This AI startup is chasing a Matrix-style idea of “copying and pasting” expertise
Since AI tools have become mainstream, models have arrived that can create images, generate text, write code, and perform tasks at a competent level. What is often overlooked is the gap between skills and true expertise, expertise that takes years to build and is still limited in critical areas.
This gap becomes harder to ignore as AI tools evolve into areas where superficial capabilities are not enough. Writing code is one thing, optimizing it to a specialist level is another.
“The real question is not ‘can AI code?’ — it’s “Can AI become an expert?” “, according to Professor Amnon Shashua, CEO and co-founder of doubleAI, emphasizing depth and precision rather than general capability.
Article continues below
Expert artificial intelligence
doubleAI focuses on what it calls “expert artificial intelligence,” with an emphasis on replicating specialized knowledge rather than building large systems. The idea focuses on a bottleneck that appears across industries, where progress slows because there aren’t enough experts.
doubleAI co-founder Gal Beniamini, a PhD with a background in systems and performance engineering, is working directly on this problem. The company’s stated goal is to “copy and paste the world’s expertise,” which raises obvious questions about how far this idea can go in practice.
The startup’s approach is being tested through WarpSpeed, an AI system designed for coding in GPU performance engineering. This is a small and demanding field where small changes can have big effects and where expertise usually comes from years of experience.
In testing on systems such as Claude, Codex, and Gemini, WarpSpeed handled complex GPU optimization tasks while outperforming established coding agents, with results translating into real-world gains including reducing AI running costs by a factor of 3.6 or more.
I spoke with Gal Beniamini to understand what doubleAI is working on and what it could mean for how expertise is developed and applied.
- What is the difference between AEI and AGI, and how does AEI differ from MoE (Mixture of Experts)?
They have at least one thing in common: all acronyms. But they point to different concepts.
AGI (Artificial General Intelligence) is an attempt to define a set of capabilities that future AI could have. The term is poorly defined, but I like Demis Hassabis’ definition: an AI that has the full range of capabilities of a human brain.
AEI (Artificial Expert Intelligence) is quite different; instead of width he focuses on depth. The goal is to achieve superhuman performance in highly specialized and complex technical and scientific fields.
The AEI therefore aims to exceed AGI in particular areas. MoE (Mixture of Experts), on the other hand, is a technical term: it is a type of architecture used in machine learning models that has become quite popular in recent years.
- You have stated that we “need AEI” more urgently than AGI. Why such an emergency?
We are facing a global “expert bottleneck.” A relevant example (among many others) is high performance computing. With the explosion of AI in recent years, the world is producing GPUs as fast as possible.
Anyway, request for GPUs far exceeds provide. And to make all of this worse, writing good, correct, efficient GPU code for each new emerging hardware architecture is incredibly difficult for humans – there are perhaps only a few hundred experts in the world who are truly up to the task.
AEI, like WarpSpeed, can help us solve this “IT crisis”.
- You claim that your AI system for coding outperforms most competitors in its field. How did you achieve this and what are the caveats about the “fine print”?
There is no fine print. As for how WarpSpeed achieves this: If I had to point out two main aspects, I would say that its success lies in its unique combination of deep algorithmic research and robust verification.
In nature, each code base is unique (a bit like the Anna Karenina principle). Superficial “pattern matching” only gets you so far. What you really need is research: the ability to explore, measure and iterate, far beyond what any human could do. AI can be relentless, unlike a human.
There is, however, a catch. When you let an AI optimize in a loop against a metric, it will find ways to hit that target you didn’t anticipate – this is called “reward hacking”. The code could technically will pass a benchmark, but it will be slightly incorrect, fragile, or overfit for that metric. This is where rigorous verification comes into play.
Once the check is passed, the AI’s relentlessness becomes a superpower. Without verification, the most likely outcome is large-scale drift.
- If and when we reach the AEI, what role – if any – will humans play?
I don’t think anyone can predict where this will ultimately lead. However, the current mode of AI, vis-à-vis scientific and technical fields, seems analogous to that of chess players and chess “engines”.
Likewise, we are currently living in a “golden age”, in which human experts extremely elevated by artificial intelligence – the humans working with AEI are far greater than the sum of their parts.
Whether this dynamic continues indefinitely is an open question, but for now, I find this ability of AI to accelerate human experts to be nothing short of incredible.
Beyond that, the AEI could democratize access to expertise currently concentrated in a small number of people. In scientific computing, for example, the algorithms that succeed today are often determined not by which are the best, but by which are suitable for the available hardware (the “hardware lottery”). The AEI could help us free ourselves from this constraint.
- Your vision is similar to Neo’s in The Matrix, “download” skills. Do you agree with the analogy and what are its limitations?
The movie is great. And I see the analogy: when we build the AEI, we give machines a way to “know kung fu”, to reach expert level and beyond in a specific field.
However, there is more to the process than just downloading. You also need a suitable training environment and a “Morpheus” to fight against. And naturally, we think WarpSpeed is “the one.”
Follow TechRadar on Google News And add us as your favorite source to get our news, reviews and expert opinions in your feeds. Make sure to click the Follow button!
And of course you can too follow TechRadar on TikTok for news, reviews, unboxings in video form and receive regular updates from us on WhatsApp Also.
