AI model Maxwell makes existing software faster while it runs

Computing power for AI is expensive, and getting more expensive. According to the research firm SemiAnalysis, the rental price of a widely used Nvidia AI chip on a one-year contract rose by 60 per cent in twelve months. Companies are therefore looking for ways to get more work out of the same chips.

On 19 October, the Tel Aviv company doubleAI came up with its own answer: Maxwell. The company calls it a 'code execution model', a new kind of AI that does not write programs itself, but makes existing programs run faster.

How does it work?

Software is written by people in a programming language. The computer translates that text into instructions for the chip. That translation is general: it has to work on many kinds of computers and with all sorts of data. As a result, speed is often left on the table.

Maxwell watches while a program runs and looks for the parts that take the most time. For those parts it writes a faster version, tuned precisely to the chip in that computer and to the data passing through at that moment. That version is swapped in while the program simply keeps running. The source code, the text the programmer wrote, stays unchanged. You can compare it with a route planner that does not give the standard route, but the fastest route for your car at this time of day.

A faster version must of course do nothing different from the original. According to doubleAI, Maxwell provides a mathematical proof with every change that the new piece of code always gives the same result. If the conditions for which a version was made no longer apply, the program switches back to the original code. Maxwell works both on ordinary processors and on graphics chips, which are used mainly for AI.

The results

doubleAI says Maxwell cut the computing costs of one customer, whose name has not been disclosed, by 67 per cent. According to the company, programs that process large amounts of data became up to 38 times faster, and parts of AI models ran more than four times as fast.

The company LTX, which makes open AI models for video, also tried out Maxwell. Making a five-second video clip in 4K quality on two AI chips then took about 134 seconds instead of 250. According to LTX, the result cannot be told apart from the clip the same software made without Maxwell.

doubleAI does not want to charge a fixed amount, but a share of what customers save. Maxwell is now available to a first group of users, who can sign up.

The next steps

For now, the figures come from doubleAI itself and from a customer. Independent tests have to show how big the gain is for very different kinds of software. It will also become clear how widely the mathematical proofs work once Maxwell runs at more companies.

What does this mean for the road to the singularity?

AI research is held back mainly by computing power: there are not enough chips and they use a lot of electricity. Every per cent that can run faster means more research with the same chips and less energy per answer. What is special about Maxwell is that here an AI improves the software that other AI runs on, while proving that nothing goes wrong. That is a step towards a cycle in which AI keeps improving its own tools. If that kind of self-improvement stays reliable and verifiable, progress in AI could speed up even further.

What is the technological singularity? Read the explanation →