Aleph Alpha and Mistral: two figureheads for Europe in 2024

Faced with the American AI giants, Europe has two champions: France’s Mistral AI, which has already given in to the siren calls from across the Atlantic, and Germany’s Aleph Alpha. Two very different strategies.

490 million in two rounds for the French company; 494 million in two rounds for the German company: that’s where the similarities between the two European champions of artificial intelligence, Mistral and Aleph Alpha, end, or almost end. Founded in April 2023, the flamboyant French nugget has welcomed to its capital the finest American VCs, Andreessen Horowitz (A16Z) and LightSpeed Ventures. Founded in 2019, the more discreet German start-up is already generating sales; it relies on an incubator in Baden-Württemberg and above all on the flagships of German industry: Bosch, SAP, Burda, Schwartz (Lidl)…

Start-up models

Buoyed by its investors’ marketing machine, Mistral is making an increasingly frequent appearance in the Silicon Valley buzz. Its latest creation, Mixtral 8x7B, is inspired by the architecture credited with GPT-4: a ‘mix of experts’ where several models are networked together. What’s more, the SMOE (Sparse Mixture of Experts) framework chosen means that the entire neural network is not required to process each token. Mixtral 8x7B thus achieves performances comparable to those of much larger models, such as Meta’s Llama 2 or OpenAI’s GPT 3.5.

What’s more, Mistral is a pioneer in small language models: Mistral 7B is well ahead of the models from Microsoft and Google, which won’t be launching its Gemini Nano until next year. However, since the success this spring of Alpaca, a degraded model of Llama that ran on a PC, there has been a race to create a language model capable of animating a mobile phone (rather than paralysing it, as is the case today). Its main focus is on English and code, the two dominant languages in AI today.
For its part, Aleph Alpha is developing a series of large multimodal language models, Luminous, with the ambition of creating a ‘sovereign’ AI that respects the linguistic diversity of the Old Continent. The German start-up’s main model has been trained on text data in English, French, German, Spanish and Italian. The company explains that its ambition is to develop the application of generative AI in complex sectors such as finance, health, law and security, with a focus on data confidentiality and interoperability.

American market

Very different, the two start-ups nonetheless joined forces this autumn to fight Parliament’s proposals on the AI Act, deemed too harsh for foundation models. We saw the Franco-German couple reunite at the European Council to obtain a more favourable framework. The precise details have yet to be published, but it seems that the joint efforts of the two start-ups have borne fruit.

While both sides of the Rhine are rightly celebrating their respective champions, the road to competing with the American giants remains titanic. Even with the backing of venture debt, the funds raised by Mistral and Aleph Alpha would appear to pale into insignificance in the face of OpenAI’s 11.3 billion and the power of groups such as Google and Microsoft.

Company

Funds raised since inception ($M)
OpenAI 11 300
Anthropic 7 600
Inflection AI 1 500
Mistral AI 657,8
Aleph Alpha 642,8
Cohere 434,9
AI21 Labs 326,5
Stability AI 173,8
Character AI 150
Jasper 131
Source : Crunchbase

European start-ups, on the other hand, compare favourably with Canada’s Cohere ($435m), Israel’s AI21 ($326.5m) and the UK’s Stability AI ($174m). They joined forces this autumn to lobby on the AI Act.

‘In France, we don’t have oil but we have ideas’ was the advertising slogan in 1977, at the height of the oil crisis. Half a century later, Mistral AI and Aleph Alpha will have to prove their worth.

By Maurice de Rambuteau

 

To subscribe to Qant free of charge, click here: Qant: Cognitive Revolution and Digital Future – The QANT newsletter: Cognitive Revolution and Digital Future (kessel.media)

To find out more

Noam Shazeer et al, Outrageously Large Neural Networks: The Sparsely-Gated Mixture-of-Experts Layer, Arxiv 2017