doubleAI Maxwell rewrites running code. Profile before you pay.

Every AI coding tool so far changes how code gets written. doubleAI Maxwell goes after how it runs. On October 19, the Tel Aviv company co-founded by Mobileye's Amnon Shashua opened early access to what it calls the first code execution model: an AI that rewrites a program's machine code while it runs, leaves the source alone, and bills you a cut of whatever compute it saves.

What actually happened

Per doubleAI's announcement on BusinessWire and The Next Web's coverage, Maxwell watches a program on real hardware with real data, then swaps in faster low-level code — SIMD routines on CPUs, kernels tuned to the specific GPU on GPUs. It targets Python first: data pipelines and PyTorch workloads.

The interesting part is the safety model. doubleAI says every rewrite is formally proven to behave the same as the original, not just tested. When a faster version only holds under a condition (say, a certain data shape), a guard sends execution back to the original code when the condition breaks.

The numbers on Maxwell's site: 38x faster data pipelines, 4.25x faster AI inference kernels, Whisper transcription about 2.1x faster on an RTX 4090. The site itself says these are specific experiments, not guarantees. doubleAI also says one early customer cut compute costs 67%, and video-model company LTX told the WSJ Maxwell doubled its inference speed.

Pricing is a share of your compute savings. The percentage is not public. Access is a request form, and acceptance is not guaranteed.

Why a code execution model matters for your business

GPU rent is going the wrong way. TNW cites SemiAnalysis data showing one-year H100 contract prices up 60% in a year. If your AI feature runs on rented GPUs, speed is margin.

But "we take a cut of what we save" is only a good deal if you know the baseline. Before you talk to doubleAI — or any savings-share vendor — do three things:

1. Profile first. Most slow Python we see is slow for boring reasons: N+1 queries, row-by-row pandas, no batching. Fix those yourself and keep 100% of the savings.
2. Pin the baseline in writing. Savings-share contracts live and die on how "before" is measured. Agree on workload, hardware and metric up front.
3. Keep the guard in your monitoring. A rewrite that falls back to original code is safe, but slow again. Track fallback rates like you track errors.

The binary-level rewrite is a real idea. Treat it as the last 2x, not the first.

Sources: BusinessWire, The Next Web, Maxwell.