Apple has quietly acquired another AI firm involved in reducing the size of Large Language Modules for the iPhone, this time Paris-based Datakaleb.
Following news that Apple secretly bought the Canadian AI firm DarwinAI at the start of 2024, it has now been revealed that shortly before then, it acquired a Parisian company. French publication Challenges says it bought Datakaleb in December 2023.
Just as DarwinAI works on making deep neural networks smaller and more efficient, so Datakaleb is now focused on the compression of algorithms. The smaller a process algorithm can be, the less energy it uses — and Apple has repeatedly been shown to be working on Large Language Modules (LLM) that are made small enough to run entirely on device.
"With compression, you can go from an Nvidia semi-processor at 10,000 euros [$10,650] to a [regular] STMicroelectronics chip for a few cents," a source said to be familiar with the deal told Challenges in translation.
Datakaleb, based in Parc Monceau within two miles of the Apple Opera store in Paris, also has a history of using AI in facial recognition. That work included how in 2021 the city of Cannes used it to monitor an entire concert hall to confirm all attendees were wearing masks.
Face recognition regulations have not been established in France. In the absence of them, Datakaleb reportedly moved to work on compressing AI algorithms.
The company has around a dozen expert staff, including ex-students from the Sorbonne University, Centrale and Polytechnique.
Datakaleb has a reported annual revenues of one million euro, leading to the speculation that Apple will have had to spent several million to buy it. According to Challenges, one Datakaleb founder left in 2020, with the rest of the initial founding staff expected to depart.
Apple keeps buying companies — and only declaring that it has when the size of the deal means that it must. This is also far from Apple's first work in Paris, as the company has had a base of operations there for over 40 years.
2 Comments
Nice, Apple is really looking at tech level 2 for AI. Not just brute force "proof of concepts," they seem to be seriously trying to drive refinement, cleaner, smaller and more usefully integrated AI.
Why do the graphic and the linked website say "Datakalab", but the article says "Datakaleb" everywhere?