Since AI tools entered the mainstream, models have arrived that can create images, generate text, write code, and complete tasks at a competent level. What often gets overlooked is the gap between competence and true expertise, the kind that takes years to build and is still in limited supply across critical fields.
That gap becomes harder to ignore as AI tools move into areas where surface-level ability isn’t enough. Writing code is one thing, optimizing it at the level of a specialist is something else entirely.
“The real question isn’t ‘can AI code?’ — it’s ‘can AI become an expert?’” according to Prof. Amnon Shashua, CEO and cofounder of doubleAI, putting the focus on depth and precision rather than general capability.
Article continues below
Artificial Expert Intelligence
doubleAI is focused on what it calls ‘Artificial Expert Intelligence’, with an emphasis on replicating specialized knowledge rather than building broad systems. The idea centers on a bottleneck that shows up across industries, where progress slows down because there aren’t enough experts.
doubleAI cofounder Gal Beniamini, a PhD with experience in systems and performance engineering, is working directly on that problem. The company’s stated goal is to “copy and paste expertise into the world,” which raises obvious questions about how far that idea can go in practice.
The startup’s approach is being tested through WarpSpeed, an AI system built for coding in GPU performance engineering. This is a narrow, demanding area where small changes can have large effects, and where expertise usually comes from years of experience.
In testing against systems like Claude, Codex, and Gemini, WarpSpeed handled complex GPU optimization tasks while outperforming established coding agents, with results translating into real-world gains including cutting AI running costs by a factor of 3.6x 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 do point at different concepts.
AGI (Artificial General Intelligence) is an attempt to define a set of capabilities that future AI might have. The term is ill-defined, but I like Demis Hassabis’ definition: AI that possesses the full range of capabilities that a human brain does.
AEI (Artificial Expert Intelligence) is rather different; instead of breadth it focuses on depth. The goal is to achieve superhuman performance in highly specialized, complex technical and scientific domains.
So AEI aims to exceed AGI in particular domains. MoE (Mixture of Experts) on the other hand, is a technical term — it’s a type of architecture used in machine learning models that’s become rather popular in recent years.
- You’ve stated that we “need AEI” more urgently than AGI. Why such an urgency?
We are facing a global “Expert Bottleneck.” One pertinent example (of many) is high-performance computing. With the explosion of AI in recent years, the world is producing GPUs as fast as it can.
Even so, demand for GPUs is far exceeding supply. And to make this all worse, writing performant, correct and efficient GPU code, for each new emerging hardware architecture, is incredibly difficult for humans — there are maybe only a few hundred experts worldwide that are really up to the task.
AEI, like WarpSpeed, can help us solve this “Compute Crisis.”
- You claim your AI system for coding outperforms most competitors in its field. How did you achieve that, and what are the “small print” caveats?
There’s no small-print. As for how WarpSpeed gets there: if I had to point to two core aspects, I’d say its success lies in its unique combination of deep algorithmic search and strong verification.
In the wild, every code-base is unique (much like the Anna Karenina principle). Shallow “pattern-matching” only gets you so far. What you really need is search: the ability to explore, measure, and iterate, far beyond what any human would. AI can be relentless in a way that a human is not.
There is a catch, though. When you let an AI optimise in a loop against a metric, it will find ways to hit that target that you didn’t intend — this is known as “reward hacking”. The code might technically pass a benchmark, but it will be slightly incorrect, brittle, or overfit to that metric. That’s where strong verification comes in.
Once you get verification right, the AI’s relentlessness becomes a superpower. Without verification the more likely outcome is slop at scale.
- If and when we reach AEI, what role — if any — will humans play?
I think no one can predict where this ultimately goes. However, the current mode of AI, vis-a-vis scientific and engineering domains, seems analogous to chess players and chess “engines”.
Similarly, we are currently at a “golden age”, in which human experts are hugely elevated by artificial intelligence — humans working with AEI are far greater than the sum of their parts.
Whether that dynamic holds indefinitely is an open question, but right now I find this ability of AI to accelerate human experts to be nothing short of incredible.
Beyond that, AEI could democratize access to expertise that’s currently concentrated in a tiny number of people. In scientific computing, for example, which algorithms succeed today is often determined not by which are best, but by which happen to suit the available hardware (the “hardware lottery”). AEI could help us break free from that constraint.
- Your vision sounds like Neo in The Matrix, “downloading” skills. Do you agree with the analogy, and what are its limitations?
The film is great. And I do see the analogy — when we build AEI, we’re giving machines a way to “know kung fu,” to become expert-level and beyond in a specific domain.
The process isn’t just a download, however. You also need the right training environment, and you need a “Morpheus” to spar against. And naturally, we believe WarpSpeed is ‘the One.’
Follow TechRadar on Google News and add us as a preferred source to get our expert news, reviews, and opinion in your feeds. Make sure to click the Follow button!
And of course you can also follow TechRadar on TikTok for news, reviews, unboxings in video form, and get regular updates from us on WhatsApp too.
