Ramin Hasani：302 神经元线虫如何催生 Liquid AI

Alessio

我们在远程棚，请来新任 Head of Editorial Richard——打个招呼。

Okay, we’re in a remote studio with our new head of editorial, Richard. Say hi.

Richard

大家好。我从英国连线。

Hi, everyone. Dialing in from the UK.

Alessio

Ramin，Liquid 的 CEO，欢迎。

And Ramin, CEO of Liquid. Welcome.

Ramin Hasani

谢谢邀请。

Thanks for having me.

swyx

很多人听过 Liquid，知道 LFM，或听过 LFS，却未必亲自试过。公开叙事里还有一点：灵感来自线虫 DNA 之类——我通常当类比，但你可以告诉我们该当真到什么程度。

Most people when they’ve heard about Liquid know about LFM, or they’ve heard of LFS. They maybe haven’t personally tried it. The other interesting public data point is that it is very inspired by worm DNA or something. I never really pay much attention to that because I think that’s an analogy—but you can tell us how seriously to take it.

Ramin Hasani

完全可以当真。大约十二年前我们开始这项研究，想把大脑的启发带回机器学习——一切从这里起。说「我们」，是我和现任 CTO：在维也纳工业大学（TU Wien）起步；我读博，联合创始人当时是硕士生。我们想把连续时间数学放进机器学习系统，于是去看神经元如何交换信息——那由非常简朴的一阶微分方程支配，也正是神经系统在曲线神经元与突触层面的动力学。

Yeah, absolutely. When we started our research about 12 years ago, we wanted to see if we can bring inspirations from brains back to machine learning. That’s where everything got started. When I say we—my current CTO and myself—we started back at Vienna University of Technology. I started my PhD; my co-founder was a master’s student there. We wanted to bring continuous-time math inside machine learning systems. We looked into how neurons exchange information with each other, governed by very simplistic first-order differential equations—that’s how nervous-system dynamics work at the level of neurons and synapses.

Ramin Hasani

我们本想从人脑入手，但人脑是发放型神经元，很难学——对人脑里「竞争」如何发生我们其实仍知之甚少；果蝇脑知道得更多，人脑仍不够。我们按第一性原理找到了这种动物：秀丽隐杆线虫（C. elegans），全身只有 302 个神经细胞，体长约 2 毫米。美妙之处是：它的神经元行为很像人工神经元——它们不发放峰电位。神经科学里叫神经元的电紧张（electrotonic）行为：分级、连续。若系统受这种蠕虫脑启发，就能做非常连续时间、且可微的系统，可以反传。这就是当时的模式生物。

We wanted to start with the human brain, but the human brain has spiking neurons and it’s very difficult to learn—we still don’t know that much about how computation happens in the human brain. We know more about a fruit fly’s brain. We got to this animal from first principles: C. elegans, which has only 302 nerve cells in its body, about 2 mm long. One nice thing: its neurons behave very similar to artificial neurons—they don’t spike. In neuroscience they call it electrotonic behavior: graded, non-spiking. If you build systems inspired by this worm’s brain, you can build continuous-time systems that are differentiable, so you can do backprop. This was the model organism we were looking at.

Ramin Hasani

最令我们着迷的是：凭 302 个神经元，它控制约 95 块肌肉细胞，比我们当时（2016、2017）行星上任何机器人系统都好。另一点：约六亿年前进化树上人类与这种蠕虫分道；神经科学家对它脑与细胞结构的发现，在一定程度上可迁移到人类——与人类基因组约有 78% 相似性。它已为人类拿下四项诺贝尔奖，是极受欢迎的模式动物。我们甚至与研究这些神经元的神经科学家合作。简言之：我们按对蠕虫脑的了解，数学建模两个神经元如何交换信息，再建成所谓 liquid neural networks——liquid time-constant neural networks。「Liquid」强调灵活性：前向是微分方程支配的高度非线性系统，又可微，能很好学习表征。

The most fascinating thing: with 302 neurons it could control 95 muscle cells better than any robotic systems we had on the planet—this is 2016 and 2017. Another interesting fact: about 600 million years ago humans split from this worm in the tree of evolution, so discoveries about its brain and cell structure are transferable to humans to some extent—78% similarity to the human genome. So far it has won four Nobel prizes for humanity. We even worked with neuroscientists on these neurons. Long story short: we mathematically modeled how two neurons exchange information from what we know about the worm’s brain, and built liquid neural networks—liquid time-constant neural networks. “Liquid” was basically for flexibility: highly nonlinear systems governed by differential equations in the forward pass, yet differentiable, really strong representation learners.

Ramin Hasani

我们意识到：用这类网络的一小撮——大约几十到几百个神经元——就能像蠕虫那样控制机器人。机器人上算力有限，能把足够智能塞进最小算力单元，极有吸引力；效率与「把智能塞进最小计算单元」成了持续课题。2017 年，MIT CSAIL 的 Daniela Rus 教授联系我在维也纳的导师——她也是联合创始人，可以说是机器人学之母，对分布式与软体机器人贡献巨大。她被我们的研究震惊，问我们愿不愿来 MIT，把技术用到真车、无人机等载体上。Matthias 和我加入 MIT；我还遇见第四位联合创始人 Alexander Amini（当时还是 MIT 博士生）。自 2017 年起四人一起做连续时间动力系统与连续时间学习，并扩展到多领域——序列建模成了我们的主线。

With a small subset of these networks—tens to hundreds of neurons—you can control robots very similarly to how the worm did it. On a robot you don’t have that much compute; packing that much intelligence into the smallest unit of compute was very attractive. Efficiency and bringing intelligence into the smallest compute unit became a research topic for us. In 2017, Professor Daniela Rus of MIT CSAIL reached out to my professor in Vienna—she’s one of my co-founders as well, arguably the mother of robotics, with tremendous contributions to distributed and soft robots. She was blown away and asked if we’d consider coming to MIT to apply this to cars, drones, and many vehicles. Both Matthias and I joined MIT. I met my fourth co-founder, Alexander Amini, then a PhD student at MIT. The four of us since 2017 worked on continuous-time dynamical systems and continuous-time learning systems, scaling them across domains—sequence modeling became our thing.

Ramin Hasani

不只 MIT：Yoshua Bengio 实验室、David Duvenaud 实验室、多伦多大学、斯坦福等许多人也在推进连续时间动力系统。可以把它想成：循环网络是离散化的动力系统，一步一步算，步与步之间的时间差变化不大；连续时间系统则让你对「事件何时发生」有更多控制，离散事件之间的时间也被计入——对连续演化的序列，在时间维上适应性更强。复发与反馈计算，是我们带回机器学习的直觉。同时期还有 Neural ODEs、状态空间模型（SSM）等名字盖在这类动力系统上。

Not only from the MIT side—researchers from Yoshua Bengio’s lab, David Duvenaud’s lab, University of Toronto, Stanford, and others were contributing to continuous-time dynamical systems. Imagine recurrent nets as discretized dynamical systems computing one step at a time, where the time difference between steps doesn’t change much. Continuous-time dynamical systems give more control over when events happen and account for time between discrete events—more adaptability in the time dimension for continuously evolving sequences. Recurrence and feedback calculation were the intuitions we brought to ML. People also put names like Neural ODEs and state-space models on top of these dynamical systems.

Ramin Hasani

大约 2020 年人人谈可扩展性。MIT 这边从研究角度在和 OpenAI、Google 们比算法效率与机器人新学习场景的表现——2017 到 2020 OpenAI 在强化学习和 OpenAI Gym 类机器人上很活跃。我们的论题是：用连续时间算法控制机器人，展示小系统如何更好做分布外泛化。但扩展一开始，连续时间、高非线性的循环回路立刻变得计算上难处理。RNN 的大障碍是顺序计算；非线性循环算子在 GPU 上并行极难。这正是 Transformer 架构之美：无偏置、本质是矩阵乘，能扩展——硬件彩票，也是解锁通用行为的路径。

Around 2020 everybody was talking about scalability. From the MIT side we were competing with the OpenAIs and Googles of the world from a research point of view—efficient algorithms that perform well in new robotics learning scenarios. Between 2017 and 2020 OpenAI was very active in RL and robotics like OpenAI Gym. Our thesis was controlling robots with continuous-time algorithms and showcasing out-of-distribution generalization with small systems. But as soon as you start scaling continuous-time, recurrent, highly nonlinear loops, computation becomes really not tractable. RNNs’ major blocker has been sequential computation; parallelizing nonlinear recurrent operators on GPUs is really difficult. That’s one beauty of the transformer: unbiased, basically matrix multiplications you can scale—the hardware lottery—and unlocking general-purpose behavior.

Ramin Hasani

但复发仍极诱人：算力不会像注意力那样爆炸——吃更多数据时注意力的二次代价人人都看到了。复发把计算压得亚二次或接近线性，理论上更高效；可若不能并行做运算，就扩不起来。一条路是把非线性系统线性化，好用线性代数技巧并行——SSM 某种意义上是 RNN 的线性版，却失去参数间的非线性关系。没有免费午餐：线性化复杂动力学就丢表达力。用 SSM 做语言时长程依赖吃力——算得高效，表达却往往逊于 Transformer。于是我们从 liquid nets、动力系统、SSM、卷积等学到的各种算子，开始拼一台通用计算机。

But recurrence is so attractive because computation doesn’t explode—when you consume more data, the quadratic cost of attention is something people have paid attention to. Recurrence brings computation lower, sub-quadratic or closer to linear as you consume more data. Theoretically more efficient—but if you cannot parallelize operations, you cannot scale. One way: linearize nonlinear systems so you can apply linear-algebra tricks in parallel. State-space models are a linear version of RNNs; they lose that nonlinear relationship between parameters. No free lunch: when you linearize complicated dynamics you lose expressivity. Applied to language, SSMs have trouble learning long-context dependencies—they compute efficiently but their expressive power falls short of transformers. With everything we learned from liquid nets, dynamical systems, SSMs, CNNs—many operators—we thought about building a general-purpose computer.

Ramin Hasani

Liquid 第一天——大约三年前——我们不想押单一「圣杯」架构。我们把另类架构的「复仇者」聚到一处：来自 Bengio 实验室、斯坦福、MIT 的发明者——Hyena Hierarchy、Jimmy Smith 等的 SSM、我们自己的 liquid nets。目标是建一个无偏的元 AI，按部署环境在感兴趣的算子空间里搜索。想部署在 GPU 或 NPU 上，就对构建混合架构做大规模搜索。优化四件事：一、相对纯 Transformer 不牺牲质量（推理、语言理解、知识等约上百条标准）；二、最小化内存；三、最小化延迟；四、最大化计算速度。相关论文可搜 STAR Liquid AI——synthetic design of tailored hybrid architectures。这套元算法有点像递归自我改进：三年半前我们用它为给定芯片搜混合架构，其计算图成了历代 Liquid Foundation Model 的骨架。

Day one at Liquid—about three years ago—we didn’t want to bet on a single holy-grail architecture. We brought Avengers of alternative architectures into one place: people from Yoshua Bengio’s lab, Stanford, MIT—inventors of Hyena Hierarchy, state-space models like Jimmy Smith of Stanford, and ourselves on liquid neural networks. Let’s build an unbiased meta-AI that searches through the operator space given the deployment environment. Imagine deploying on a GPU or NPU—we applied a massive search algorithm on top of building hybrid architectures optimizing four things: no sacrifice on quality versus a pure transformer (reasoning, language understanding, knowledge—on the order of 100 criteria); minimizing memory; minimizing latency; maximizing speed of computation. There’s a paper—search “STAR Liquid AI”—synthetic design of tailored hybrid architectures. That meta-algorithm is kind of a recursive self-improving system. Three and a half years ago we applied it to search hybrid architectures for a given chip; that became the computational graph of different generations of liquid foundation models.

Ramin Hasani

LFM1 是架构搜索的第一实例：有卷积、注意力、也有原始 liquid 成分——什么都有一点的「大杂烩」。第二代专为 CPU 跑得极好优化：LFM2 架构很简洁优雅——大约 80% 双门控一维卷积，加上大约 20% grouped-query attention，又快又高质量，能在 CPU 上跑得很好。架构搜索之美在于：算子空间有新发明就加进元搜索；每一代 LFM 相对上代必有改进，但选哪些算子、目标设备暗示何种混合，由搜索决定。今天我们在 LFM 2.5；我们是从零预训练、自己跑 scaling laws 的基金会模型实验室，规模大约从数千万参数到 70B。目前开源/发布的是较小的 LFM2 与 LFM 2.5；下一代 LFM3 会再改一改，加质量、稳定性，以及在我们关心的部署框架上更好的延迟控制。

LFM version one was the first instance out of this architectural search—elements of convolution, attention, and original liquid; a mess of everything. The second generation we optimized specifically for running really well on CPUs. LFM2 has a very elegant, simple architecture: basically 80% double-gated convolutions—1D convolutions—plus some grouped-query attention, basically 20%. Very fast, high-quality foundation models you can run on a CPU really well. The beauty of architecture search: when a new invention in the architecture space happens, we add it to this massive meta-AI search. Every generation of LFM has improvement over the previous; operators and the target device imply what the hybrid should be. Today we’re at the LFM 2.5 instances. We’re a foundation-model lab building everything from scratch—we pre-train, run scaling laws, scale from tens of millions of parameters now to 70 billion. So far we released smaller LFM2 and LFM 2.5 instances; next is LFM3—slightly modified for more quality, stability, and better latency control on the deployment frameworks we care about.

swyx

你刚给了我们一份 Liquid 的「博士摘要」。业务侧也更新一下：上次播客讲 Liquid 是和 Shopify 那边的朋友——他很粉你们。这不只是研究项目；人在真正用，基准也不是空刷。

You just gave us a PhD abstract of what Liquid has. For people on the business side—last time we covered Liquid on the podcast was with someone from Shopify who’s a big fan of you guys. This is not just a research project. People are actually effectively using this. The benchmarks are not just goal-seeks; they’re reflective of the world.

Ramin Hasani

绝对是。Shopify 那位（ASR：Mikuel）现在像我的导师之一——行业里少有的 no-BS 型，建议极直接。他说：我们 Shopify 有很强团队也在建模型；你说模型好，我们团队要亲眼测。他们第一次测就对质量和效率正惊喜。在另类架构空间里，我们模型确实很强。业务上，我们优化含 CPU，要把智能带出数据中心——与关心「数据中心之外部署」的企业合作：车厂、机器人厂商、笔记本厂商、手机公司。它们在数据中心外有算力；我们的基金会模型能在这些设备上带来一流智能。

Absolutely. Mikuel is one of my mentors now—one of those no-BS people in the industry; advice very direct. He told us: we have a very competent team at Shopify building models; if you tell me you have models, our team has to see and test whether it’s good or bad. They tested the first time and were positively surprised by the quality and efficiencies. In alternative architectures we have really good models. On the business side, we’re optimizing for CPUs too—bringing intelligence outside of data centers. We started working with enterprises that care about deployments outside data centers: car manufacturers, robot manufacturers, laptop producers, mobile companies. They have compute somewhere outside data centers; our foundation models bring the best of intelligence onto those devices for various applications.

Ramin Hasani

伙伴例子：AMD 在 AI PC 侧。全球每年大约产生 3 亿台 AI PC；笔记本大约 3 到 4 亿台，其中大约 1 亿可能是 AMD 核心。我们正把 LFM 直接跑在笔记本上，做本地智能：例如隐私过滤——你要往 ChatGPT 或云端 coding agent 发内容前，先过一层 LFM，智能挡掉敏感信息；也可做设备上的主动智能体，访问本机信息做分析，具备离线与气隙能力。今天的设备智能多是混合：云端一个模型、端上若干模型协作。苹果智能、Galaxy AI 这类最常用手机方案里，大约 90% 调用上云、10% 在端——因为端侧模型还不够强、不够可靠，扛不起更多云侧用例。

Examples of partners: AMD on the AI PC side. About 300 million AI PCs are generated on the planet every year; 300 to 400 million laptops, and from there about 100 million could be AMD-powered at the core. We’re bringing liquid foundation models directly onto the laptop for local deployment—privacy filters that smartly guardrail what wants to get out of your computer before you send to ChatGPT or a cloud coding agent; or proactive agents running on-device with access to information on your laptop, offline and air-gapped capabilities. Device intelligence today is hybrid: a cloud model plus on-device models collaborating. For the most-used phones—Apple Intelligence, Galaxy AI—about 90% of the calls go to the cloud, 10% on device, because device models aren’t powerful or reliable enough to take more off those cloud use cases.

Ramin Hasani

公司为何想把 AI 跑在端上？首要就是纯粹成本：若有约 4 亿部手机或数十亿客户要免费智能，而智能全在云端服务，成本会极高。端上跑 workload，等于给自己创造利润率。此外还有离线、隐私敏感等优势。但为设备建模型并做到生产级极难——不是从开源下个模型塞进去就成功。要做热管理；长上下文下模型行为会变，需要可靠度。过去约两年半我们经历多轮生产循环：生产级 AI 远不止在设备上跑几项质量基准。

Why do companies want AI on the device? First: sheer cost. If you have 400 million phones or billions of clients to support with free intelligence access, and you have to serve that intelligence in the cloud, it becomes really prohibitive. Bring the AI workload on device and you’re creating margins. Plus offline features, privacy-sensitive topics, and many advantages. But building a model for the device and getting it to production quality is extremely difficult—not just downloading a model from open source and putting it on the device. You have to do thermal control; behavior in long context actually changes; you need a degree of reliability. Over the last two and a half years we’ve gone through many cycles of production. Production-grade AI is way more than benchmarking some quality benchmarks on the device.

Ramin Hasani

再如车厂：Mercedes-Benz 是客户之一，我们做车内智能。去年四月宣布的合作在汽车行业算推进很快。坐在车里可以对话、「Hey Mercedes」等——这些由直接进车的基金会模型驱动。我们供电的是多模态方案：可接摄像头、做语音交互，并用文本模型做推理。模型大约 600MB——想象约 600MB 智能装进每辆车，还可 OTA 更新全球车队。今年首批规模部署会在北美所有 Mercedes-Benz Generation 3 车上；跑在大概一百美元成本的芯片上。把生产级高质量 AI 做到这种效率并进车，感觉非常好。

Another example: car manufacturers—Mercedes-Benz is one of our clients. We bring in-car intelligence. We announced a partnership last April that happened really fast in automotive. Sitting in your car, talking—“Hey Mercedes” features—those happen with a foundation model that goes directly inside the car. The model we power is a multimodal solution that can connect to cameras, do voice interchange, and has reasoning through text-based models. The model itself is about 600 megabytes—imagine 600 MB of intelligence inside every car, with over-the-air update of every car on the planet. First deployments on Mercedes this year will be on all North America Mercedes-Benz Generation 3 cars—first deployment at scale of AI actually inside the cars—running on a chip that is probably 100 bucks in cost.

Ramin Hasani

Shopify 是难以置信的伙伴——他们不做端侧 AI，但在乎延迟与质量，有私有部署服务 Shop App 的商家端与客户端。大量 LFM 已在生产：今天在 Shop App 打字搜索，信号会打到 Liquid foundation model。用量指数增长，现在已超过每月十亿次请求——仅这一条 Shop App 指标。模型本身也开源给社区，很受欢迎：总下载超过 4000 万，大约每周 150 万。已发布参数范围大约 1 亿到 24B，更大的是混合专家；并在加多模态——可同时吃音频、视觉、文本，输出音频与文本。

Shopify has been an unbelievable partner. They’re not doing on-device AI but care about latency and quality—private deployments serving the Shop app, merchant-facing and client-facing. We have a lot of liquid foundation models deployed in production. If you go to the Shop app today and type something, you send the signal to a liquid foundation model. We’re exponentially at over a billion requests per month—just that Shop app metric. We open-source liquid foundation models to the community; overall over 40 million downloads, about 1.5 million downloads per week. Released range so far: 100 million to 24 billion parameters—the bigger ones mixture-of-experts—and adding multimodal capabilities: processing audio, vision, and text as input, generating audio and text as output.

Richard

机器人是你们起步灵感之一。你们有视觉语言、音频等模型——是否用于机器人？有哪些实验？

I was wondering about robotics, because that was one of your inspirations when you started. You’ve got multimodal models—a vision-language model and audio models. Are those being used for robotics purposes, and can you talk about some of those experiments?

Ramin Hasani

可以。视觉语言模型能做视觉—动作：机器人看世界再据此行动；也可做监控——工厂摄像头盯整条产线。我们多想工业机器人而非人形。与 Roboe Tech（ASR）等合作过，AMD 与 Liquid 有 demo；意大利的 Gbionics 等用我们的 VLM 做模块化机器人身体上的动作空间与指令跟随。低延迟机器人场景里，较小实例很受欢迎。但第一天我们刻意把机器人业务按下：机器人的 go-to-market 难、周期长，生产级要验证、确认再进监管。我们自己从机器人出来，知道产品化有多难，所以先做 OEM 可落地的水平智能层，动作世界里验证更敏感的部分放后。现在模型更成熟，对数据与后训练/定制控制更好，是更认真做机器人垂直的好时机。

Absolutely. With vision-language models you can solve vision-action problems—robots seeing the world and taking action—and use them as monitors for factory cameras. We usually think about industrial robots, not mostly humanoids. We worked with a company called Roboe Tech—search Roboe Tech, AMD, and Liquid AI for a demo—and Gbionics in Italy; modular robots using our VLM for action space and instruction-following. Smaller instances are really popular for low-latency robotics. But robotics we deliberately put on pause day one: go-to-market is difficult and delayed—production-grade needs verification, validation, then regulatory aspects. Coming from a robotics background, we know how hard it is to productize a robotics vertical. We wanted a horizontal intelligence layer where we can have OEMs first, then enter the action world where V&V is more sensitive later. Now models are more mature, with more control over data and post-training/customization—it’s a good time to start the robotic vertical much more seriously.

swyx

机器人是不是还用同一套 LFM 家族？很多人把动作模型、世界模型说得很不同——多少是营销，多少真本质不同？

Do you basically use the same LFM family? A lot of people make a big deal about how different action models or world models are for robotics. How much of that is marketing versus fundamentally the same thing?

Ramin Hasani

好问题。基金会模型创新我分成三类：模型架构、算法、数据处理机制。我们此前谈的多是架构创新——效率与成为通用计算机。算法上：可用自回归损失、扩散、或世界建模（world modeling）去训同一架构——世界建模偏无监督表征学习，从中得到某种「世界视图」。你完全可以把 LFM 架构放进世界建模管线，学更好的物理理解。不同算法从物理世界抽取的基元不同；next-token prediction 也有物理理解，但与人类方式很不一样。机器人的「ChatGPT 时刻」在算法侧仍是大问号；架构侧则任何通用架构都可嵌入这些框架。

Great point. Innovations in foundation models I’d categorize in three classes: model architecture, algorithms, and data-processing mechanisms. What we talked about so far was model-architecture innovation—efficiency and becoming a general-purpose computer. Algorithms: you can train with an autoregressive loss, diffusion, or in a world-modeling context—world modeling as unsupervised representation learning so you have a world view. You could take LFM architectures into a world-modeling pipeline and train for better understanding of the physics of the world. Various algorithms extract different primitives from the physical world versus next-token prediction; next-token has physical understanding but very different from humans’. The ChatGPT moments for robots are still a big question mark on the algorithmic side; on the architecture side you could use any general-purpose architecture in those frameworks.

swyx

回商业模式：这些多是授权——气隙、端侧。定制要花多少时间？是你们人做，还是协作？

Before we go back to research—I assume these are all licensing agreements because it’s air-gapped, on-device. How much time do you spend to customize a model if I come with special compute constraints? Is it your guys doing it, a collaboration—how does this work?

Ramin Hasani

早期多是 design partner：我们派应用 ML 团队跟客户，按接触期收费；方案建成部署后，再按年、按设备收许可，视部署场景而定——模型是经常性成本。同一垂直做完后，下一客户服务时间下降，更多是方案交付。例如 Mercedes 是 design partner，汽车垂直后来者服务量减少。同时我们在建模型开发平台，目标自助：客户可自动完成过去人工定制。平台仍在 beta，我们很兴奋——定制与持续学习是部署下一关键。今天约 90% 市场还在推理 token：Frontier 实验室与 Fireworks 类推理商靠托管收推理。下一波会是定制 token——Rei、CoreWeave 自动化类、Thinking Machines、Trajectory 等；已有开源基座，企业可资本化定制。定制不只是后训练：从预训练、中训、后训、数据生成到 RL，再让智能体闭环，客户拥有自己的智能——真正的主权 AI，再往下是人人可拥有个人智能。

Early engagements are usually with design partners—we associate an applied ML team with the client and charge for that engagement duration; once the solution is built and deployed, we charge licenses on a yearly basis on a device basis depending on where it’s deployed—models as a recurrent cost. After we build a solution in a vertical, time of service reduces for the next customer—becomes solution deployment. If Mercedes-Benz is our design partner, next automotive customers get less service from our side. We’re building a platform for model development that’s self-serve so you can automatically do that job. It’s in beta right now—we’re still testing; customization and continual learning are a very important next step. Right now like 90% of the market is around inference tokens—Frontier Labs and inference providers like Fireworks host models and charge inference. The next wave is customization tokens—companies like Rei, Core automation, Thinking Machines, Trajectory. With good base open-source models, monetization of customization tokens is what people will capitalize on—next wave of decacorns; some already are; Jeff Dean just started a lab on similar topics. Customization isn’t only post-training—depths from pre-training to inference: pre-training, mid-training, post-training, data gen, RL—allowing agents to loop so customers own their own intelligence. That’s sovereign AI; one step further, every person owns their own personal intelligence.

Richard

开发者平台 LEAP 现在怎么用？部署流程是怎样的？

Before we head to the research—can I ask about your developer platform, LEAP? How are developers using this platform at the moment, and what process do they use to deploy your products?

Ramin Hasani

LEAP 像一个库加 cookbook——开源、给人用的。有大量洞见；例如 LEAP fine-tune 可微调我们的模型；也可导出推理就绪的 GGUF / llama.cpp 式 bundle，直接上设备做 CPU 部署。我们正在建的企业平台，本质上是把 LEAP 做过的事自动化、补齐缺失能力，也可交给你的云端 coding agent 去调工具微调。LFM 也接入 Unsloth、Hugging Face TRL 等；我们跟 Hugging Face、SGLang、llama.cpp 走得很近——llama.cpp 是把智能带出数据中心的了不起努力。

LEAP is like a library—we have a cookbook, an open-source place that got a lot of attention. It’s built for humans: fine-tune liquid foundation models there—LEAP fine-tune is one place—then deployments: LEAP lets you extract a bundle that’s inference-ready GGUF format, llama.cpp-ready, directly on device for CPU deployment. The platform we’re building now is literally the automated version of what we’ve been doing with LEAP—complete the tools, and give enterprises capability to use those tools or give them to your cloud agent to fine-tune. There’s also support in popular places like Unsloth and Hugging Face TRL. We try to stay connected to open source—we love the Hugging Face guys, the SGLang guys; we work closely with them. The llama.cpp project itself is an amazing effort to deploy intelligence outside of data centers.

swyx

「自助」让我眉毛一挑——我本以为你们会做很多服务，或如今所谓 forward deployed。有没有成熟度阶梯？比如量化感知蒸馏似乎容易自助，再往上走？

My eyebrows went up when you said self-serve. I expected a lot of services—maybe these days called forward deployed. Is there a maturity or hierarchy—quantization-aware distillation seems self-serve, then you work your way up?

Ramin Hasani

正中要害。我们大概接触过 Fortune 500 里约 200 家。成熟企业折腾自建智能很久：自组应用 ML、下开源、甚至从零训——多数没变成真正产线，最终退回云方案；大约 80% 的 PoC 到不了生产级。另一面：开发者疯狂要最强智能来写代码、token max，但成熟组织里 vibe coding 仓库归因后，大约 90% token 是无效的；企业要控成本，又要满足开发者，于是强烈需要能尽快兑现价值的自助平台。我们不做又一个新 UI：平台坐进客户环境，像你对接 Claude Code 一样，把 Liquid AI 工具塞进你喜欢的 agentic harness，逐步引导开发者交互式建成生产级 LFM——人在环很重要。也可「开自动档」一键交付，但质量通常不如交互式共建。

You touched on a very important problem. We’ve worked with probably 200 of the Fortune 500. Mature enterprises went through building their own intelligence for a long time—applied ML teams, downloading open-source, building from scratch—and this did not result in actual production lines; they defaulted back to cloud. The ratio: like 80% of production-grade PoCs do not hit production level. Second: massive demand from developers for the highest form of intelligence—everyone wants to vibe-code and token-max. When you attribute value in mature orgs, you end up with 90% of these tokens as useless tokens. Enterprises want to control cost while satisfying developer demand—massive demand for a self-serve platform to get to value ASAP. We don’t build another UI: the platform sits inside the client, similar to how you interface with Claude Code—give your agentic harness the tools of Liquid AI and walk step-by-step through developing, building, and deploying a production-quality LFM. Human-in-the-loop is very important; you can also put it on auto mode, one click—but quality won’t be as good as interactive building with a developer.

swyx

反驳一下：若我有好评测，为何还关心过程？它不该自己 auto-research 吗？

Let me push back. If I have good evals, why do I care what the process is? It should just auto-research itself.

Ramin Hasani

第一次部署我完全同意。但模型按当时评测上线后，新数据涌入、分布大迁移，社区想要的东西变了——静态评测会过时，必须有持续演化的系统。这也是行业走向环境、Harbor 式任务的原因。分工上不只是客户做后训练、我们做架构：我们希望客户能从零做完整模型开发——包括架构搜索。若你像 Qualcomm 一样成熟，想为下一代芯片、正在设计的 NPU 找到最佳计算图；也有公司把网络权重直接蚀刻进芯片。硅伙伴对我们极重要，因为架构搜索是硬件感知的——把下一代硬件放进环里找最优架构。这套自动化基金会设计平台硅厂很喜欢。企业拥有模型开发与部署，才是主权 AI 真义；再下一步是个人智能。

I fully agree for the first deployment. But once the model is in production based on the evals you had, new data comes in—massive distribution shift, community wanting different things. Static evals become obsolete; you need dynamism, a continuously evolving system. That’s why we’re moving toward environments and Harbor-style tasks. Division of labor: not just post-training—we talked about whole model development from scratch, including architecture search. If you’re as sophisticated as Qualcomm designing the next generation of chips and want a dedicated computation graph that works best on their NPU—or companies etching weights of a neural network directly inside chips—silicon partners matter because our architecture search is hardware-aware. They put next-generation hardware in the loop of finding the best architecture. Silicon companies really like this automated foundation-design platform. Enterprises owning model development and deployment—that’s the true meaning of sovereign AI; next level is personal intelligence.

swyx

神经架构搜索——你们大概是第二、三家我们深度聊过的；最近还有 poolside 的 model factory，以及 Apple 的 Taria 层析工具。搜索里有什么「以为有用却没用」或意外有效的？

On neural architecture search—maybe second or third company we’ve covered deeply; we recently covered Poolside’s model factory, and Apple’s Taria LM-ops layer-by-layer inspection tool. What have you found in architecture search that you thought would work doesn’t, or is surprisingly effective?

Ramin Hasani

对每个架构候选我们会跑 scaling law——固定尺寸扔更多数据，或增大模型——看候选训得怎样。学到的是：网络越大、若要通用智能，架构应越无偏——要给系统去偏置。LFM2 没有 gated delta net 之类；就是很「无结构」的一维卷积——能做的最简事物之一。注意力本身也很无结构；无偏置架构如 Transformer 才能扩上去。规模越大，架构里需要的偏置越少，要简化。另一端：极小规模、比如约 8GB 内存上限，反而可在架构上更有创意——多重反馈环、更多复杂性；小模型体制里偏置越多，动力学越有表达力；大了就把偏置拿掉。第二洞见是数据模态：音频上连续时间动力系统或 SSM 很有效，但在文本上 SSM 不行。只要音频模型，在所需规模扔循环网络，复发会很闪光——尤其大约 20B 参数以下。我们称这些为更有偏置的架构；放大时再拿掉。

When we do architecture search for foundation models, for every architecture candidate we run a scaling law—up to a certain number of tokens, throwing more data at a fixed size or increasing model size—and see how well it trains. What we learned: the larger the network, if you want general-purpose intelligence, it should be unbiased—you have to unbias your system. LFM2 doesn’t have gated delta nets and all these things; it’s literally 1D convolutions—the simplest unstructured thing. Attention itself is very unstructured; when architectures aren’t biased like transformers, you can scale them. At scale a lot less bias is needed—simplify as you make them larger. On the other side: very small scale, memory limit say 8 GB—you can get more creative: multiple feedback loops, more complexity. In the small-model regime, the more biases you add, the more expressive dynamics you get; at larger scale you remove those biases. Second insight: data modality. In audio you can use continuous-time dynamical systems or SSMs—very effective—but on text they suck. If you just want an audio model, throwing in a recurrent net at the scale needed for good audio output is where recurrence shines—again at small scale, below say 20 billion parameters. I’d call them more biased architectures; you remove those things as you scale.

swyx

小尺度加偏置、大了拿掉——我们有过类似讨论。层循环（looped layers）为何小尺度管用、大尺度不行，有点反直觉。

At small scale you add biases and eventually take them out. Layer looping—or looped layers—is a bit unintuitive why it works at small scale but not at larger scale.

Ramin Hasani

任何形式的 looping 仍是偏置，会把系统变复杂，还要找对基础设施。当前 GPU 基建为可并行的矩阵乘优化；想象光子学——若能以光速通信，串行计算的速度对人不可感知，非线性算子也许不再受限于硅基底。基底会塑造下一代架构；我约十四年前听 Christopher Koch 一类神经科学讨论说「智能是碳的属性」——自然生命在碳上，我们在硅上建智能，基底不同。

Looping is again a bias—any format of looping. You make it complicated and you’ve got to find infrastructure. Current GPU infrastructure is optimized for parallel matrix multiplication. Imagine photonics: if you communicate at the speed of light, the speed of serial computation moving at light speed is humanly not perceivable—that could unlock nonlinear operators. Substrate will play a role in the next generation of architectures. I read something like Christopher Koch years ago—intelligence as a property of carbon; natural life on a different substrate, we’re building intelligence on silicon. The base of intelligence is very different.

swyx

SF 圈子里不少人对量子计算不太兴奋。若真有规模化量子比特，会改什么？

A surprising number of people in my SF circles aren’t that excited by quantum computing. If we did have at scale a bunch of qubits, do you think it would change anything?

Ramin Hasani

我认为它会大幅改变的是搜索——搜索极其重要；若能并行搜索……优化归根结底是在参数空间做大规模搜索，我们现在屈从于梯度下降，是因为没有更好的规模化搜索。量子有望让搜索大进化。其他瓶颈量子解不了——同意。2017 年 NeurIPS 上 IBM 展出吊在天花板上的量子设备，我问何时人人能用，对方说 2035；现在说什么都还太早。

The only thing it changes massively would be search—search is one of those algorithms that matters so much. Imagine parallel search. What is optimization at the end? We’re submitting ourselves to gradient descent because we don’t know any better way to perform massive search at scale in parameter space. With quantum computers, search is going to evolve massively. There are many other bottlenecks not solved by quantum—absolutely. In 2017 at NeurIPS, IBM was showing off their quantum machine hanging from the ceiling; I asked when everybody can use something like this—he told me 2035. I think it’s still too early to say anything.

swyx

我们做过 RWKV、Tree Dow（Cartesia）等；你说 SSM 很贴声音——连续推断、状态不多，听来合理。若连续时间这条路错了，还有什么该挖的维度？

We’ve done RWKV and had Tree Dow on; Cartesia not yet. When you said SSMs are attuned to voice—continuously inferencing, not that much state—relatively obvious once explained. If we’re wrong about continuous time, what else?

Ramin Hasani

Liquid 的想法一开始就有两轴：智能的效率，与智能的适应性。适应性我们挖得还不够。很多人说上下文学习已解决：测试时塞更多数据，某种伪反传、像最小二乘。但能同时结合前向与后向的路径，会是新一代架构。听说 SSI 也在做接近原始 liquid neural networks 的想法。也要默认：Anthropic、OpenAI 等所有基金会实验室都有团队砸另类架构。我们要另类架构的第一理由是效率；别人现在多在做事后优化——推测解码、量化、量化训练——内核科学，更好的 AI 也能把现有架构压出效率。若要解锁下一代能力，则可能是元算法，或系统级：编排器加多智能体——认知架构，用神经元/突触、或一个 LFM、或小模型作计算单元，捆成系统的系统。智能体应能把训模型当日常工作的一部分；那需要的不只是 LFM，而是整套 LEAP 平台去适配部署情境。

When we started Liquid as an idea we were thinking about efficiency of intelligence and adaptability of intelligence. We haven’t explored too much on adaptability. A lot of people say in-context learning solves that—more data at test time, somehow a pseudo-backprop, least-squares-like. But approaches that combine forward pass and backward pass at the same time would be a new generation of architectures. I’ve heard SSI is working on similar architectures to the original ideas of liquid neural networks. Assume all foundation-model labs—Anthropics and OpenAIs—have teams on alternative architectures. Why alternative architecture? First for us: efficiency. Everybody else seems focused on post-hoc optimizations—speculative decoding, quantization, quantization of training—kernel science; with better AIs people can optimize current architectures for efficiency. To unlock the next generation of capabilities—that could be a meta-algorithm. Think about coding workflows: an orchestrator and lots of agents—multi-agent architectures, cognitive architectures: units of compute as neurons and synapses, or an LFM, or a small model, bundled into systems of systems. Agents should be able to train models casually as part of their job—and it wouldn’t actually be the LFM alone, it would be LEAP, because you need the full platform to adapt to whatever situation you’re deploying in.

swyx

这些是命名过的方向；Liquid 更近的研究议程是什么？

Of all these named directions, you also have more near-term research directions. What would you say is the current research agenda for Liquid?

Ramin Hasani

我们非常聚焦大规模多模态——所有模态一起训，同时做更长程推理；可靠的端侧助手需要更高可靠度。有专门效率部门：内核设计、推理、基础设施、更快的 RL 基础设施——从预训练到交付生产的全流程提效。研究上也看非人类可读数据，例如信号、生物；我们在建 DNA 基金会模型——用眼睛读 DNA 不像文本，词表大约四个字母，但序列极长；我们的动力系统/长序列技术天然往这类数据推。多模态研究现在在 Liquid 极热。

We are very much focused on massively multimodal systems—training all data modalities at the same time, and longer-horizon reasoning. Reliable on-device intelligence services like assistants require higher degrees of reliability. We have a department for efficiency—kernel design, inference, infrastructure, faster RL infrastructure—making all processes from pre-training to delivering a model to production more efficient. From a research point of view we’re also looking at non-human-understandable data—signals, biology. We’re building DNA foundation models: with your eyes you don’t really read DNA like text; vocabulary like four letters, but sequence length very large. Our dynamical systems and longer-sequence technology naturally pushed us into alternative data. Multimodal research is extremely hot right now at Liquid.

swyx

视频/音频进，视频/音频出——文本默认。输入端还好，输出端是新的。从约 7B 可用地出图、出声、出视频，现实吗？

Video, audio in and video, audio out—text is assumed. The in is fine; the out is new. Is it realistic to have image out, audio out, video out from a 7B model that’s usable?

Ramin Hasani

方向上是。我不认为能从那个量级挤出真正高保真视频；但若要解决物理问题、要有模糊的世界理解，在约 7B 以下用世界建模语境训 LFM，能编码某种物理知识——模型未必能高保真输出，却能增强内部表征。对工厂里接地的工业机器人，视频理解被低估了；现在多半还是交错图像，我们要更多。高度多模态系统仍很少，对基金会模型是很开放的研究题。

It is—directionally. I don’t think it is feasible to get proper very high-quality, high-fidelity video out of that. But if you want to solve physical problems and have a vague understanding of the world, training LFMs in the range of 7B and below in a world-modeling context, you can encode some physical knowledge—the model would not output something very high fidelity, but it would add to its internal representation. For grounded industry work of a robot in a factory, video understanding is so underrated; there’s still not enough—mostly interleaved images. We want more. There aren’t that many highly multimodal systems; it’s a very open research topic for foundation models as a whole.

swyx

相邻空间：Modular 被 Qualcomm 买了，刚开源 1.0。你们著名地从 JAX 转到 PyTorch。有没有探索过 Mojo？你们做自定义内核、又偏 Python。

Something adjacent: Modular just got bought by Qualcomm; they released 1.0 open source. You guys famously switched from JAX to PyTorch. Did you explore Mojo? You’re doing custom kernel work, you’re Python-centric.

Ramin Hasani

Modular 在基金会模型下一层做内核，想像我们一样当「瑞士」对接所有硬件，但现在锁进 Qualcomm。几周前我和 Chris Lattner 同台 AMD 面板——他是我英雄之一。第一性原理上我们不想重复造轮：很多推理公司想替代或包装 llama.cpp 做端侧，凡试图当「全硬件通用语言」、不锁单一生态的，深挖后几乎都不如原版 llama.cpp。我们不想换底座。锁进单一生态的反而可能更成功；我觉得 Mojo 在进 Qualcomm 生态后会更成功，平台确实让托管智能更容易，但我们自己部署还没用。llama.cpp 是高度优化的 C++ 内核编译器路径。统一极难——我不相信 Mojo 式平台的「均一性」；AMD 版、Qualcomm 版都要分别设计内核，那层脏活才是硬核。若有人能做成，是 Chris 的团队。我们会贴最原生、能尽量贴近金属的语言与平台——生态总是赢，对新人很残酷但不公平也得接受。难题反而更值得做：简单题人人会去解。

Modular operates one layer below foundation models on kernels—trying to be Switzerland with all hardware providers, but now locked into Qualcomm. I was on a panel with Chris Lattner weeks ago on an AMD panel—he’s one of my heroes. From first principles at Liquid we don’t want to reinvent the wheel. A lot of inference companies try to replace llama.cpp or build wrappers for on-device AI; almost every one that isn’t locked into one ecosystem and tries to be a general-purpose language for all hardware—if you dig in, they’re worse than the original llama.cpp. We don’t want to change base. Successful ones lock into one ecosystem; I feel Mojo is going similar now locked into Qualcomm—there’ll be more success; the platform makes hosting intelligence easier, but we haven’t been using those for our own deployments yet. llama.cpp is very optimized C++ kernels. It’s extremely difficult to unify; I don’t believe in the uniformity of the Mojo platform. You need custom work—kernels for AMD-specific or Qualcomm-specific Mojo. If anyone can pull it off, it’s Chris’s team. We stick to the most native language and platform that lets us touch the computer as close as possible. Ecosystem always wins—brutal, not fair to new entrants. If a problem is hard to solve, that makes it more attractive—easy problems everybody else will go solve.

swyx

收尾：明年是 Transformer 十周年。Attention 是否一切你需要？

The question I was going to end with: next year is the 10-year anniversary of the transformer. Is attention all you need?

Ramin Hasani

Attention 肯定是你需要的要素之一；我们还会与基于注意力的系统共存很久，它会出现在几乎任何系统里——像眼睛怎么看。但人并不是看一整帧：眼睛不断跳动，大脑在合成；你看到的是两帧之差。所以 attention 是组件；此外还有硬件感知、定制、具身——部署在数据中心之外的社会智能需要某种具身；虚拟世界里的智能体也要对「虚拟物理」有理解。具身智能与具身推理很重要；再上一层是适应性智能——能持续演化、持续学习、持续改进的系统。

Attention is definitely one element that you need. It will be a long time that we will be living with attention-based systems—attention is definitely going to be inside any of them. Like how your eyes work: you’re not seeing the whole view; your eyes are bouncing around and your brain is compositing—you see the difference between two frames. So attention is one component. Then again: hardware awareness, customization, embodiment. Intelligence that wants to get deployed in society outside of data centers—you’ve got to have some sort of embodiment. Agents inside the virtual world also need understanding of their virtual physical environments. Embodied intelligence and embodied reasoning are very important. And adaptive intelligence is another level—systems that can always evolve: continual learning, continuously improving systems.

swyx

很好的深潜。还有我们该问却没问的吗？

That was a really good deep dive. Is there anything we should have asked that we didn’t cover?

Ramin Hasani

没有，这是一场很好的对话。非常感谢。

No, I think that was a great session. Thank you so much.