More than 38 times faster: that’s what doubleAI says Maxwell did to real-world data pipelines. And the startup calls Maxwell the first Code Execution Model, or CEM. Introduced Oct 19, 2026 and described by The Register as a system that “evaluates, optimizes. And rewrites binary code during execution,” Maxwell lives inside a running program, watches the exact hardware it landed on plus the data flowing through it. And swaps in new machine code while the program keeps running. doubleAI, a Tel Aviv startup founded in 2023, formally calls the model Maxwell’s Daemon and says it works on both CPUs and GPUs.
Your source file stays untouched the whole time.
The speedup number is the least interesting part of that pitch. It’s company-reported, single-source, and untouched by independent testing so far. The interesting part is the category flip: every AI tool you’ve used this year writes code for you. Maxwell works for the machine.
What a Code Execution Model Actually Does
A Code Execution Model is not a chatbot that emits source from a prompt.
Per doubleAI’s own description, a CEM “lives inside the running program, watching the data flow through it and reshaping the algorithms and instructions as it goes.” Its inputs are the running program, the exact hardware it has landed on. And the actual shapes and types of the data moving through it. That last input is the whole trick, because as The Register notes, that information is only available at runtime.
The output is a binary that runs directly on the hardware.
And the source, as The Next Web put it, “stays exactly as it was.” Hold that shape in your head, as it inverts how you expect AI to touch software.
In doubleAI’s framing, coding agents write the program and Maxwell makes sure the machine runs it optimally.
The two are complements, not competitors.
My read: the source file stops being the product and becomes a spec.
The binary becomes negotiable, rewritten per machine, per dataset, per run.
Compiler engineers spent decades making one binary run acceptably everywhere. This trades that portability for fit, and the code you ship becomes a suggestion the runtime finalizes.
It’s a JIT Compiler That Proves Its Own Work
The Register’s comparison is the fastest way to understand the mechanism: Maxwell is similar in concept to a just-in-time compiler.
But with more flexibility and formal verification that replacement code is equivalent.
JITs have recompiled hot paths at runtime for decades inside JVMs and browsers. The difference here is the guarantee, since doubleAI says Maxwell formally proves each rewrite leaves the program’s behavior unchanged.
That proof is the load-bearing wall of this entire category. A system that rewrites live machine code under a running program only ships if a machine checks every rewrite first. Formal verification is what turns a terrifying idea into an engineering idea. And I suspect it becomes the standard safety story for anything AI-flavored that touches production binaries.
Now apply that same discipline to the numbers. The data-pipeline claim of more than 38 times faster and the inference-kernel claim of more than 4.5 times faster both come from doubleAI, relayed by The Next Web, on real-world workloads as the company defines them. I treat vendor benchmarks as marketing until a third party runs them. The one testable artifact is doubleGraph, released on GitHub as a drop-in replacement for NVIDIA cuGraph with no code changes required, per doubleAI’s research page. That’s where the claims go to get checked. And if you run graph workloads you can settle this yourself in an afternoon.
The Install Is Light. The Questions Aren’t.
The Register describes installation as light-touch: developers install the runtime and run their code, with the Maxwell model hosted by doubleAI. Read the second half of that sentence twice. The optimizer is a hosted service, which means the thing watching your program’s data flow is doubleAI’s model, not something running on your own iron.
If you’re a solo operator pushing customer data through pipelines, that’s a question you answer before install: what leaves your box.
And what does the provider see?
Then there’s the missing line item.
The announcement material I read carried no pricing for Maxwell. A tool whose entire value proposition is cutting your compute bill will carry a price of its own. And until that’s public, “free speedup” is a placeholder, not a fact.
The economics are still why this category matters to small shops. doubleAI claims more than 4.5 times faster AI inference kernels on real-world workloads, across CPUs and GPUs.
If that survives independent testing, the cheapest GPU in your stack becomes the one you already own. And the upgrade decision turns into a runtime decision. I run automations that eat inference budget every month. And an effective multiplier on existing hardware beats a capital request for new hardware every single time.
The Name Is the Thesis
Maxwell’s Daemon is named for Maxwell’s Demon, the creature physicist and mathematician James Clerk Maxwell imagined in 1867 as a thought experiment on entropy. The demon sorts fast molecules from slow ones and appears to beat the second law of thermodynamics. The catch in the original paradox is that watching and remembering costs energy of its own. And an optimizer that burns compute to save compute sits inside the same joke.
The name is doing real work anyway.
This daemon sorts hot instructions from cold ones, per machine and per dataset, and hands you the difference. Watch the overhead question as this category matures, given that a runtime optimizer that costs as much as it saves is just a JIT with extra steps.
Final Thoughts
The verdict, plainly: the speedup numbers are vendor-reported and unproven, and I’d hold your skepticism until third-party benchmarks exist. Design: the actual news, as a model acting on the machine’s side of the compile, with formal equivalence proofs as its safety story, is complementary to the coding agents you already pay for rather than a replacement for them. This is your cost curve if you ship data pipelines or carry an inference bill. File it and move on if you ship CRUD apps.
Your move is concrete:
– Pull doubleGraph from GitHub and benchmark it against cuGraph on your own workload before believing any number in this post, mine included.
– Ask the hosted-model question in writing before you install any runtime optimizer: what data leaves your infrastructure?
– Track whether independent benchmarks of the 38x and 4.5x claims ever appear, since their absence tells you as much as their presence would.
The next time a vendor slide says 38 times faster, ask who measured it. I ask that question for a living, and it has never once bothered anyone whose numbers were real.
