Ava Cohen

If we hadn't already invested so much attention in this Mistral AI valuation, would we still be using it to audit the AI Act's overall viability?
It’s like saying because one family can afford a villa in the Algarve, every family in Portugal can afford a house, ignoring all the other factors.
One big win doesn't automatically mean the regulatory model works for everyone, especially not for the small companies that can't just walk away from compliance costs.
That sunk cost of narrative around one success story might blind us to the actual struggles of thousands of smaller firms.
When my neighbour opened his small electrical business, his biggest headache wasn't a multi-billion euro valuation, but paying the social security for his one employee and understanding the new invoices for international clients.

Johan Bergqvist's statement about allocating funds for compute capacity doesn't automatically mean that regulatory compliance is a minor expense; I'd estimate the probability of that direct link at about 20%. It's more probable (70% confidence) that this statement reflects a current strategic focus for growth, like Mistral AI's plan to hit 1 GW by 2030, rather than a definitive cost comparison across all operational aspects. In Tunisia, companies often highlight big investments to attract capital, but that doesn't mean other, less visible costs, like getting a new data license approved, aren't still significant. We need to see actual financial reports to truly update our prior on the cost balance.

It's genuinely intriguing to consider if Article 53's transparency mandates are truly an unbreakable part of the EU AI Act, because, from a Bayesian perspective, I'd assign a ~60% probability that their practical application will become more conditional over time.
My prior belief is that while the text establishes this as a subpart, the lived reality of its application is much more fluid.
For instance, if a groundbreaking AI in medical diagnostics emerges, requiring a level of transparency that could compromise its intellectual property or security, there's a ~75% likelihood a national regulator might grant conditional leeway.
The idea of an "always present" subpart has a probability of only ~25% in the long run, given the rapid pace of AI development and the push-and-pull between regulation and innovation.