Title mistral.ai strategic memo.pdf ✅

mistral.ai strategic memo Generative AI is a transformative technology The last year has seen a spectacular acceleration in generative AI: systems able to generate text / image conditioned on text and images. Those systems can help humans: ● produce superb creative content (text, code, graphics) ● read, process and summarise unstructured content streams thousands of times faster than humans ● interact with the world (exposed through natural or application interfaces) to execute workflows faster than ever before. The power of generative AI was suddenly demonstrated to the general audience with the release of ChatGPT; this kind of product has been in the making by only a few small teams across the world — the few researchers of these teams are now the limiting factor to create new economic actors in the field. Generative AI is about to boost productivity in all sectors and create a new industry (10B market size as of 2022, projected to be 110B by 2030, with an estimated growth rate of 35% per year), by seamlessly enhancing the human mind with machine capabilities. It is a transformative technology for the world economy that will change the nature of work to bring positive societal changes. An oligopoly is shaping up Generative AI technologies are based on years of research made in many parts of the industry and academia. The final breakthroughs, i.e. scaling training to internet-wide data and aligning models with human feedback, finally made these technologies usable by many; these breakthroughs were made by very few actors, the largest of which (OpenAI) appears to have hegemonic intention over the market. These very few actors train generative models and hold them as assets; they serve it to thousands of third-party productivity enhancing products, in addition to also serving first-party chatbot-like products. Dozens of third-party startups are created every month to build various interfaces to these generative models. We believe that most of the value in the emerging generative AI market will be located in the hard-to-make technology, i.e. the generative models themselves. Those models need to be trained on thousands of very powerful machines, on trillions of words coming from high quality sources, which is one factor that sets a high barrier to entry. The second 1
important barrier lies in the difficulty to assemble an experienced team, something that mistral.ai will be in a unique position of doing. All major actors are currently US-based, and Europe has yet to see the appearance of a serious contender. This is a major geopolitical issue given the strength (and dangers) of this new technology. mistral.ai will become a European leader in productivity and creativity enhancing AI, and guide the new industrial revolution that is coming. Current generative AI do not meet market constraints OpenAI and its current competitors have embraced a closed technology approach, which will dramatically reduce their market reach. In that approach, the model is kept secret and is only served through a text to text API endpoint. This raises the following important concerns for businesses: ● Businesses wishing to use generative AI technology are forced to feed their valuable business data and sensitive user data to a black-box model, typically deployed in the public cloud. This creates safety issues: models kept secret cannot be inspected to guarantee their outputs to be safe, thereby preventing them to be deployed in safety-critical applications. It also raises legal problems, in particular the one of falling under extraterritorial reach when sending personal data out of a company's legal territory. ● Only exposing the output of models, instead of exposing the model entirely, makes it harder to connect with other components (retrieval databases, structure inputs, images and sounds). Hundreds of products are currently built by interconnecting model outputs and inputs to create composed capacities (memory, vision, etc.). Those products would work much better and faster were the models available as white boxes (see for instance the Flamingo model, that combines white box vision and text models into a text+vision model). ● The data used to train the model is kept secret, implying that we rely on a machine that has unidentified sources, and can produce uncontrollable outputs. Filtering efforts to address this issue are only a slim and breakable guarantee that the model will not output sensitive content on which it may have been trained. As of April 2023, this issue formed the basis of ChatGPT ban in Italy. Disrupting the market from Europe By creating mistral.ai, we intend to train state-of-the-art models with counter-positions to closed-model current offerings. Our vision is to become a leading actor in the field, while developing a very valuable business around integrating these models in the European industry and beyond. mistral.ai will become a research leader in the generative AI field, eventually offering the best technology within 4 years. For this, we will first focus on several key 2