At a panel earlier this month hosted by the European Commission, Mistral’s Audrey Herblin-Stoop, Peter Sarlin from Silo AI and a relatively unknown tech entrepreneur — Karim Nouira, the cofounder of Sics — sat down to discuss the future of European tech.
Unlike Mistral or Silo, Sics (short for Superintelligence computing systems) has neither raised massive amounts of capital nor anticipates a lucrative exit any time soon, but it has just secured nearly €4m in a mix of grants and equity from the European Innovation Council (EIC) and a group of Swedish angel investors.
So what was a guy like Nouira doing next to AI hotshots talking about the future of European AI?
Sics was one of 71 startups selected by the EIC Accelerator last year — awarded equity and grants out of 1,200 applicants — and was later invited to the panel. The company also says it may be the first to develop artificial general intelligence (AGI).
If a European company were the first to make such a breakthrough, it would completely reshape the AI map.
The race to AGI
AGI refers to the hypothetical intelligence of a machine that possesses the ability to understand or learn any intellectual task that a human being can. Companies and backers of large language models (LLMs) such as ChatGPT, Mistral AI and Llama have claimed to be on the cusp of solving the AGI conundrum. Others, like Meta’s chief AI scientist Yann LeCun, argue that true reasoning capabilities will never be reached through LLMs.
According to Nouira, GenAI is more about mimicking and using large quantities of data to “reason” than about understanding and learning things independently.
“Our AI is not a language model and does not use reinforcement learning, but learns through interaction and feedback. We are not copying American technology but developing a new technology that can reason and disrupt AI development with the goal of being the first in the world to reach AGI,” Nouira says.
Nouira compares it to human learning and says the missing link in today’s AI is that it doesn’t have a structured “world map”.
“It is like a big black box with billions of unmarked knobs. A supercomputer sets these knobs through a long trial-and-error process that can take months. This is called reinforcement learning. It is like waiting long enough for a monkey with a typewriter to write Shakespeare’s plays,” he tells Sifted.
Sics is developing next-generation AI, which they call an “AGI foundation model for robots”, with the goal of reaching AGI or superintelligence within 18–24 months.
It does this by developing a robot brain that can control all types of robots, from industrial machines to humanoids.
“The path to that goal is to solve the hardest problem in AI — to develop an AI that interacts with the real world and controls robots and solves problems, not just looks at text and images and recombines human input,” Nouira adds.
To develop more advanced AI, Sics is building a model that can create a model of the world from scratch, similar to how a child learns about the world around them.
“The best path to AGI is to emulate a child, with embodiment and real-world interaction. The AI starts by solving simple tasks, such as moving objects, which is difficult for AI but easy for humans. Once the AI understands basic physics and relationships between objects, it learns more complex relationships,” Nouira says.
A robot brain to rule all robots
To test its digital brain, it has been put in control of a pick-and-place robot at a warehouse run by Nowaste Logistics in southern Sweden, where it picks orders for e-commerce clients. Within a “chaotic warehouse environment,” the pick rate is around 2,000 picks per hour with nearly 100% accuracy, according to Nouira.
“While most pick-and-place robots can achieve similar outcomes with tedious reinforcement learning, our digital brain operates autonomously in a human-like manner and learns new objects quickly on-site in real time,” he says.
“One robot could replace up to 16 warehouse workers in continuous 24/7 operations, offering significant cost savings.”
Speaking about Sics to a deeptech investor, who spoke to Sifted on the condition of anonymity, says the startup isn’t alone in this space, but it would be seen in a different light if it appeared in another context.
The company’s leaders don’t look like run-of-the-mill AI startup founders. They aren’t young, nor outwardly naive in the way most young entrepreneurs tend to be.
Instead, both core founders share a curiosity about what can be done but also have plenty of exits behind them. Nouira is a serial entrepreneur, as is his cofounder Per-Eric Olsson, cofounder of MySQL and an AI veteran since the 1980s.
“If it was an American company that had been accepted to YC, then it would have been a hundred-million-dollar company today,” the investor says.
“But, like for many other deeptech startups, it’s a bit binary for Sics. If it works, then this company, or the way they do it, will be amazing. If it doesn’t work, it won’t have anything at all.”
Read the orginal article: https://sifted.eu/articles/can-this-small-european-startup-solve-agi/
