Tech
doubleAI’s Maxwell promises faster software without changing a line of code
Mobileye founder Amnon Shashua’s startup launched a model that tunes programs as they run. Its 38x claims are the company’s own, and early access has just begun.
Most of the AI coding boom is about writing software faster. A startup from one of Israel’s best-known tech founders is betting on the other end of the problem: making software that already exists run faster, without touching a line of it.
doubleAI, co-founded by Mobileye co-founder Amnon Shashua, launched Maxwell on Oct. 19, a model it calls a “code execution model,” TokenPost reported. The company says it works on machine instructions, the commands a computer actually runs beneath a program’s source code, and improves performance while programs run, without changing the source. It’s available in early access.
“AI has changed how software is written, but not how it runs,” Shashua said, according to TokenPost.
I read through the launch coverage and the funding reports. The pitch is big. The proof so far is thin, and the company’s own numbers deserve the usual caution.
What it claims to do
Maxwell sits between code and hardware. Rather than suggesting edits for a developer to review, it adjusts what the machine executes as the program runs, and doubleAI says it mathematically verifies those adjustments, according to TokenPost. That verification step is the important part. A tool that silently rewrites what a computer does needs to prove it hasn’t changed the result.
The company’s own evaluations claim more than 38 times faster data pipelines and 4.5 times faster AI inference kernels, the routines that run AI models, TokenPost reported. doubleAI also says one early customer cut compute costs by 67%, Runtime Wire reported, though that customer wasn’t named. A company called LTX told The Wall Street Journal that Maxwell doubled its model-inference speed, according to Runtime Wire.
Who’s behind it
Shashua is doubleAI’s co-founder and CEO, and Shai Shalev-Shwartz is co-founder and chief technology officer, according to Lightspeed’s company profile. Shashua is best known for Mobileye, the driver-assistance company, and has stepped down as its CEO to focus on doubleAI, Globes reported. In an earlier trial on Nvidia software, the company reported an average speed improvement of 3.5 times and said it found bugs in the original code, according to Globes, which also reported, citing sources, that progress has been slower than Shashua predicted.
The startup has raised serious money. A $200 million Series A was co-led by Lightspeed Venture Partners and Bessemer Venture Partners, with investors including Nvidia, Dell Technologies Capital, Greenfield, Pitango and BRM, TokenPost reported. Israel’s Globes reported that Nvidia joined a $200 million round at a valuation of about $1 billion, and that the company had raised $221 million to date. FourWeekMBA reported that doubleAI spent about half of that on Nvidia B300 chips.
Why the idea matters
Computing is expensive, and AI has made it more so. Companies are spending fortunes on chips and data centers, and much of the software running on that hardware wasn’t written to squeeze every cycle out of it. A tool that could make existing workloads run several times faster, safely and automatically, would be worth a lot. Every percentage point of efficiency is money not spent on another rack of GPUs.
That’s the business case Maxwell is built on, and it’s a good one. It also explains why Nvidia, which sells the chips, would invest. Cheaper AI, my guess is, simply means more AI gets run.
Read the numbers carefully
The 38x and 4.5x figures are doubleAI’s own evaluations, not independent benchmarks. The 67% cost cut is one unnamed customer. Gains that large may depend heavily on how well-tuned the original code was, and results on already optimized software could be much smaller.
Runtime Wire put it plainly: the results “will need to hold across customers’ workloads.” The company hasn’t published pricing or a timeline for general availability, according to the coverage.
From the engineer’s desk
My read: for an engineering team, the appeal is obvious. No rewrite, no migration, just faster code. The worry is equally obvious. Letting a model change what runs in production, even with mathematical checks, means trusting a system you didn’t write to touch systems you’re responsible for. That trust gets earned in production, over time, not in a launch announcement.
What we don’t know
The coverage I read doesn’t explain how Maxwell’s verification works in detail, what it costs, which hardware and languages it supports, or how many customers are testing it. doubleAI’s headquarters also wasn’t stated in the reports.
Shashua has built a large company once before. Maxwell is a bet that the next fortune in AI isn’t in writing code. It’s in making code cheaper to run.
Sources
- TokenPost: doubleAI opens early access to Maxwell
- Runtime Wire: doubleAI launches Maxwell, citing a 67% cut in one customer’s compute costs
- BigDATAwire: doubleAI unveils code execution models
- Globes: Nvidia joins $200m round raised by Shashua’s AI startup
- Globes: Amnon Shashua doubles down on doubleAI
- Lightspeed: doubleAI company profile
- FourWeekMBA: Nvidia-backed doubleAI spent about half its $220M on B300s
