Apr 29, 2024, 10:00 AM ETInterview, Fireside Chat

Arthur Mensch: Open vs Closed - Who Wins and Mistral's Position | E1146

Mistral's core mission is to bring "freedom to developers" by enabling them to own, modify, and deploy models, rather than relying on centralized API providers.
The company positions itself as an enabler for vertical applications, believing value will accrue at the application layer where specific data and user feedback create differentiation.
Mistral aims to make frontier AI ubiquitous, targeting a future where general-purpose models serve as the starting point for a broader ecosystem of tools and lifecycle management platforms.
The organization operates on a "platform approach" rather than a verticalized one, focusing on providing the infrastructure for others to build specialized AI.

Compute availability remains the primary bottleneck; Mistral currently possesses approximately 1.5K A100 GPUs, which is a small fraction of its competitors' capacity.
Arthur Jaffe notes that raising $2 billion in a seed round was not feasible in 2023, limiting the speed at which the company could scale infrastructure and hiring.
The company faced delays with compute providers that impacted operations, though significant improvements in capacity are expected in the coming months.
Mistral has maintained a lean team structure (initially 25 people, now expanding) to prioritize speed and agility, operating on the philosophy that "a team of five is faster than a team of 50" if properly uncoupled.
The company acknowledges that while capital is correlated with compute, quality is not solely dependent on scale, allowing Mistral to compete via efficiency and algorithmic improvements.

The release of the 7B model was strategically targeted at the "efficiency-performance" gap, allowing developers to run models locally on consumer hardware like MacBooks and smartphones.
Mistral's approach emphasizes "compute multipliers," focusing on extracting efficiency gains through better algorithms and data quality rather than simply increasing raw compute spend.
The company is shifting toward a hybrid open-source and commercial model, releasing large models like the 8x22B under open licenses while maintaining commercial licensing for specific enterprise needs.
Arthur Jaffe predicts the end state of the AI landscape will not be commoditized models, but rather specialized models built by application developers to reduce latency and improve domain-specific performance.
Model quality is currently bottlenecked by data quality and evaluation mechanisms rather than just compute, requiring more sophisticated methods to refine high-quality datasets for specific tasks.

Mistral learned from DeepMind that organizing into sufficiently uncoupled teams (sharing infrastructure and code but working independently) maximizes innovation speed in general-purpose model development.
The company bridges the gap between science and sales by ensuring the science team has direct exposure to user problems and the sales team understands the technical nuances of the product.
Management has evolved to operate with "almost fully transparent" feedback mechanisms, which Jaffe cites as a key factor in scaling without breaking organizational culture.
The company is actively recruiting senior AI scientists primarily in the US (Silicon Valley) while leveraging junior and mid-level talent pools in France, Poland, and the UK.
Jaffe advises that product development and go-to-market strategies should be more staged, noting that launching sales motions before having a mature product created initial organizational friction.

Mistral respects competitors like OpenAI, Anthropic, Google, and Cohere, viewing the landscape as collaborative rather than purely adversarial, with shared goals for AI advancement.
Brand trust is a critical determinant of adoption; Mistral leverages its open-source releases to build community vouching and trust, which is essential for enterprise adoption.
The company anticipates that marginal costs for AI applications will decrease as models become more efficient, but foundational layer margins will not drop to zero to ensure fairness and sustained innovation.
European enterprises are currently behind the US market by approximately one year in AI adoption, though executive support is growing and moving from experimental budgets to core operational budgets.
Mistral warns against the "lethargy" of European enterprise adoption but remains optimistic that local talent and a growing ecosystem will support a serious European AI industry.

Mistral expects to remain a leader in open source while licensing unique assets to cement strategic relationships with major cloud providers (Azure, AWS).
The company foresees a future where application makers build their own low-latency, vertically specialized models, with Mistral providing the "foolproof" tools to do so without needing expert AI knowledge.
Jaffe predicts that job displacement will occur but views it as a shift toward higher abstraction levels where humans focus on creativity rather than being replaced by AI.
In 10 years, Mistral envisions a strong developer platform that allows users to create, evaluate, and customize AI applications, with the company providing both commercial and open-source models as the foundation.
The company acknowledges that while the ecosystem in Europe is smaller and more fragmented than the US, the "revolution" in software offers new opportunities for European actors to grow rapidly.

Arthur Jaffe cites the high demand for Mistral's models as the most unexpectedly challenging aspect of scaling, exceeding initial capacity management expectations.
He reflects that he was initially unaware of the immense energy required to care for his child, noting parenthood has been a significant life adjustment.
Jaffe admits to underestimating the speed of community adoption and the speed of his own management learning curve, specifically regarding organizational scaling and stakeholder management.
His personal methods for managing stress include running, cycling, and prioritizing time with his family.
He warns that global warming represents a "race for survival" where AI must play a critical role in bringing efficiency and control to the planet's processes.