Mistral AI Rejects US Tech Monopoly With Open-Source Models

Mistral AI以开源模型对抗美国科技垄断 联合创始人Arthur Mensch认为，开放模型权重是打破正在形成的美国人工智能双寡头垄断、解决结构性幻觉并确保有效安全审核的唯一途径。 在本次炉边谈话中，Arthur Mensch讨论了Mistral AI致力于构建顶级开源大语言模型系列的战略布局，以此抗衡专有的“黑盒”方案。针对安全和幻觉等行业挑战，他主张通过开放权重、未审查的基础模型以及检索增强生成（RAG）来赋予开发者控制权。Mensch还探讨了法国人工智能生态系统的优势、训练架构的技术现实，以及工程师向人工智能领域转型的进阶路径。 Mistral AI通过采用开放核心模式与美国科技巨头形成差异化，通过发布模型权重构建高品质的开源模型系列，而不是隐藏在黑盒API背后。 (1:12) Meta发布Llama 2被视为合作契机而非威胁，这创造了社区协同效应，让开发者能够复用海量算力成果并在其基础上继续构建。 (2:22) 开源模型让开发者能够直接访问模型权重，从而提升AI的安全性和可控性，便于他们实施自定义审核策略并有效打击虚假信息。 (4:04) 解决企业级幻觉问题可采取多管齐下的方案，包括通过延长训练以嵌入深层知识，以及微调开源模型使其严格遵循检索增强生成（RAG）的上下文。 (6:15) 法国人工智能的复兴得益于其注重数学和计算机科学的高标准国家教育体系，以及科技初创生态系统的蓬勃重振。 (7:55)

Arthur Mensch of Mistral AI argues that enterprise artificial intelligence will be won by open-weight models, controllable safety mechanisms, and cost-effective production deployment rather than proprietary black-box APIs or speculative existential risk debates. Open-Core Strategy Against Proprietary Lock-In (1:12) Mensch positions Mistral as an open-core alternative to proprietary US tech giants like OpenAI and Anthropic, whose closed APIs prevent developers from inspecting model weights or controlling operational costs. Comparing generative AI's evolution to the historical transition of operating systems and cloud orchestration to open source, Mensch asserts that releasing weights enables community-driven cost-sharing and compute reuse. He treats Meta's Llama 2 release as an ecosystem opportunity rather than a threat, anticipating mutual technical synergies. John Collison notes that while businesses frequently adopt proprietary APIs for early proofs of concept, Mensch emphasizes that production deployments inevitably hit strict infrastructure and financial boundaries. In Mensch's view, open models provide the granular control required to adapt systems to specific data flows. He also frames European open-weight development as an urgent counterweight to what he warns is a looming US oligopoly of Microsoft and Google seeking regulatory capture. Safety as Controllability and Mitigating Hallucinations (4:04) Challenging the argument that open-source models compromise AI safety, Mensch contends that open weights actually empower it. He argues that current industry practices confuse safety with blanket censorship: to reliably filter hateful content or combat misinformation, base models must first be able to process those concepts. Direct weight access allows downstream developers to implement nuanced, application-specific guardrails rather than relying on generic, closed API filters. To solve LLM hallucinations—a primary blocker to commercial adoption—Mensch advocates a twofold approach. First, Mistral maximizes pre-training compute relative to parameter count to embed deep foundational knowledge into models. Second, they rely on retrieval-augmented generation (RAG), which injects external context directly into prompts. Mensch notes that fine-tuning open models to strictly adhere to retrieved data makes RAG substantially more reliable, supported by Mistral's planned release of dedicated embedding models. Production Realities, Efficient Decoders, and Specialization (13:02) Mistral's development roadmap focuses on moving systematically along the cost-performance curve, starting with highly compressed, efficient models that can run on accessible hardware like a MacBook Pro before scaling up. Mensch defends the standard decoder architecture as exceptionally resilient when trained from scratch, explaining that Mistral's technical advantage lies in rigorously distilling web datasets, given that uncurated internet data is overwhelmingly low quality. He expresses skepticism toward neuro-symbolic AI and knowledge graphs, calling them late to the party compared to deep learning. Looking beyond the cloud, Mensch forecasts a hybrid operational model where compressed sub-billion-parameter models run locally on-device to handle privacy and latency demands. While open source currently lags behind proprietary models in voice and vision, Mensch notes that Mistral is intentionally focusing on text first—the most compute-intensive foundation—before expanding into multimodal architectures. Grounded Regulation and Dismissing AGI Narratives (8:48) Mensch strongly opposes proposed AI regulations that impose hard compute thresholds or benchmark limits on frontier models, warning against divergent regulatory regimes between Europe and the US. Instead, he advocates for good rules centered on proportional documentation and auditability requirements that scale strictly according to a model's public user base. Mensch explicitly rejects the goal of Artificial General Intelligence (AGI) as an ill-defined semantic trap that fuels speculative sci-fi and existential risk narratives. Describing current LLMs as soft key-value networks that interpolate sentences rather than conscious agents, he calls for harder, modernized benchmarks to measure real-world enterprise utility rather than theoretical panic. He attributes the rapid rise of France's AI sector to this grounded focus, underpinned by strong national education in mathematics and computer science. While Mistral targets delivering a leading open-source model and expanding developer orchestration tools, open-weight deployment faces unresolved friction with European policy, particularly the technical paradox of reconciling internet-scale model training with the GDPR right to be forgotten.

Mistral AI 联合创始人 Arthur Mensch 阐明其核心论点：与主导市场的美国闭源大模型相比，开源核心（Open Core）模式是打破黑盒垄断、降低企业级生产成本并赋予开发者真实控制权的必然选择；欧洲应当依托本土数学与工程基础推动务实落地，而非受制于通用人工智能（AGI）炒作与监管俘获。 开源核心模式与企业级生产成本控制 (1:12) Arthur Mensch 指出，诸如 GPT-4 与 Claude 2 等专有模型均依托黑盒 API 运作，使开发者失去对内部机理的把控。Mistral AI 确立了“开源核心”（Open Core）定位，即在公开发布模型权重的同时提供专有托管方案。针对 Meta 推出的开源模型 Llama 2，Mensch 认为这并非威胁，而能形成社区协同并复用既有算力。他援引操作系统及 Kubernetes 编排的发展史断言，专有技术终将被开源社区的分摊成本模式取代。John Collison 亦补充说明，多数 AI 企业片面追求性能上限，而企业客户的核心诉求在于性价比与防止厂商锁定。 Mensch 分析称，企业在概念验证阶段通常对调用成本不敏感，但推向终端生产后必将遭遇严苛的基础设施与开销瓶颈。Mistral 的发展路线图沿“成本-性能曲线”推进，率先打造可在 MacBook Pro 等消费级硬件上运行的小型高品质模型，计划于次年第二季度推出顶尖开源基准模型，并逐步自底向上构建面向开发者的编排与微调软件栈。 重构模型安全、控制权与监管逻辑 (4:04) 针对开源模型可能危及 AI 安全的普遍质疑，Mensch 反驳称开源反而在深层强化了安全机制。他主张安全的核心本质在于“可控性”（Controllability），闭源 API 施加的严厉过滤阻碍了企业根据具体场景构建审查规则。开发者必须直接掌控未审查的基础模型权重并赋予其识别恶意内容的能力，方能在下游应用中架设稳固的防护栏以遏制虚假信息与有害言论。 在政策与战略层面，Mensch 明确反对依据算力消耗或基准测试跑分等绝对数值划定监管红线的做法，认为此类规则极易扭曲创新。他倡导推行以透明度与可审计性为核心、依用户规模逐级递增的文档披露机制。此外，Mensch 批评 AGI 是缺乏清晰定义的“语义陷阱”，并驳斥了科幻式的存在性风险假设，敦促行业将注意力从恐慌情绪转移到底层缺陷的扎实攻关上。 架构务实主义：数据提炼与幻觉抑制 (6:15) 针对企业落地中最为棘手的幻觉难题，Mensch 提出了涵盖充分预训练与检索增强生成（RAG）的工程解法。Mistral 的技术差异化在于针对既定模型体量极限推高训练算力以深度注入知识，同时配套研发专用嵌入（Embedding）模型将外部事实引入提示词中。由于开源模型支持底层微调，开发者能够强制模型严格遵从外部上下文以消除胡编乱造。 在底层架构上，Mensch 确认 Mistral 坚持自零开始预训练并沿用经典的解码器（Decoder）路线，认为该架构高度稳健且难以撼动，而神经符号（Neuro-symbolic）与知识图谱方案在深度学习面前已显著落伍。他强调互联网语料中绝大部分为无价值垃圾，Mistral 的关键竞争力在于构建精密的提炼管道。关于机器意识与图灵测试，Mensch 将大语言模型本质定性为运行时进行知识检索与插值的“软键值网络”，并无真正的概念理解。 欧洲生态战略、隐私合规与端侧演进 (7:55) Mensch 将法国当前的 AI 繁荣归结为强调数学与计算机科学的严格国家教育体系，以及重新被激发的科技初创生态。但他同时对欧洲决策层发出警告：美国巨头正与白宫互动形成寡头垄断与监管俘获，若欧洲无法在一年内迅速行动建立独立生态，行业格局将彻底固化。在隐私法规方面，Mensch 承认通用数据保护条例（GDPR）的“被遗忘权”在制定时未考虑大语言模型，在数据端尽量过滤之外，企业生产环境更需依赖推理端过滤栈来拦截包含用户信息的输出。 面向未来硬件部署，Mensch 预测大语言模型将演变为云端与本地并存的混合架构。随着模型压缩技术的进展超出预期，他预言数十亿参数量级的专精小模型将很快在端侧设备独立运行，从而彻底规避将敏感数据传输给第三方的合规隐患。对于软件工程师转向 AI，Mensch 建议从训练微型模型与硬件内存交互入手，培养不可替代的工程调优直觉。 Arthur Mensch 试图破除围绕大模型与 AGI 的神话色彩，确立以工程控制权、深度数据清洗与欧洲技术主权为支撑的开源替代方案。然而该愿景仍面临重大现实考验：团队在资源受限下能否按期在多模态等计算密集领域匹敌专有巨头，以及如何在欧洲严苛且未匹配现代 AI 机制的 GDPR 框架下化解模型记忆带来的合规风险。

[2:22] Arthur Mensch: by giving to the community, you can take from the community. [3:23] Arthur Mensch: small model, but it's a very good small model [4:04] Arthur Mensch: playing with text and closed source APIs. [6:15] Arthur Mensch: cram some knowledge in them [11:19] Arthur Mensch: ill-defined goal [11:19] Arthur Mensch: semantic trap [13:02] John Collison: rooting around [his] underwear drawer. [17:15] Arthur Mensch: intuition of what works, what does not work [19:19] Arthur Mensch: is not going to happen in two days [23:36] John Collison: like a kid would play [26:07] Arthur Mensch: regulatory capture [28:36] Arthur Mensch: GDPR wasn't designed with language models in mind. [30:06] Arthur Mensch: pretty strong and very hard to beat, if you know how to train them. [30:06] Arthur Mensch: on the internet you have mostly garbage [30:06] Arthur Mensch: very late to the party [31:56] Arthur Mensch: soft key-value networks [36:20] Arthur Mensch: wanted to not be alone [37:59] John Collison: the precise day is also a surprise

[0:08] Hello everyone! I am very excited for this. [0:12] We're joined by Arthur Mensch, [0:13] who was originally in academia but was rescued from academia [0:18] and joined the ranks of the working society. [0:21] And then worked at DeepMind as a researcher. [0:25] That was where you led the chinchilla paper, correct? [0:29] Yeah, that was one of the main. [0:31] A very foundational paper in AI, and now he is the co-founder of Mistral. [0:38] Who here knows what Mistral is? Okay, good. [0:43] And the other people were lost and wandered into AI Day by mistake. [0:47] We're not going to spend a whole lot of time on the basics here, [0:51] but I'm super excited to jump right in to the important questions, [0:55] like, why the name Mistral? [0:58] So Mistral is a French wind and it has two vowels, which are I and A. [1:03] And I and A is AI in French. [1:06] And I think everybody here knows what that means. [1:12] You released this kind of technical paper and described [1:16] what you think Mistral's market opportunity is, [1:18] or maybe what you think everyone else is doing wrong. [1:20] And, some of these were, you know, [1:22] you think it's very important for there to be an European AI lab. [1:25] You know, you think everyone else is underrating open source. [1:29] How would you describe what you think [1:32] the most important differentiators are for Mistral today? [1:36] I guess our larger differentiator is that we take [1:38] a different approach to technology. [1:42] There were some demonstrations around GPT-4 and Claude 2. [1:46] Those two are very proprietary models. [1:48] You don't know anything about them. Everything happens behind a black box API. [1:53] What we want to promote is an open approach to models. [1:56] We want to create models and release their weights. [2:00] We want to have a family of models which is open source, [2:04] and we want it to be the best of the open source models. [2:08] We are also developing a proprietary family of models [2:12] and some hosting solutions. [2:13] But we really want to be an open core LLM provider. [2:17] And this is quite differentiating compared to the big US actors [2:20] that are currently in the space. [2:22] Obviously, the notable exception to the US providers... [2:29] You know, you look at Google, which has its successful closed-source Bart [2:34] and OpenAI, but Meta released Llama 2. [2:38] And so does that change anything there, or how do you view the Meta releases? [2:42] We view it as a strong opportunity. [2:44] The good thing about models is that once you have access to a model, [2:48] you can retrain it. You can reuse the compute that has been spent on [2:51] the model itself to actually build something better. [2:53] And that's really where the open source community play is important. [2:58] We are about to release some models that [3:01] we hope that the community will take and make better. [3:05] So improve them on the controllability side. [3:08] Fine tune them for specific purposes. [3:11] We think that by giving to the community, [3:13] you can take from the community. And we are lucky that Meta also thinks the same. [3:17] And we believe that there will be some very interesting synergies to be built [3:21] in the coming years. [3:23] I find it very exciting seeing all the open source progress. [3:27] And by the way, we will have time for questions at the end. [3:29] So if you're thinking of questions, we will get to them shortly. [3:33] I asked the person I was with just before this, what questions I should ask you. [3:38] And they said, when do we get to use the Mistral product? [3:42] So you will have to wait a couple of days more because we have something coming. [3:46] But only a couple of days? [3:48] Yeah, we have something cooking in a couple of days. [3:52] It's only a small model, but it's a very good small model. [3:55] That's very exciting! [3:57] And I won't say more, but something interesting is coming. [4:01] Okay, you heard it here. [4:04] You describe the importance of open source models in your paper and your materials. [4:12] You also describe the importance of AI safety. [4:18] How do you square those two? [4:20] It's a very good question. [4:23] We think that the benefits of providing open source models to the community [4:28] are great for safety, because if you have access to the weights of the model, [4:32] you can tweak it to do a moderation: a model, for instance, which counteracts, [4:39] I think, the biggest bad usage of generative AI today, [4:42] which is production of misinformation. [4:44] And if you do censor the model heavily, behind proprietary API, [4:50] this is not something you can do because you're blocking any kind of content [4:54] that you may want to moderate. [4:56] So safety is, I think, empowered from open source. [5:01] So you think basically the most important thing is to be able to understand [5:05] what the model is doing. And then you can encode whatever policy [5:09] or content goals you have. [5:10] But it's very bad if that's happening in a closed source way? [5:13] Well, I don't know if it's bad, [5:15] but I think it's important to understand that a strong model should be able to, [5:19] process bad content, hateful content, [5:24] because it needs to understand that it's bad, that it's hateful, [5:26] and if it understands that, [5:28] then you can control it to block it from end users applications. [5:32] And so you do need to have access to the underlying weights to do that. [5:36] And I think there's a strong misunderstanding on safety: [5:42] what we really want from models is for them to be controllable [5:46] and you want the downstream users, [5:48] the users that create some applications, to take these models [5:52] and control them into an end user application. [5:57] And so if it's moderation, for instance, [5:59] you need to start from a model which is uncensored from the beginning. [6:04] If it's a chatbot, you do want to impose some guardrails, but to impose guardrails, [6:10] it's much better if you have access to [6:11] the weights than if you are just playing with text and closed source APIs. [6:15] What is your view on how we solve hallucination? [6:21] Stripe uses LLMs as you saw in one of the demos earlier, [6:25] Stripe uses and it's rolling out more uses of LLMs in lots of parts of the product, [6:30] but it seems one of the main inhibitors to serious business applications [6:36] is the fact that they just sometimes make stuff up. [6:38] Yeah, I think there's several things in which you can reduce these hallucinations. [6:43] The first thing is you need to train the models for a long time. [6:46] You need to to cram some knowledge in them. [6:49] And that's what we've been doing. [6:51] And this is important, right? [6:52] Because this is one of the places that you differentiate against [6:55] the other people producing models is the amount of training. [6:58] Yeah, exactly. The amount of training you put on a certain size of models. [7:01] So this is one of the things you can do. [7:04] I think one thing that is doing that very well is to use retrieval augmentation. [7:09] A way of getting some knowledge, [7:11] putting it in the prompt of the model and then fine tuning the model [7:15] or modifying it so that its only goal is to follow the context it has [7:19] to answer a question. [7:21] And this is something that is made easier if you can fine tune [7:24] the model to actually follow the context. It seeks to answer questions. [7:27] If you only use a ready-made model, it's slightly harder. [7:31] So we think that hallucinations is also something that open models will promote. [7:36] For this you need two kinds of models. It's quite interesting. [7:39] You need a generative model, but you also need an embedding model. [7:42] So a model that takes some text and basically produces an embedding. [7:47] So some floating point representation of it. [7:50] And this is something that we will also provide very soon. [7:55] It seems like France is having very much an AI renaissance at the moment. [8:02] Why? [8:06] This is very true. [8:07] And this is very exciting for us [8:10] as a company to be in such an ecosystem. [8:13] I think this is the result of several factors. [8:16] One of them is the French institution: [8:20] the French system of education, [8:24] which promotes mathematics and computer science quite a lot. [8:27] And as it turns out, these are [8:29] the two competencies that are very important to make models work. [8:32] This is one of the reasons and I think it also goes with, [8:37] the general renaissance of the French tech ecosystem. [8:42] So the two things are now connecting with each other, [8:44] and I expect to see a lot of interesting things in generative AI in France. [8:48] What do you think we need now to further promote the ecosystem? [8:52] Obviously, there's a lot of energy, as you say, the French technical talent [8:58] and quality of education has always been very good, [9:00] but it feels like it's now being unleashed in the startup scene. [9:03] The capital is much better. [9:04] And so if you were encouraging the VCs and policy makers and founders now, [9:09] what would you like to see happen from here? [9:12] I think we need boldness. The capital is good. [9:16] It could be better. There's really a lot of opportunities [9:19] to found some new AI companies. [9:22] We have very interesting ideas, [9:24] and I encourage all of the VCs in the room to take another look. [9:28] I also think we need [9:30] some control on the regulation that is going to happen in Europe. [9:34] There will be some differentiated rules between Europe and the US. [9:38] They shouldn't be too differentiated, [9:42] if Europe wants to play a role. [9:45] So this is something that we are actively discussing. [9:49] Um, and what else? [9:51] I guess these two things would already make quite a difference. [9:56] I'm curious on the regulation. [10:00] Can you give us an example of a good rule that you think we should have for AI [10:04] and also a bad rule that you think we should not have, [10:07] but might end up with? [10:10] A good rule would be to say that models should be well documented, [10:15] that they should be auditable so you could have access to the weights, [10:21] and be able to explore them in depth and without restriction. [10:27] This is one good rule. [10:28] All models, even a model Stripe trains internally [10:32] or only a model that's released to the public? [10:35] So documentation in our mind should be proportional to the user base. [10:41] If it's internal, you don't need to document it too much. [10:43] If you give it to many people, you should document it more. [10:47] A bad rule is a rule that is based on some hard threshold. [10:51] So let's say your model is above a certain performance [10:54] or a certain benchmark, It should be heavily regulated. [10:57] This is a rule that's being discussed in the United States, I think. [11:02] Or maybe also here? [11:03] Any kind of method rule where you have numbers, [11:06] you have this amount of compute makes you a frontier model [11:08] and you should be regulated. [11:09] This capacity makes you a frontier model and you should be regulated. [11:12] We don't think that this is a proper way of accompanying the growth [11:16] of the technology. [11:19] People love to talk about, you know, artificial general intelligence. [11:22] And that was kind of central to the founding mission [11:28] of many of the foundational model companies. [11:32] I don't think I saw mention of it in your strategy memo. [11:36] Should I read something into that? [11:39] The thing with artificial general intelligence is that [11:43] it's an ill defined goal. [11:46] I think so, so I'm glad to hear someone saying so. [11:49] So you have many people talking about it, but meaning different things. [11:52] And so we don't want to to go into that semantic trap [11:56] of talking about things that are ill defined. [11:58] What we really want to make are models that can be adapted to [12:01] the enterprise use case to make everybody's life easier. [12:05] We don't believe in existential risk, [12:07] and the fact that the model is going to go berserk [12:09] because it becomes too intelligent. [12:10] I mean, if you look at it, [12:12] even the best models still have huge limitations. [12:15] And these limitations need research. [12:17] And research should be done by companies like ours. [12:21] That makes a lot of sense. [12:23] So you don't think that much in terms of, you know, [12:28] Mistral producing AI that can get artificial general intelligence [12:31] because you think that's just not a well specified question? [12:33] Yeah, I think we want Mistral to produce [12:36] we are producing models that will empower people [12:40] to work better, to work faster, to focus on the most important things. [12:43] But really, generative AI should be thought about [12:48] as a tool that just makes us superhuman in a sense. [12:52] But it's never going to be an independent active berserk Terminator. [13:02] You also heard it here first. [13:05] You laid out in your... [13:07] What do you call it, your strategy paper or your launch paper? [13:09] It's the leaked strategic memo, destinated to our seeding investors. [13:13] Oh, it was leaked? I assumed it was released. [13:15] It was leaked... [13:17] I'm sorry, I feel now that I'm kind of rooting around your underwear drawer. [13:25] We'll have more material defining our strategy coming. [13:29] Well, I look forward to reading that also. [13:32] Anyway, you lay out in that [13:34] you would like to have the best technology within four years. [13:38] You know, you lay out some of these milestones, [13:40] the best open source model by Q2 next year. [13:44] Are we still on track for that? [13:47] We are very much on track with that. [13:48] We are starting with... basically models are [13:52] on a cost performance curve. [13:53] And the higher it costs, the, the best performance you get. [13:58] And we are going to go from left to right on this curve. [14:02] So starting with some small models [14:04] and we are currently training much bigger models. [14:06] And are small models targeted at on-device usage? [14:10] On-device is a bit ambitious, but it can run on some. [14:15] Okay, small, just for cheaper inference. [14:16] It can run on a MacBook Pro pretty well. [14:20] It's an expensive device. [14:21] And then you said within four years having the best technology. [14:24] Is that a well-defined goal? [14:26] Because presumably, you know, as the AI space gets bigger, [14:32] there will be a lot of niches, a lot of subspecialties. [14:34] And so do you think about the best technology being [14:37] the best performing text model or how do you define that? [14:40] So it's a very open space. [14:42] The use cases we are currently discovering, [14:47] I think we do have a good understanding of the basic use cases like, [14:51] knowledge discovery, marketing, customer assistance, [14:55] the hard use cases, which is empowering research and development departments, [15:01] for instance, is still very open. [15:03] And so I can't tell the future [15:06] whether it's going to be one big model that can do everything, [15:09] or whether we'll need to scatter into niches. [15:14] But in any case, we are there [15:17] to make models that are very good in a very big part of the market. [15:22] What has surprised you since founding this company? [15:26] I was surprised at the speed of our employees because [15:31] I didn't expect to... we have models that are pretty good at chatting, [15:37] and I wasn't expecting it for the size of the models we targeted [15:41] for our first trainings. And so this was a pleasant surprise. [15:47] I'm still surprised at the speed of the field in general, [15:49] you see things happening every day, and it's a challenge [15:54] to stay focused on what we are building. [15:58] But I think we do a good job at that. [16:02] We'll go to audience questions in a second. [16:05] Do folks have questions? Yes, I see we have. [16:08] Actually we'll go to them now. Question from over here. [16:12] You talk about models running on device versus in the cloud. [16:17] During the 2010s a lot of the computing has gone to the cloud. [16:21] But it feels like it's unsustainable for generative AI at scale. [16:25] What's your thought process about that? [16:28] The question was, just for anyone who didn't hear it: [16:31] We've seen in the 2010s everything moving to the cloud, [16:35] you're talking about on device. [16:36] How do you think about the need for on device or that trend? [16:40] I expect that we'll see the two groups grow, [16:43] because when you have very large privacy concerns, [16:46] it can be good to be completely on device. [16:49] I think in the current state of the technology, [16:52] the models that are being used are much too big, [16:55] but they can be made much smaller for a specific use case. [16:59] So as long as you can make a model smaller than a few billion parameters [17:03] and you can make it solve your task, then you can deploy it on device. [17:07] And I expect that this will happen sooner rather than later. [17:10] The speed of compression is actually faster than what we expected. [17:15] There's probably a number of engineers in this audience who do not currently [17:19] work in AI, but are interested in getting into it. [17:22] If you're a software engineer curious about becoming good at AI, [17:29] curious about learning how to dabble in this, [17:31] where would you encourage people to go? What would you encourage them to do? [17:36] So I think it's very good to start with the basics. [17:40] So current AI is built on deep learning. [17:43] So if you want to understand what's going on, [17:46] you should do a little bit of deep learning first. [17:50] Say to train your own. [17:51] Train your own small model, try to understand what's going on, [17:54] how the technology is made, what are the limitations of it. [17:58] Before going in full engineering mode, [18:02] you do need to understand the underlying technology behind it. [18:05] And it seems very easy today because [18:08] the models are pretty well controllable and you can quickly make [18:11] an application out of it. But if you want to make a strong and robust application, [18:15] you do need to understand better the underlying parts. [18:19] When I cook, I follow the recipe very closely and weigh everything. [18:24] And I'm very careful and follow the steps and my girlfriend, [18:28] who is a much better cook, when she cooks, she just throws everything together. [18:31] And she knows things and she knows what the right flavors are [18:34] and how long things need. And, you know, yes, [18:36] you can just skip that step and everything. [18:39] And it feels like in AI when you say it's useful to build a small model yourself, [18:44] a lot of the people who are very good just kind of know things in that same way, [18:49] and they know how to tune the hyperparameters, [18:51] and they know that this approach won't work, and they have a lot of intuitions. [18:55] How do you develop the intuitions to become a better model chef? [18:58] It takes time. I think it's the biggest value in the good AI engineers [19:04] to have this intuition of what works, [19:06] what does not work. How do you make it better? How do you iterate? [19:08] How do you train a model? Look at the performance. [19:11] Decide what you should change in your training setting to make it better. [19:15] And that's what makes the value of the team that we hired, in particular, [19:19] all of these people, all of our employees built some experience in usually large [19:23] tech industries where they had access to a lot of compute [19:26] and they could build their intuitions. [19:28] So to build intuitions, unfortunately, I think it takes compute. [19:31] So it's nice if you can get a hand on a couple of GPUs. [19:34] But that's really important to be able to iterate fast [19:37] and to understand how the models actually work and how to make them better. [19:41] Okay. So start by building a toy model, [19:43] but there's no substitute for actually running something in production where [19:47] you need to make the performance better and you start to develop these intuitions. [19:51] Yeah, I mean, you have intuitions [19:52] or you can have intuitions in how do you train a model, [19:55] how do you orchestrate a model with retrieval augmentation, [19:59] for instance, you can also build a lot of intuition on that. [20:02] And for this you need to experiment. You need to do A/B testing. [20:06] You can build intuition on how to make it faster, [20:08] how to accelerate the inference part. [20:11] For this you need to understand very well how the hardware works, [20:15] because it's very tied to how GPUs are connected, [20:19] how GPUs use the memory. So you have in all of these parts, [20:23] you need to build some intuitive knowledge and it takes time. [20:27] It's not going to happen in two days. [20:30] I mean, I know the AI looks a bit like magic, [20:32] but in order to do robust things and move things into production, [20:36] you need a little patience. [20:38] It makes sense. Sorry, we had a question up here. [20:41] Why do you bet on open source for AI? [20:47] We think it's the best way of being cost efficient [20:50] and of breaking the barriers that are still there in the proprietary technology. [20:56] So breaking the problem of hallucinations, making the model smaller, more efficient, [21:01] making them easier to align. [21:03] We think that if you look at the history of software, [21:06] this has always been how it ended up with barriers. [21:10] They were owned by proprietary big technology company. [21:14] They ended up being open source. Operating systems: the same. [21:18] And I think the most recent example is cloud orchestration. [21:23] So Kubernetes is open source. [21:25] Today we are on the brink of a new revolution. [21:29] And as always, the big proprietary solutions were there first. [21:33] And we're very glad they did because they showed [21:35] the world how generative AI worked. [21:37] But we think there's so much to be taken from the community, [21:41] from cost sharing across different actors that open source is going to prevail. [21:46] And that's what we are here for. [21:48] It also seems it's in response to very strong customer demand, [21:52] where what everyone wants is better performance and lower cost [21:56] and kind of less vendor lock in. [21:58] And right now, many of the companies are focusing on the first [22:01] and no dedicated AI companies, except for Mistral, [22:05] are focused on the second of lowering costs and reducing lock in. [22:08] Yeah, it's very true. And based on our customer exploration, [22:11] we've seen that most businesses are still in the proof of concept phase. [22:16] So they don't have that much cost. [22:18] They are not too cost sensitive. When we move into production, [22:23] when those enterprises decide to deploy to their end users [22:27] some generative AI solutions, they will need to control the cost. [22:31] They will need to control the infrastructure. [22:33] They will need much finer control on the model. [22:35] They will need to make them smaller and adapt it to their use case. [22:38] And so I think that's where open models are going to have a lot of value. [22:44] All right, your question. [22:47] You mentioned that your differentiator is that your models can be fine tuned, [22:51] you can adjust weight, etc. [22:53] Do you plan on building some kind of abstraction on top of that, [22:58] or is your target persona mostly machine learning engineers? [23:02] So our MVP has some deployment software on top of it. [23:07] So you can easily spin up an API internally and have full control over it. [23:11] We are strengthening our software arm, software-as-a-service arm, [23:15] to actually propose some customization on specific data flows. [23:19] And this will happen. But we started with the models because that's what [23:23] the team was the best at. So we start with very strong models [23:27] and progressively we'll build some developer-oriented [23:30] orchestration stack on top of it. [23:34] Over here. [23:36] Hi. Thanks for the panel. 15 minutes ago, OpenAI announced that ChatGPT accepts [23:43] images and voice as input and voice output also. [23:49] Does Mistral have any plan to get into multimodal models? [23:55] And open source seems to lag a bit in that field. [23:59] Yes, we are focusing on text right now. [24:03] I was personally working on a lot of multimodal things [24:06] when I was at DeepMind. [24:08] The architecture behind GPT-4 vision is probably one of the things [24:12] that I developed. [24:14] So we will build that. Obviously, we are a very lean team [24:18] and we are focusing on the text first. [24:20] It's actually the hardest part and the most compute intensive part [24:24] because, as Mathieu can testimony, [24:28] it's actually cheaper to train models on vision than on text. [24:31] And we will move to that eventually. [24:34] And we are aware that there's a huge gap in the open source world, [24:38] on multimodal models. [24:40] And that's a gap that we will want to fill at some point. [24:44] It also feels like for the industrial use cases that you talked a lot about [24:48] in your leaked strategy paper, and, you know, [24:51] text is a useful encoding layer for a lot of other cognition, can be like, [24:55] I always like the demo of, you know, [24:57] you can use text to play chess [25:00] with the AI model where you just say moves in the written notation. [25:04] And they're not amazing chess players, [25:06] but they know the rules and they can play, [25:09] and they play like a kid would play. [25:10] And so I think this is a good example of text is just our encoding [25:15] for a lot of other thinking that is happening. [25:17] Yeah, that's for sure. [25:18] You can play chess with these models, but it's not the best architecture. [25:24] I know, but it's just like the text is an encoding layer for the cognition. [25:28] Yeah, and the way we think about open weights model is that [25:32] it provides basically a decoder. So it's the text oriented part. [25:35] But if you want to do specific workloads that involve images, that involve chess, [25:40] you probably want to connect them to existing deep learning models. [25:44] So you should see the decoder part, as long as you have access to the weights, [25:47] as a gateway to any kind of data workload that you want to run. [25:53] But putting everything in the single decoder through a text bottleneck [25:57] is very unlikely to be optimal in the chess part, in particular. [26:00] Chess shows that very well, [26:02] where after a certain point they make an illegal move. [26:05] You had a question. [26:07] Arthur, John, thanks for the panel. My name is DK, [26:10] I'm building a company called MarqVision. [26:12] We are building a AI powered product that helps companies to remove [26:17] and enforce counterfeits in the online space. [26:19] In fact, we're backed by YC, just like you did, John. [26:22] So my question is for you, John. [26:25] At YC, we were taught to build products that the customers love. [26:29] That's the number one, the utmost priority of all startup founders. [26:33] I was wondering if you could share some of your experiences when you were doing [26:36] go to market, expanding outside the US, [26:39] when you ran into the struggles [26:41] and hurdles of meeting that expectation of global customers, [26:44] that would be very helpful. [26:46] I think the biggest thing is just... you are the product's... [26:51] It depends on the space. But certainly for us with Stripe, [26:55] the product we built for the US market was not the right product [26:59] for the European market. And so as we were expanding to the European market, [27:02] all the considerations around needing to handle separate payments and mandates [27:06] and VAT and everything like that were all totally different from [27:11] what American customers want. [27:13] And so that's why we have engineering here in Europe, [27:17] and there is a product development cycle that is very important [27:21] in terms of getting close to your customers, [27:23] because the first version of a product you ship [27:25] will probably not be that compelling, and you have to constantly refine it [27:29] based on what you're hearing from customers to actually make it good. [27:32] But I think you shouldn't expect to just be able to take [27:35] the product that you launch in one market and have it succeed in the other market. [27:38] And again, that's what I found very interesting in your paper, [27:41] was this notion of kind of a European AI [27:44] and this very strong kind of European view that this is what European customers [27:48] would want. I thought it was striking how much that came up in the paper. [27:52] Yeah. I think the European angle we have is really very related [27:57] to the different business model we want to create, which is around open models. [28:01] Currently, we are seeing on the US side some kind of oligopoly shaping up, [28:07] you have Google on one side, Microsoft on the other side. [28:11] They are talking a lot with the White House, [28:14] a little bit doing regulatory capture. [28:16] And we think that in Europe, [28:18] if we don't do something about it in the coming year, we will be too late. [28:24] And that's why we started Mistral AI [28:26] and we started it with our conviction that a different approach [28:30] to the business was possible and that an open core [28:34] business model was actually valid. [28:36] And it also feels like a reflection of the fact that people are [28:39] so aware right now and so hung up on what is happening with data [28:43] on individual people. And then you have AI come along into this environment, [28:48] this feels relevant there. [28:51] Yeah. I think that if you're concerned about data privacy, [28:57] you should be concerned about sending it to [29:02] or sending it without control to a third party [29:06] and eventually, especially in production. So for proof of concept, it's okay. [29:09] I mean, you can, but if you want to move into production, [29:13] you definitely want to remove this liability at some point. [29:15] What is the technical solution? [29:18] You have large language models trained on data from around the Internet. [29:22] Then you have the GDPR right to be forgotten. [29:26] It's a tough question. It's a tough question: I think... [29:31] the GDPR wasn't designed with language models in mind. [29:34] We do our best, but obviously statistical things, [29:38] but we do our best to remove user data from what we find, [29:42] from the data we gather. But it's a tough question, I believe. [29:47] And one solution is to have a filtering stack. [29:50] So a way of verifying whether [29:53] the model is not outputting anything that is user related. [29:57] And I think it's the most robust way of building things. [30:00] That make sense. More questions... [30:04] We'll go to the back in the center. [30:06] Hi! Regarding the new models you're training: are you training [30:10] from scratch or are you using an existing model [30:12] and you are basically training on top of it? [30:14] So we are training from scratch. [30:16] And regarding the architecture, [30:18] what's your view about GPT like architecture versus new components more [30:23] exotic that are being now developed and researched? [30:26] Decoder architecture is pretty strong and very hard to beat, [30:29] if you know how to to train them. [30:32] Do you have any different perspective on scraping and how you scrape the internet? [30:37] The construction of a data set is part of a proprietary, [30:40] very strong asset in our proprietary tech stack. [30:43] So there's some stuff that you can't tell us. [30:47] No, I mean, it's very important to... [30:50] on the internet you have mostly garbage [30:52] and it's very important to take this garbage and distill it very heavily [30:56] so that you do get the things that are actually interesting for a model. [31:00] And this is something that we've been pretty good at doing. [31:06] Especially regarding trustworthiness [31:09] and reliability of generative language models. [31:13] You've mentioned retrieval augmented generation. [31:16] I was wondering if you see any future and perspectives [31:19] on neuro symbolic approaches or other. [31:22] I mean, other approaches incorporating knowledge graphs or some other technology. [31:29] No, I'm not sure about that. [31:35] I think this kind of exotic approaches are very late to the party [31:44] and deep learning basically makes things work. [31:47] And that's what we are betting on. I'm not saying that it doesn't work. [31:50] I'm saying that we won't be working on it. [31:52] Yeah. Okay. Thank you. [31:55] Question here at the front. [31:56] So in a few days, there's going to be a first release from Mistral. [32:01] Is that going to be just like an upload to Hugging Face or something, [32:04] or will there be an interface? How is it going to look like? [32:08] Or can you just not talk about it? [32:10] It will be a surprise. [32:11] All right. Cool. [32:13] Hello, thank you for the panel. [32:15] I have a more philosophical question. [32:17] So I would say, how would you define human consciousness? [32:21] Do you think that humans will ever be capable of creating artificial [32:25] consciousness? And if not, what do you think it's the biggest barrier to that? [32:30] You're talking to the wrong person. [32:33] I guess I'm not able to give any interesting insight on that. [32:40] If the question is whether large language models are conscious, [32:44] I tend to say that they are not. They are mostly like soft key value [32:48] network that is storing knowledge and retrieving it at runtime. [32:54] the question whether there's some internal world representation in it, [32:58] I think it's open, but it is quite unlikely. [33:01] There are some very interesting papers that came out this year, and this week, [33:06] showing that really large language models understand nothing [33:10] and are mostly about interpolating sentences, [33:15] which is the way we train them, so it's not that surprising. [33:20] I think we are pretty far away from an artificial consciousness if you ask me. [33:26] That is the beginning. [33:27] It is interesting though. [33:30] Everyone used to talk about the Turing test, and then we just blew past [33:34] the commonly used definition of the Turing test without much commentary on that. [33:40] Do you have thoughts on what a good new Turing test would be? [33:44] Yeah, I think this relates to the remark I did on regulation, [33:47] when you have a technology that grows, [33:49] you need to grow the difficulty of the benchmarks. [33:52] And obviously the Turing test is not good enough to measure consciousness. [33:57] But back in the 70s, it's understandable that one could not have a... [34:03] Could not have an idea of what could trick the test. [34:07] And so we need some better benchmarks. [34:11] And we need some better benchmarks also to compare the models [34:14] and the different solutions, because right now you have all of the solutions, [34:17] all of our competitors that produce things. [34:19] But it's pretty hard to actually compare actively what's going on [34:24] and what solution should be taken. [34:26] And I really hope that the space is going to be clearer as we grow, [34:30] as we become more mature. [34:31] Got it. I guess... [34:34] we still have a pretty good version of the Turing test. [34:36] If the AI is inside the box asking to be let out, you just ask it [34:39] something that happened after its knowledge cut off [34:41] and it will give itself away. [34:43] Yeah, that's a very good Turing test. [34:44] But again, this is something you can beat if you put some retrieval augmentation. [34:49] Will you guys fix the knowledge cutoff problem? [34:52] It's pretty much fixable if you plug it to the internet. [34:55] So if you make the model able to query things and then put it in the context [34:58] and use it. It's not perfect, but it works pretty well. [35:02] And so it would probably trick your Turing test. [35:04] You probably want to think of something even harder. [35:07] But you could probably tell what the difference is in latency. [35:09] Yeah, obviously, you can do that. [35:12] You could look at the network traffic coming out of the box. [35:14] More questions? I have heard questions coming from over here. [35:19] Hello. So what's your take on LLM versus specialized models? [35:24] I think LLMs are meant to be specialized to your use case. [35:28] So as it turns out, you can specialize them on many use cases and they can solve [35:33] basic use cases without changing the weights, which is good. [35:37] But if you have a specific thing in mind [35:38] it's always going to be better, if you have data, [35:42] you should put this data in your model in some way in order to make it better, [35:47] to make it smaller with the same performance. [35:49] So specialization is indeed needed for many use cases, [35:53] and especially as we move to production. [35:56] It's not to say that a model which has been specialized to 200 tasks [36:02] is not going to be able to solve these 200 basic tasks, [36:05] but as long as you have a specific task, [36:07] and I think most enterprises have very specific tasks [36:09] that are outside of the training set, then you should figure out a way [36:13] of putting this knowledge and the comprehension of the task [36:17] into the model itself. [36:20] Okay, I'll do one more question... Lady, all the way at the back. [36:25] Thank you for the panel. [36:27] Why did you decide to do the leap from academia to the industry? [36:33] And how do you forecast or how would you like to envision [36:36] the collaboration of academia and industry in the future? [36:40] So I took the leap during Covid. [36:44] So I wanted to not be alone and work with the team. [36:47] So that's my personal story. [36:50] I think there are a lot of very good collaboration ways between the academia [36:56] and the industry. [36:57] In particular on the inference part. [37:00] So how do you take the model and make it an API? [37:02] You see many very interesting things, [37:05] interesting software that is very efficient coming from Stanford. [37:11] And we're working with them to make it better [37:15] and to integrate it into our own solutions. [37:18] So on many parts of the stack [37:21] and in particular on the non-too-costly part, [37:24] on the parts that are not too compute intensive, [37:27] you have a lot of interesting ideas that come out of the academia, [37:31] and that's something that we want to foster, [37:33] and that's something that we also want to give to. [37:37] So when we give models to everyone with permissive licenses, [37:41] we also expect that academia is going to pick it up [37:44] and to bring new ways of aligning it, of instructing it, [37:48] of specializing the models. [37:50] I think we'll have to leave it there, but I just want to say thank you, Arthur. [37:53] This is the most exciting technology trend that I think we've had, [37:57] at least in the time I have been working in technology. [37:59] And so it's amazing that so much of it is happening here in Paris. [38:03] We're obviously very excited about it from a product point of view, [38:06] as you saw in the demos earlier and a lot of the stuff we have going on. [38:11] It's very exciting to see what you're doing. [38:13] And I can't wait to see the surprise you have for us later this week. [38:17] And so when should we be setting our calendars to check in on the internet? [38:22] Just out of curiosity. [38:24] The precise day is also a surprise. [38:28] It's a great week for AI and lots of good surprises. [38:32] Thank you, John!