speedups

doubleAI, a Tel Aviv-based startup founded in 2023, on Monday introduced Maxwell, a code execution model (CEM) that evaluates, optimizes, and rewrites binary code while a program is running. The company claims the model has made data pipelines more than 38x faster and AI inference kernels more than 4.5x faster, with overhead capped at 2 percent. That headline number, 38x, is the kind of gain that makes engineering leaders sit up, but the real story is how Maxwell gets there: not by tweaking a compiler's output, but by reasoning about the algorithm itself.

Maxwell is named for Maxwell's Demon, the 1867 thought experiment by physicist James Clerk Maxwell about a creature that sorts particles to reverse entropy. In the same spirit, the model preprocesses algorithms and machine instructions to maximize performance on whatever hardware the program lands on. Co-founder Gal Beniamini, who holds a PhD in theoretical computer science, told The Register that Maxwell works like a just-in-time (JIT) compiler but with a critical difference: 'A compiler can't do high-level reasoning. It can't, for example, change the algorithm. Maxwell can and does.'

The distinction matters. Traditional compilers make decisions before a program runs and must be narrow and conservative to guarantee correctness. Maxwell, by contrast, operates continuously, seeing the exact hardware, the actual shapes and types of data flowing through it, and the state of the environment. 'This allows it to specialize in ways a compiler never can,' Beniamini said. Because it can swap out a general-purpose sorting algorithm for a different one mid-execution, it can chase bottlenecks as they emerge.

The performance claims are striking. Beyond the 38x data pipeline figure, doubleAI says Maxwell made AI inference kernels more than 4.5x faster and helped LTX-2.5, an open world model that generates 4K video, run almost twice as fast. The optimizations reduced data exchanged between GPUs and dropped unnecessary token padding. Beniamini noted that Maxwell does best on algorithmic code that can be tailored to hardware in non-trivial ways, but even high-level Python and PyTorch graphs typically have 'far reaching optimisations' available.

Formal verification is the safety net that makes this possible. Since Maxwell rewrites code on the fly, it must prove that its optimizations are sound. Beniamini said the model can also eliminate issues like memory errors by proving their non-existence. 'Anecdotally, we have found such issues in Python itself, through the use of Maxwell, i.e., cases in which Python was not memory safe, and where Maxwell proved that to be the case,' he said. That security angle could appeal to enterprises wary of AI-generated code.

Deployment is deliberately light-touch. Developers install the runtime and run their code, with the Maxwell model hosted by doubleAI. That arrangement requires trust, since the code is being rewritten by a cloud-hosted AI. But Beniamini said the overhead is minimal: 'The maximal overhead of Maxwell's tracing and profiling is 2 percent (when it isn't tracing, that falls to zero, of course), and the optimisations that Maxwell yields far outweigh that.' He added that doubleAI applies Maxwell to its own runtime, so it is 'never slower.'

The strategic framing comes from Amnon Shashua, co-founder and CEO of doubleAI, founder and departing CEO of Mobileye, and a professor of computer science at the Hebrew University of Jerusalem. 'CEMs, and Maxwell in particular, are agents on behalf of the hardware as opposed to code agents that are on behalf of the human user,' he said. 'In other words, they are the first AI that works for the machine, not the programmer.' That inversion, AI serving the runtime rather than the developer, could reshape how companies think about performance engineering.

For engineering leaders, the implication is clear: the next frontier in optimization may not be better compilers or hand-tuned kernels, but AI that reasons about code at runtime. The formal verification piece addresses the trust gap that has kept AI-generated code out of critical systems. And with a founder like Shashua, who built Mobileye into a driverless-car powerhouse, doubleAI has credibility to push this into production environments. The question is whether enterprises will hand their binaries to a cloud-hosted model, and how much speed they are willing to trade for that trust.