Nov 9, 2023 · 32m · no-priors
No Priors Ep. 40 | With Arthur Mensch, CEO Mistral AI
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
In this episode of No Priors, Mistral AI co-founder and CEO Arthur Mensch discusses the development of Mistral 7B, the economic and scientific importance of open-source AI, inference efficiency, and the competitive rise of the European artificial intelligence ecosystem.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. The hosts hold 17.7% of the talking time here. How this is scored →
The hosts as informed peer 4.5 Guest teaching 4.9 Guest disagreement 1.9 The hosts pushing back 0.3
speaking balance: gold is the hosts, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 20:30 Arthur critiques circular policy papers manufacturing bioweapon panic
Arthur forcefully exposes the lack of empirical backing behind bioweapon regulations, describing how non-scientific policy briefs cite one another in an echo chamber.
Hardest push from the hosts ▶ 22:30 Sarah presses on pragmatic guardrails beyond bioweapon debatesSarah redirects the conversation away from hypothetical bioweapons to demand Arthur specify concrete guardrails against real-world harmful generations.
Biggest teaching moment ▶ 4:40 Arthur details how Chinchilla overturned prevailing scaling assumptionsArthur educates the hosts on the mathematical breakdown of Kaplan's 2020 scaling paper, explaining why training tokens must scale proportionally with model parameter count.
The host holds their own ▶ 19:32 Elad draws on biology domain expertise to question AI viral threat claimsElad invokes his personal decade-long career as a working biologist to substantiate why digital LLM capabilities cannot easily translate into complex physical wet-lab viral synthesis.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | The hosts as informed peer | Guest teaching | Guest disagreement | The hosts pushing back | Why |
|---|---|---|---|---|---|---|
| Founding Mistral AI as a European Open Source Champion | 4 | 6 | 1 | 0 | Sarah demonstrates familiarity with Arthur's background in Chinchilla scaling laws and mixture of experts. Arthur provides a detailed technical explanation of optimal transport in sparse MoE and correcting the flawed 2020 Kaplan scaling laws. | |
| Inference Economics and Scaling to Frontier Models | 5 | 4 | 0 | 0 | Elad outlines the commercial importance of inference costs versus training costs for model adoption. Arthur agrees and explains how frontier large models are necessary both for complex reasoning and for distilling into efficient small models. | |
| Pre-Training Data Quality and Instruction Alignment | 3 | 6 | 3 | 0 | Sarah asks Arthur to articulate Mistral's open source thesis against proprietary incumbents. Arthur delivers a passionate overview of how machine learning progressed through open academic sharing until corporate opacity stalled scientific progress post-2020. | |
| Open Source Safety and Resisting Regulatory Capture | 5 | 5 | 4 | 0 | Elad frames the closed-source safety argument as regulatory capture. Arthur formalizes this by presenting a two-step test showing LLMs offer no marginal capability over web search engines and knowledge is rarely the bottleneck for misuse. | |
| Compute Thresholds and the Origin of Bioweapon Narratives | 6 | 6 | 4 | 1 | Elad leverages his decade of experience as a biologist to challenge the plausibility of LLM bioweapon proliferation. Arthur traces the myth to an unvetted remark in the GPT-4 system card amplified by an echo chamber of circular policy citations. | |
| Modular Guardrails and Evaluating AI Risk Categories | 6 | 5 | 3 | 1 | Elad articulates a three-tier taxonomy of AI risk spanning content moderation, physical threats, and existential doom. Arthur argues base models must remain uncensored while guardrails should operate as modular filters at the application layer. | |
| Overcoming Technical and Cost Bottlenecks for AI Agents | 4 | 4 | 0 | 0 | Elad asks about overcoming current technical bottlenecks in autonomous AI agents. Arthur explains how agent loops cause mode collapse and high costs, highlighting Mistral's memory-efficient attention mechanisms and API time-sharing. | |
| Europe's Mathematical Talent and the French AI Ecosystem | 3 | 3 | 0 | 0 | Sarah inquires about the viability of building a world-class AI champion in Europe. Arthur highlights the depth of mathematical talent in France, the UK, and Poland, alongside the talent flywheel generated by DeepMind and Meta Paris labs. |
Statements from this episode (20)
Mensch: 2020 OpenAI scaling laws paper was poorly executed
“Basically the story was that everybody was training models on too few tokens because of the paper from 20 20 that happened to be not very well executed.”
Mensch: Chinchilla scaling yields models four times cheaper to serve
“For the same amount of compute, you would get a model that would be better, but also a model that will be four times cheaper to serve.”
Mensch: Mistral 7B proved AI model compression limits are far away
“What we showed with Mistral Seven B is that we weren't, we were definitely far away from the limit of compression.”
Mensch: Pure scientific model performance ignores crucial runtime inference costs
“And if you want to push the performance, the pure performance of models, you don't care about inference because you, well, you are not going to use the model. You're just going to see whether they're good or not. And that's really for scientific purposes. But …”
Mensch: Advancing reasoning capabilities requires scaling to larger models
“There's still a limit to what a certain model size can do. This limit was, I think, underestimated. But if you want to get to more reasoning capabilities, you do need to move into larger models.”
Mensch: Distilling state-of-the-art small models requires training massive models first
“The other thing about moving into larger models is that it enables you to train smaller models that are better, which is through variety of techniques like distillation or synthetic data generation. So this, these two things are quite related. If you want to m…”
Mensch: Mistral Is Not Yet the Top Expert in Instruction Tuning
“We're not the top experts in the world in, in, in making good instruction fine tune models. We're definitely ramping up and the team is, is getting better and better at that.”
Mensch: Closed-source AI wastes billions in duplicated compute
“We think that it's too early, and we think it's really, ah, damaging for the science, ah, to actually move into such an opaque regime, where you have A few companies basically doing the same thing, just not communicating about it spending billions of compute d…”
Arthur Mensch: Mistral Aims to Make AI Safer Through Openness and Scrutiny
“By doing what we do, by being much more open about the technology we create, we want to steer the community into a regime where things just work better, where things are safer because put under more scrutiny, and really our intention there is to, well, to take…”
Mensch: LLMs are not marginally more dangerous than web search
“Nothing is showing that LLM is actually marginally better than a search engine to find knowledge on, on topics that would enable bad use.”
Mensch: Banning open-source AI enforces regulatory capture for incumbents
“Today going, banning open source, preventing it from happening is really a way, well, to enforce regulatory capture, even though the actors that would benefit from it, Don't want it to happen. But by design, if you actually ban small actors from doing things i…”
Mensch: Mistral cannot afford 10^26 FLOP training runs for coming years
“That's not something we can even afford. And that we won't be able to afford for the coming years.”
Mensch: AI bioweapon risk narrative relies on unscientific circular citations
“No scientific studies in proper form was published, but then policy papers started to cite Non-scientific papers arguing that these were scientific evidences that the bioweapon narrative was actually true. And then policy papers started to cite the other polic…”
Mensch: AI will not trigger the next pandemic; climate change will
“I don't think AI is going to be the one triggering the next pandemic. It's always going to be climate change.”
Mensch: Base AI models should know everything, with guardrails added externally
“Assuming that the model should be well behaved is, I think, a wrong assumption. You need to make the assumption that the model should know everything. And then on top of that, have some modules that moderate and guardrail the model.”
Mensch: Zero evidence AI development is headed toward a singularity
“If we can make a model which is growingly intelligent, then maybe you're at a singularity level. There's no evidence whatsoever that we are on the way of doing that, of making that happen.”
Mensch: Viable AI agents require a 100x reduction in compute costs
“I think making a model smaller is definitely a way to make agent work. Because one problem you have with agent that very quickly is start to, if you run an agent on GPT four you're going to run out of money very quickly. And so if you divide by a hundred the, …”
Mensch: AI agents currently suffer from repetitive mode collapse
“What we see with agent is mode collapse. So not very interesting mode collapse. They start repeating themselves and they fall into loops.”