Valarian

Harry Stebbings Verified account @HarryStebbings · 2d

Honestly, 12 months ago I thought “sovereignty” was a BS marketing message for founders to leverage to get funding.

Now I think it is a strategic imperative and will breed some of the largest companies of our time.

Mistral is the poster child of sovereignty. The last remaining hope for Europe when it comes to frontier model providers.

I sat down with @MistralAI CEO, @arthurmensch to debunk the hype from reality and ask the uncomfortable questions. My notes from the discussion below.

1. Why Energy Is Not the Ultimate Cost That Matters

Energy makes up around 10% of the total cost of ownership of compute, so it is not where most optimization can happen. The cost of energy is not the one thing that matters in the economics of a project.

Love to hear your thoughts on this @alanchanguk @ChaseLochmiller @ZachBDell

2. What Would Mistral Do Differently If They Were Totally Resource-Unconstrained?

Mistral would build much more capacity, with capital as the main bottleneck. Customers are running out of compute, and once one use case scales they need thousands of GPUs rather than eight. Europe does not have enough compute, so Mistral is putting equity to work to build capacity ahead of demand.

Love to hear your thoughts on this, if you were resource unconstrained today, what would you do differently @alexmashrabov @bavaria_kush

3. How the Inference Market Behaves Very Much Like a Commodity Market

On-demand inference behaves like a high-margin commodity market where pricing matters a lot and demand comes from non-committed customers. Mature customers need service level agreements, reserved physical capacity and guarantees on latency and generation speed. The more precisely they describe the shape of their demand, the better Mistral can price it.

What does no one see about the inference market that everyone should see @tuhinone @neilmovva @lqiao

4. Models Have Gotten So Strong They Are Being Treated as Defense Systems

Models are now strong enough to carry out cyberattacks, and the US began banning them outside the country in May. China is also looking at restricting exports beyond a certain threshold. As with any dual-use technology, customers now want security of supply on both infrastructure and model IP.

What does no one see about the future of security that everyone should see @zeyu1337 @nikesharora @pilvar222

5. Will We End Up With a Model Landscape That Has Just Two to Three Providers?

If every company relies on two or three centralized AI providers, those providers gain enormous leverage. They will use the data their systems produce to build better systems and go after their customers' business, an end game Arthur sees as priced in today.

What will the model landscape look like in 5 years @emostaque @AnjneyMidha @mmurphy

6. The Biggest Problem With AI Effectiveness Within Enterprise Today Is...

Most enterprise problems are not well defined, and ambiguous objectives with humans in the loop cannot be solved through self-improving loops. They only get solved by deploying models that interact with people, which is why Mistral uses FDEs to define problems better.

7. Why You Can Go Public in Europe and Negativity Around European Stock Markets Is BS

Mistral can absolutely go public in Europe, because the bottlenecks of 10 years ago are no longer there. There is a huge appetite from investors worldwide to invest in Europe in both private and public equity. Mistral is preparing for the operational constraints of being public while delivering on its commercial milestones.

You think this is true, you really think you can go public in Europe @eldsjal @klarnaseb

(links in comments)

View quoted post on X

Joel de la Garza Verified account @YODELAGARZA · 2d

Our systems are not designed to contain agents.

View quoted post on X

Ariel Herbert-Voss Verified account @adversariel · 6d

We found an 0day in CPython that turns your LLM's chat template into a code execution primitive. A backdoored model on Hugging Face can own your inference server 🧵

View quoted post on X