Liquid AI: Product Launch Webcast 10/23

Visibility: Public Uploaded by: Liquid AI Uploaded at: 2024-10-27 Published at: Length: 3:06:10 Views: 9128 Likes: 88 Category: Science & Technology

Description

Find out more: https://www.liquid.ai

Outline:
00:00 - Introduction
01:56 - Keynote by Ramin Hasani, CEO and Cofounder of Liquid AI
13:18 - Science and Tech by Jimmy Smith, Founding Scientist
19:30 - Post-training by Maxime Labonne, Head of Post-training
25:06 - LFM Development by Alexander Amini, CSO and Cofounder
48:18 - LFM Deployment by Mathias Lechner, CTO and Cofounder
01:00:24 - Summary and Firesides by Ramin Hasani
01:03:15 - Intermission
01:32:47 - Panel with Maura Healey, Governor of Massachusetts
01:43:43- Panel with Ranjit Bawa, CTO, Deloitte
01:53:09 - Panel with Raymond Liao, Managing Director, Samsung Next
02:05:15 - Panel with Daniela Rus, Director, MIT CSAIL & Co-Founder, Liquid AI / Joscha Bach, AI Strategist, Liquid AI / Mikhail Parakhin, CTO, Shopify / Sebastien Bubeck, VP of AI, Microsoft
02:17:20 - Panel with Ichiro Tsuge, Chairman, ITOCHU-CTC
02:25:37 - Panel with Keith Williams, CTO, Capgemini Engineering
02:35:13 - Panel with Stuart Schreiber, CEO, Arena BioWorks
02:46:59 - Panel with Ralph Wittig, Head of R&D, AMD
02:58:05 - Closing with Steve Pagliuca

Liquid AI, an MIT spin-off, is a foundation model company headquartered in Boston, Massachusetts. Our mission is to build capable and efficient general-purpose AI systems at every scale.

Transcript

[Moderator] Eesteemed colleagues thank you for taking your seats. This is a reminder once again, please silence all mobile devices. Thank you. [Background] To make the leap from what's possible with AI today. To push beyond limitations like privacy, computational power, and the confines of the cloud, we've built something revolutionary, something dynamic, fluid, liquid. The Liquid Engine is a true innovation that will allow every company to own their own intelligence. It allows every step of model design and training to be tailored to the specific needs of your company — whatever they are. Today or in the future. Models that are more memory efficient, more explainable, and higher quality. Giving you the power to scale up or down, at an affordable cost. Making what has up to now been impossible, possible. Like enabling autonomous drones to find brush fires before they become wildfires. Analyzing a patient's entire medical history against their unique genome. Or detecting manufacturing anomalies using multiple modes of sensor data. Whether you need a model that fits entirely in the palm of your hand or models that will solve the world's biggest problems, the future of AI should not be limited, it should be Liquid [Moderator] Please welcome Ramin Hasani, CEO of Liquid AI. [Muisc] [Music] Good morning everyone! Welcome to the first Liquid AI event where we're going to talk about a lot of cool things we've been doing the last year and a half at the company. So as you all know we are in transformational times we are in a new era of statistics. What we are experiencing it's amazing because we started a scaling Foundation models. We started building models larger and larger we put a lot of energy into breaking the barriers and getting into the regimes where general properties, general knowledge can actually emerge from scaling neural networks. And that allowed us to get into a new era. This in this new era AI is changing everything. AI is helping us to discover new math discover new physics discover new drugs. AI is allowing us to solve science a lot of of the impacts that AI has is going to help us become the next you know contribute to the next generation of evolution. But AI also has some limitations. So what we have to think is that scale doesn't come for free you're spending a lot of energy to actually get there at the same time when you scale systems you're getting into troubles like carbon and the sustainability aspects of the whole technology. It is good the goal is amazing but what we have to deal with is also some of these challenges you know it's not only this but everything is increasing so another aspect is which is very important is that we are scaling black boxes. What we're trying to do we we're basically like building systems with elegant simple mathematics we take them to the next stage but then it becomes harder and harder as the scale of these systems become larger to understand what happens so we losing control so it's not that um easy to be uh ignorant about these issues these are from the scientific and technological kind of perspective on AI but when we talk also to Enterprises let's see like some of those challenges you know you're hearing like when we interacting with Enterprises we see people say the codes from people that we talk to what if I don't have um internet connection if I want to use these powerful AI models what if I want to personalize these models what if I want to have ai systems but I don't want to spend that much energy that that much kind of capital into into designing these systems why do we have to spend millions of dollars to get there now and a lot of other codes that we're hearing about the implications of this thing these are the realities we're facing when talking to people and these are the underlying consequences of scale or this one you know gpus what if I want to run my systems on CPUs and a very important matter what if my data can't leave my premise due to file size laboratory re requirements or privacy concerns we built Liquid AI to actually enable us to get access to generative AI in a smarter way we want to enjoy the benefits we built the company with the mission to build very capable and very efficient AI systems that are general purpose at every scale so if you look at this Vision that we have we want to get smarter about scaling properties of these Smalls this trend is there is not stopping but what we are trying to do is that can we do more with less that means if we are in smaller parameter regime can we unlock better capabilities with AI if you think about it from first principles and we did just that we built something that we call liquid foundation models we released 3 weeks ago the first batch of language lfms these models actually elevated the scaling laws that we've seen before into regimes that higher quality models are now possible at smaller scale at every scale we released a 1.3 billion parameter model released a three billion parameter language model and a 40 billion mixture of expert system that allows you to have this technology at every scale enabling general purpose Computing be it a chat applications be it a reasoning or a mathematical kind of question or if you want to put it in production for performing a task for you liquid foundation models are neural networks they're large neural networks that their computational units are rooted in control theory and in U signal process processing and this is a completely new way to look at AI systems with with lfms we want to change the base of AI so another property of lfms is that we are not just looking at increasing the quality of models but we want to also be better and more efficient at the same time so usually when you use a foundation model you are giving it a couple of like contexts and then the model generates for you more kind of tokens or outputs and then that tokens are increasing as those tokens as you use the model the amount of kind of information that you actually extract from the model use increases now that would lead to a much higher kind of memory capacity requirement on device that's why we need higher and higher or better and better Hardware but a liquid we enabled something that up to a million tokens you can fit the data that you're using and you're feeding inside the network to map on 16 gab of memory this was something that was not possible before while achieving the quality of the models that we get we released these models as I mentioned three weeks ago and we became the models became trending on uh you know we we're not a a consumer facing company we we're serving Enterprise is but I mean it seems like the consumer also like love playing with models like right next to Gemini Flash and and llama 3.1 we have had a trending model so I I guess today actually if you go to open router as well you would be able to see like our API uh is still trending I think and uh that's a testimony for us that yes so the work is actually paying off lower energy lower cost the best quality you can unlock from the systems I'm very excited to today tell you that we're releasing more we releasing multimodel lfms we're releasing an audio lfm and a vision lfm the audio lfm is just short of a 4 billion parameter model that can fit on a device that can give you seamless audio kind of generation it can do speech to speech and it can perform you know speech to text kind of functionality the vision lfm is going to process text and image and can produce context on top of that kind of image there a lot of applications that we can unlock on the edge and on Prem using this New Foundation models we're going to talk about this today and the team is going to tell you a lot more but there is also more so we're going to release a library of elephants we have been applying our technology to solving multimodal data data of any nature any sequential nature this data could be biological DNA sequences this data could be Financial time series this could be U you know in B in any kind of for example in autonomous driving could be signals that you're collecting for driving we're building a full Library as products of liquid and we're going to talk about this the team in a second but before getting there just want to position also the company itself so you should not think about liquid as an AI architecture company we've innovated an AI architecture for sure but architecture is one element of a foundation model company so we are a foundation model company we are putting together a full ecosystem of model development we have innovated across not just architecture but on learning algorithms how to train how to pre-train how to post-train how to evaluate the models we have thought about this deeply from first principles to get into these qualities and this shift that you see in scaling laws so you're going to see a lot today more details about what what I showed you're going to see a little bit about science and technology behind these models how you're building them you're going to see a library of lfms we're going to talk about how we can enable this and how we can give people access to to the technology so that they can build with this technology through liquid def kit and we're going to tell you what kind of infrastructure is ready to be used deployment and after that we're going to have a couple of uh luminaries on stage to tell you some of our partners some of our future partners and some of the key figures of the industry that have experienced with the you know with lfms and will experience with lfms on the stage okay so I'm excited we are going to Deep dive in science and technology let me ask Jimmy to show you more [Music] thank you from a technology perspective much of what ramin has just shared with you is the result of us building a way to develop high quality and efficient models that do not require you to have to choose between the two so I'd like to take a few minutes to tell you more about what we mean by this how we think about the problem by quality we mean the capabilities of the model and the usefulness of the outputs it can produce for various tasks we can consider many aspects of quality with a few important examples listed here the first is knowledge capacity or given a specific model size for a specific regime we want to deploy in how much relevant knowledge across a variety of domains can be packed into the model weights from the training data this is important to ensure the model has a strong base of factual knowledge to be used for various quaries another important aspect of quality that we consider is multi-step reasoning this ensures that the model is capable of breaking down complex task into simpler problems that are easier to solve we also prioritize Lan context capabilities which ensures that the model can effectively use information from throughout its context to solve a task this was particularly important for us to focus on because note that just because a model can efficiently process aone context does not mean it can necessarily effectively make use of information throughout that context when we talk about efficiency we think about both inference efficiency which determines the cost and resources required to deploy the model as well as training efficiency which determines the costs and resources required to develop the model and it's important to note here that there have been many prior attempts at developing efficient Transformer Alternatives in fact many of the members of our team have performed foundational research in this area however prior attempts at scaling these prior methods both by many others and ourselves has always led to the ob observation that you have to sacrifice on at least one of these aspects of quality if not many and this is one of the challenges we have been solving at liquid and we have developed a way to deliver high quality and efficient models that do not require you to have to choose between the two and we have accomplished this by developing a holistic framework for high quality and efficient model design given a particular deployment scenario we do not only consider the model architecture which we do and is important but also how everything interacts together but also how everything interacts together from the data the training algorithms and the post-training procedures and how all these interact with the architecture and the scenario we're trying to deploy in and this view has allowed us to innovate in each of these areas and a key driver of this innovation has been our careful design of fine grained internal evaluation valuations which allow us to pinpoint the strengths and weaknesses of different approaches to architecture data and training designs so I'd like to walk through just a concrete example to give you a better sense of how we think about this and let's consider one of our internally developed L context capability evaluations which measures the model's ability to use information from throughout its context in a complex way to solve a task and I will start by plotting the performance of a strong Transformer Baseline on this metric throughout a training rout the y- axis represents a quality score and the x-axis represents the number of tokens seen during training and we'll note that at the end of training the Transformer achieves what we would consider to be a high quality score and we found that in practice when combined with high scores on our other evaluations this is predictive of strong performance on real world applications that we care about now on the other side we'll also plot the memory footprint that the Transformer requires to solve this task and we'll note that of course the Transformer requires what we consider to be a large memory footprint which makes it makes it expensive to deploy but now we can also plot the performance of two previously proposed efficient Transformer Alternatives one a modern input dependent State space model variant and the other a modern gated linear attention variant and we see that while on the on your right the memory footprint of these efficient Alternatives is very favorable compared to the Transformer Baseline we also notice a large quality gap with the between the performance of the Transformer and these efficient Alternatives and we found in practice this Gap was significant and predictive of very important failure modes when trying to apply these previously proposed efficient alternatives on Downstream task and of course this is just one of our many evaluations that has allowed us to identify other important failure modes not not previously recognized and this created clear paths for innovation of course the first release of our lfm models are a very early result of these Innovations focusing in on the performance of the one second focusing in on the performance of the lfm versus the Transformer we see that the lfm actually improves upon the quality score while maintaining the very favorable memory footprint allowing us to deliver on the cost savings so these are the types of challenges we are solving at liquid and this is just a very brief preview to give you a sense of how we are delivering on the promise of high quality and efficient models that do not require you to have to choose between the two now another important core component of our framework that I have not spoke much about are our post training procedures so I'll now invite to the stage our senior scientist and head of post training Maxim Leon to tell you more about this thank you [Applause] okay thanks Jimmy Hi everyone so here I'm going to talk a bit about uh the post trining work that we do at liquid AI so during pre-training models are trained on a lot of data to learn General capabilities but we need to tune them to turn them into useful assistance that can answer questions and that can for follow instructions this is the goal of post training by creating uh high quality examples we challenge our models to learn how to answer complex questions with multiple steps and with Advanced knowledge but this is an iterative process and to monitor our progress we need the best evaluations so in addition to the internal evales that Jimmy talked about we have an entire stack of evaluation at every stage of the training process these evaluations can guide the pre-training and post trining process to know when we achieve the right level of performance public benchmarks allow us to confirm our internal scores and also compare our models with our competitors but this is not enough to make great models to make great models we also need to evaluate that chat capabilities how they interact with users so to do this I'm really excited to talk about the liquid Arena that we created so in this Arena we have two Anonymous models and they provide an answer to the same instruction then we ask a human to choose the best answer we can repeat this process with a lot of instructions and a lot of answers to understand where our models are good and when they underperform but to scale it to millions of instructions and answers we created an automated version of this Arena where we replaced humans with llms this gives us A fine grain understanding of the capabilities of our models all these evaluations provide feedback on areas of improvement for us so for example if we see that our models are not great at summarizing text in German for example we can create better data retrain our model and fix this issue so to do this we have a custom preference algorithm that can process and learn from Human preferences very efficiently in addition liquid training recipe also includes distillation and model merging so with distillation we have a big teacher model and a small student model the student directly learns from the outputs of the teacher this effectively impresses the higher performance of the big model into a much smaller one which is both cheaper and faster to run then there's model merging this one I'm the most excited about uh so during model merging the the idea is that we merge the parameters of different versions of the same model and this creates a single unified high quality model this combines high performance across different tasks and this also allows us to um upcycle our experiments into better models so all of these techniques in the post trining World they allow us to create and deliver best-in-class models we also have a strong focus on on evaluations to Target fine grain capabilities in our models finally for n users this delivers better customers ability to adapt lfms to particular task and domains of Interest a good example of this is the Japanese model we co-created with CTC here we fine tune lfm 3B on Japanese data and we created the best Japanese Edge model on the market this is confirmed by both public benchmarks and also human preferences we got from the liquid Arena okay so now we've covered all of the model cycle that we do at Liquid AI. Now let me introduce Alexander, our CSO and Co-Founder to talk about the principles behind LFMs that go beyond language. Thank you. [Music] Thank you Maxime, and hi everyone! This is so exciting because what you've all seen so far today is how Liquid is building best in-class language LFMs that are state-of-the-art in both efficiency as well as accuracy. But today, we all know that large language models (LLMs) are extremely overhyped. Liquid AI is not a language modeling company. Liquid AI is a foundation model company. We build LFMs for all types of sequences. And well language is just one type of sequence. In reality, sequences are everywhere in the world. And in order for AI to be everywhere in the world, we need LFMs to be highly capable on all sequences. That's exactly what we're doing at Liquid AI. Today we unveil a suite of LFMs that are going to revolutionize across domains and industries. Let me tell you a bit more about how liquid foundation models are revolutionizing biological and therapeutic design. We've already seen so far today how LFMs have learned the language of our natural language but actually LFMs can learn the language of biology as well. Here you can see BioLFM. BioLFM learns the language of proteins to learn how to generate brand new proteins from looking at the entire existence of all of these proteins that exist in the natural world. BioLFM is highly capable and highly efficient to generate brand new proteins that have never been created before but we can now create them in the lab BioLFM is so efficient and scalable that we can do this not just for one protein, not just for five proteins, but at large-scale across biology and accelerate biological discovery and design. It doesn't stop with just efficiency, BioLFM scales to capability and accuracy as well. BioLFM is state-of-the-art best-in-class accurate protein foundation model at this scale and we are still scaling and improving this model to bigger scale and moving these digital generations into the actual lab. Now liquid Ai and lfms not only enable everything from understanding the biological world but we extend to also understanding our physical world as well this is critical because our physical world is so complex it's so dyamic and Ever Changing that in order to deploy autonomous systems into the physical world we need to be able to understand and simulate our physical world this is so challenging because of all of these complexities that the physical world entails we built drive lfm to understand our physical world so that we can test and simulate new environments of our real world for these autonomous systems to be stress test and verified before they go out into reality Drive lfm learns from our world how the world Works to understand the world using a sequence of past video here we look at driving as a particular impactful example to understand the world and then learn how to predict the future learn how to generate brand new futures for every single scene every single environmental condition time of day weather condition and so on bio lfm excuse me drive lfm extends to all of these different scenarios and again it doesn't stop here lfms enable us to understand our physical world but also understand our own interactions and behaviors this is so important because when we think about how we interact in our world it's also so personal and so complex every single day each one of you is is transacting is interacting with other individuals through financial transactions as well as other businesses and in reality if we look at this sea of transactions identifying and and isolating a single fraudulent transaction a single mistake in this massive sea of transactions that occur is a classic needle in the hay stack problem it's a huge undertaking a huge challenge but identifying this one fraud is one of the most critical things to ensuring and safeguarding our personal finances and the finances of all Enterprises in fact if we zoom into these transactions what we can really see when we zoom into the sea of transactions we can actually see that every individual transaction is nothing more than a composition of what makes that transaction a transaction how you what you purchased how much you spent when you made the purchase where you were all of these go into that that single transaction but of course we all know that transactions don't exist in isolation these this one transaction was actually part of a much larger sequence of transactions that you participated in understanding this sequence and predicting what comes next in the sequence is the core problem that we solve with transaction lfm transaction lfm is able to take as input all of your past sequences or excuse me past transactions and be able to predict what comes next which enables us to predict when there was a potentially fraudulent transaction in the mix so that we can identify and Safeguard every individuals and every business's financial transaction lfm is able to scale towards large scale transaction histories which means that our fraud detection models are very very accurate because it learns from all of your personal past history of transactions using the liquid technology because these models are so efficient we can scale to these regimes but also it means that they are highly accurate now what we have seen is that lfm scale to language and far beyond to all of these different sequences but in reality when we look at this landscape we can actually see that lfms are not only efficient and accurate they are highly multimodal which means that we can combine these different sequences together and learn how to talk to sequences we can leverage language on one side and go into an lfm and also communicate and talk with other sequences here for example I'm going to show you a demonstration of how we've built a time lfm a Time series lfm that you can communicate with directly with natural language time lfm is able to ingest both language as well as time series which is a highly structured and very complex data type let me show you an example time lfm comes with a very simple interface because it's rooted in natural language but has a backend of Time series you can directly describe what you want to search for or identify with time lfm for example here identifying key financial Trends using the language model to query in a multimodal fashion all of the underlying underpinnings of our time series model it doesn't stop there we can identify these Trends but we can also use use time lfm to analyze key and very precise time and metadata information but of course the most important part is because this is not just a wrapper on top of a language model this is a multimodal Time series combined with language model it means we can actually understand the Dynamics the behaviors of what's happening in the time series things are not things that are not only located in The Meta data description of the time series that you could get from a Google search but simp but actually going deep into the time series and understanding the behaviors of the time series and then interfacing with this through natural language for example here looking at a funny funny use case of trying to uncover cryptocurrencies that have had Market manipulation all of these are examples of how time lfm are revolutionizing the ability to not just have language models that you can communicate in and out with language but can communicate through all of these different sequences that we've been talking about today and this is just an example of how we've been scaling lfms to achieve highly efficient highly capable and accurate modeling that is fundamentally multimodal now how did we do all of this lfms we built lfms using what we call the liquid devkit kit the liquid dev kit is a developer package that allows us to build and and deploy lfms for any domain and any scale let me tell you a bit about the liquid devkit the devkit sits on top of P torch as well as optimized kernels that we have created at liquid for all of these efficient and accurate operations it is comprised of a variety of different abstraction levels that allow our internal engineers and scientists to easily build and scale these models across a variety of different levels starting first from the lowest level of operators and then moving up to blocks and backbones from there our engineers and scientists can Leverage devkit in a very easy to use way so that we can achieve all of these incredible lfms that we're releasing to the world today in just a single line instantiating and building these lfms and I'll actually walk you through an example right now of how we can have the entire pipeline of building state-of-the-art examples with the devkit scaling them to large scale highly capable systems and explaining the outputs every single output every single prediction that comes from this model so that we can certify the model for safety let me walk you through an example of how we build with liquid devkit the devkit remember I told you is a composition of layers that make it very easy to scale these systems across all of these levels of complexity I'll start first with the lowest level which is the operator level operators allow us to take as inut a sequence and learn how to transform that sequence now operators typically are extremely complex operations they take a lot of very low-level optimized code and kernels that's very very hard to not only write but write efficiently the liquid dev kit takes all of this complexity and removes it it abstracts all of this complexity away and moves it into easy to use Liquid operators that can be easily instantiated for your application we do this for a variety not just one operator but a variety of different liquid operators that we package together in the devkit this includes both operators which we have created and invented at liquid as well as uh leading open-source highly efficient alternative architectures that exist in the world we can bring this up a level from operators now to blocks blocks allow us to elevate the level of complexity that our models can handle and they do this by by very straightforward gating mechanisms that are abstracted also into the devkit in the same way and finally going up one more level Beyond this we can combine these blocks into backbones and the unique thing here is that devkit is able to handle these levels of abstraction and scale them up into the backbone level in ways that goes beyond vanilla Transformer sequence insequence backbones that we see in the world today the liquid devkit allows us to build backbones that have hybridization or feature sharing or depth sharing allowing us to move beyond the state of the art today this is how we build lfms at liquid and how we work with our partners to build lfms but it doesn't stop there the liquid devkit enables us to not just build and instantiate an lfm but we now need to train it and to scale that model across Hardware so that we can achieve the incredible results that you're seeing today this is done by actually enabling highly scalable and distributed training uh Pipelines with the liquid devkit you can see an example here on the left and right side of building that model that you saw in the previous slides we built an lfm here a very small lfm I'm going to show an lfm that's only about 150 million parameters but now let's ask a question of okay what if I wanted to build a bigger lfm what if I wanted to scale this into a multibillion parameter lfm well this requires me to basically scale the number of blocks because all the complexity is actually abstracted away into the operators and the blocks I can just scale the blocks to scale my lfm so let me increase the number of blocks from two to 12 this makes my lfm about 12 billion parameters now but what happens is actually this lfm if I put 12 blocks in it it no longer fits in my GPU so now with devkit we make it very easy to add one extra line using a devit k. distribute line and take those 12 blocks and split them to distribute them across our Hardware so that we can scale to multiple gpus multiple nodes and have all of those nodes seamlessly communicate with each other to handle this much bigger model finally I'll tell you about how we're able to also take these models that we train with the devkit and explain every single answer that comes out of their uh predictions that comes with their predictions this is because liquid neural networks were invented to be highly explainable systems we achieve this explainability through the devkit and we package the explainability through the devkit using what we call model. exlain model. exlain is a very easy to use function that comes with every single model that is built with devkit so any model that that is built with devkit actually comes with this very special function the do expained function do explain takes as input the input of the model but unlike other neural networks other AI systems that just take the input and predict an output explain explains how this output was generated with respect to the input and with respect to the model it can do this with respect to going inside of the model understanding what in the internals of the model made this prediction predict what in the data did it see to make this prediction get outputed right and all across the board looking at all of these different axes of interaction that our model has with itself and with its data let me show you one example of how we've used model. exlain with one of our lfms so we built a Vision language model this is a model that takes as input both an image as well as a textual description or a textual question about that image and then the model is supposed to respond in text about this question so we can ask a question to the model about this image and you can see the image on the top right here describe this image in one sentence you can see a couple things in this image the cat is very obvious but if you look a little bit closer you'll see that the cat is about to play with something in front of it there's a small insect very very hard to see let's see what happens when we feed this image to our vision language model lfm here we can see we feed it to the lfm the lfm instantly responds and it's able to say and correctly predict that in this image we can see a cat and an insect on the ground okay this makes sense but let's look inside the model let's say what happens when we use model. explain to explain how we got to this prediction for each output that the model creates we can ask model. explain to explain that output starting first with the very first token moving on to the image once it says in this image we can see that the model starts to expand its focus a little bit and then it talks about the cat it gets to the cat and you can see that now the focus has shifted it's now looking at the cat it's not looking at other places it is explaining why it's predicting the word cat it's because it's looking right right there and then you can see it shifts its focus yet again now looking at the insect which was so small in the image and finally relocalizing itself back into its context understanding that all of these things happened while on the ground and shifting its focus yet again from the cat to the insect now to the ground this is just one example of how we are using model. exlain but model. exlain is deeply embedded in all of our models that we build because they come with the liquid dev kit now again now today you have seen how we've built these lfms that are highly capable highly accurate highly efficient and across theboard multimodal for all of these different domains and industries today we release this Suite of lfm to the world and also we're going to Showcase how we actually this tweet of lfms one of the most empowering things I think about liquid AI is not only that we build these lfms for the world but we deploy them to the world at all scales and one of the most powerful aspects of our lfm library that you can see here is this part about the edge lfms today and I'll pass to Matias who's my my co-founder and CTO of liquid AI who will tell you more about how we're actually deploying this library of lfms directly onto the edge in a fully private fully capable manner what if to maximize ai's potential we need to solve for its limitations what if a major Bank could accurately analyze billions of transactions for fraud without letting a single bit of data leave their firewall on Prem or a leading Quick Service food company could power automated drive-through windows in rural areas without relying on the internet what if a chip manufacturer could build a microchip with the power of AI in it what if a regional Telecom company could process and filter their data without sending all of it to the cloud whatever your use case is liquid AI can develop and deploy solutions that Target and overcome the specific limitations of your company with efficient scalable models that turn your wh ifs into what next with liquid AI it's not if it's when please welcome CTO of liquid AI mataas [Music] lechner so we have just seen in this video why gen AI on the edge matters and I'm going to tell you a bit of progress that we have made recall the slide from before where we showed that we train uh lfms at every scale uh our currently smallest model at around 1.3 billion parameters a 3B model and also a 40b uh model are currently most capable and largest variant and the numbers the parameter size of each of these models was not chosen uh randomly but with a specific Target uh end device to run on in mind for instance the 1.3 billion model is able to run on embedded uh Hardware the 3B is optimized for phone use and the largest variant to run in the cloud or on Prem on a single GPU so at L we we don't just train at every scale but we also deploy at every scale in the in particular we deploy where no one else can so here we have a video of a Raspberry Pi a Rasberry Pi is a $60 cheap uh low power fully fledged computer with the compute capability of around a 10year old laptop and we ported our smallest model the 1.3 billion parameter model uh to run on even such small uh and low power devices particular asking about the Isotopes of hydrogen I was the wrong video before but yeah you we see that the model is responsive and can make use of even like the low power and like the small and cheap Hardware of the respirator so this is the smallest model like which is nice but let's take a a step and look at the other side of the spectrum of our currently largest uh model uh the 40b particular one thing that Jimmy mentioned is that we also look into making sure that our models can make a long context so they are memory efficient but we also made sure that they are making use of all of the context uh that they provided and in particular we can then feed additional data and information that is not in the training data into our model in particular here we can add the web search functionality and augment the model with post training like data that's not in a training uh with data that's not in a training set particular lookup of former co-founder Alexander me uh look up his profile using the web search which is quite nice so these are all uh so far text models like texting text out but Alexander mentioned to you that our models are natively multimodal meaning they uh we can have uh conversations with our model uh hello what's your name hello my name is liquid Edge I'm a realtime Voice Assistant here to help you with questions and tasks uh can you tell me one thing to do in Boston accept the Freedom Trail absolutely one great thing to do Boston aside from the Freedom Trail is to visit the Boston Common it's a beautiful Park in the Heart of the City where you can relax have a picnic or even play a game of catch it's a great place to enjoy the outdoors and take in the city's vibrant atmosphere uh can you give me one short tip for networking at a tech launch event Absol absolutely one tip for networking at a tech launch event is to be genuine and approachable start by smiling and making eye contact when you meet someone ask open-ended questions to show your interest in their work and to encourage a conversation remember to listen more than you talk and be prepared to share your own experiences and insights don't be afraid to follow up with a LinkedIn connection or an email after the event to maintain the connection uh thank you you're welcome how can I assist you further so this was our [Applause] three so this was our 3B uh 3 billion parameter speech model um and due to the native multimodal support provides low latency with uh after the user provided some um instructions for the model uh today's uh large language models are all running in a cloud um connected through an API but what if um we don't have a reliable internet connection and want to make use of gen or what if um we have privacy and security concerns that we don't want to share our data uh with a potential uh API provider we at liquid um deploy where the the device is and not where the connections are instead of telling you about it let me show it to you so here we have our app where we can chat with our offline model let's also turn off the Wi-Fi and ask it something Halloween is coming up soon so let's test the creative writing abilities of our model give me three uh AI sorry a i themed Halloween costumes so this is our liquid yeah neural network Necromancer is deep learning Droid and generative adversarial night that's nice quite creative yeah let me uh like write a scary story about what was the first one neural network NE chroman the title the neural The necromancers Awakening and yeah I'm not going to read out but I think you get a [Applause] point so this is our 3B language model running fully offline on on our [Music] [Applause] iPhone it's yeah one of the things we have achieved at liquid so far we have seen chat models so basically open-ended conversations with the with the model through speech through text and so on but what if we have a consumer a customer application uh where we want actually the model to be tailored to a specific use case and we created a hypothetical use case uh where we want to use our 3B speech model and we tailored it using a fine-tuning approach to uh uh a tech support a customer agent the customer this this this agent takes unstructured speech input irrespectively of format and translate it into a structured format that is machine processible um and readable by by a machine it uh the video also let me just customers reporting that their TV is not turning on and uh they're suspecting this is due to a recent software update and uh customers tried to turn it on and off and unplug in and replugging uh customer's full name is Robert wilmart and could you summarize the issue and give me a short classification so we have seen our uh my colleague Mark um basically using this demo um taking speech input translating into structured format using the uh the classification abilities of the language model as well as the summarization abilities um but as I mentioned before we at liquid bring the intelligence the geni to where the devices is where the devices are so let's uh run run it again but making sure and showcasing that it runs name is Lily last name bazak uh correct that to b a n a s z a k um customer is reporting uh potential malware infections so whenever they turn on the computer essentially there's only a red screen and uh basically the computer is unusable and could you provide with a classification please and can you summarize this issue [Applause] awesome so we have seen now like a potential commercial application of our speech lfm fine tuned on the specific task running offline we at liquid uh unlock the power of generative AI on devices that were once Out Of Reach in a fully private and secure manner we have seen that we train lfms and our foundation models at every scale we also deploy them at every scale in particular we have seen it as small as a Raspberry Pi we have seen that our models are multimodal taking speech and vision input we have also seen that our models can be fine tuned and tailored to customer applications very easily so let me hand it back to Ramin thank [Applause] [Music] you all right all right all right so what do you think we are a foundation model company we're building the most capable the most efficient AI systems that they you haven't seen before at every scale so as you saw we can develop and we can deploy these models today we do this for on Prem kind of solutions that you want to have or on device as you've seen throughout the talks lfms are private solutions that are enabled by lfms can be explained the dynamic can be explained they're high quality they're reliable you can use them for many many different applications now if you're really want want to engage with us we would love to work with all of you we want to um give you a couple of you know opportunities if you want to develop AI on the edge if you want to deploy that on Prem you can talk to us you can come come to our team and we can we can uh we can have like conversations about those you know use cases that you might have imagine uh uh building with Gen and you want to do that in an efficient way you want to do that with the highest quality and more more reliable version of your models this is where you can contact us the other thing is that you we also provide a codevelopment kind of opportunity because we're still early on in our journey you know to building these lfms for many many different applications there are some applications maybe we don't have even a uh visibility into but you've seen for example we wanted to Enlighten you with a kind of a taste of what kind of possibilities we what kind of things we have we are thinking about we have thought about for co-development and deployment of AI systems so you heard from our team in the next chapter we are going to see some of these lfms used by some of our partners we're going to have conversations with our partners on stage and um some of the luminaries are joining us including the governor of Massachusetts and um we're going to talk about the impact that we can bring with lfms revolutionizing the how AI has been uh deployed in the society so don't go anywhere we we're going to do an 15 minutes intermission and we're going to be back with the panel discussion thank [Applause] you ladies and Gentlemen please enjoy this short intermission we will resume programming at promptly 11:30 thank you [Music] [Music] oh [Music] [Music] yeah hey hey [Music] hey e e [Music] oh [Music] yeah I you [Music] hey [Music] hey [Music] he he he hey yeah hey [Music] [Music] e oh e you [Music] he [Music] e e e e [Music] yeah yeah [Music] yeah [Music] I you he [Music] [Music] for [Music] [Music] [Music] [Music] [Music] oh [Music] heyy hey he [Music] hey hey [Music] hey hey [Music] e [Music] e yeah [Music] I [Music] [Music] oh hey [Music] [Music] he he [Music] [Music] he yeah e e [Music] oh [Music] yeah you [Music] [Music] [Music] [Music] [Music] H [Music] hey it [Music] hey he he [Music] [Music] e e e e [Music] [Music] I he you [Music] ladies and gentlemen it's time to take your seats please silence all mobile devices once again please silence all mobile devices and take your seats the program will begin momentarily e oh ladies and Gentlemen please take your seats the program will begin momentarily [Music] please help me welcome the governor of the Commonwealth of Massachusetts Mora Healey and liquid AI CEO Ramin [Music] Hassani okay amazing amazing so Governor let's just get into it yeah why do we have to put our headquarter here in Massachusetts and on California my God where else would you want to be right I mean um look uh you you you know this right I mean you you just saw the demo um and it and it reminds me you know of of the shots first fired and conquered in Lexington Massachusetts that gave birth to the this great country the United States of America Massachusetts we have been about Revolution and Innovation from the very beginning and that's why it makes sense um it also makes sense because you know here we are at MIT right and you came here Daniela others came here to study to learn and have now found this the place to to birth a a business and not just a business but uh like that 250 years ago Skirmish on a battlefield what you guys are doing is revolutionary it's going to change the world so that's Massachusetts DNA and you know I'll also say this last night um I had a chance to see the banner raised at TD Garden for our Boston Celtics Banner 18 the most storied franchise in all of sports history and you know um I appreciate Steve and Nick puka and uh for for so many things and especially of course last night their stewardship of the Boston Celtics I think about Boston do you know who they whooped last night the New York Knicks okay so while there are some cities New York San Francisco maybe bigger you know open there's there's stuff happening there Boston I will tell you kicked New York's ass because we are more Nimble more Innovative more Scrappy just like liquid AI who's found a way to like Lea frog everything by doing things twice as fast with half the energy which leads me to my final point you know we've got big goals around climate here you know sustainability we lead the Nation when it comes to climate poy uh policy and I love that you found a way to do this and do it with half of the energy it's incredible so that's why it makes so much sense that liquid AI is here in Massachusetts thank you Governor thank you thank you so from this state point of view how are we making sure that we are positioning well in this new AI Revolution yeah that we're experiencing you know I I think it's about leadership and as secretary ofon how our great secretary of Economic Development um she uh she's doing a great job for all of us and and one of the things she put together was an economic development plan and you know Massachusetts we became we are the Global Hub of Life Sciences because there was a huge investment made on the part of state government and and funding there was work with our incredible ecosystem of colleges and universities and research institutions and there was love and support given to that ecosystem and that's what built it into what it is it's it's people like you right it's people like you our Founders our entrepreneurs um that we want to support including through our economic development plan which includes a few things you know doubling down on Life Sciences saying that we're going to be the global leader and when it comes to climate technology and importantly we want to be the global hub for or applied AI okay we recognize that others may have you know built out Ai and have some things going but the the race is on to see who's going to use AI to solve the world's most challenging problems President John F Kennedy a Massachusetts native son said we don't do things because they're easy we do things because they're hard so in the the the race the challenge to cure disease to build resilience you know to identify how to bring a higher quality of life for people across this planet something that liquid AI is going to enable in absolutely Revolutionary Ways we want to support that as a state by leaning into this space fantastic all the reasons for retaining talent but what are you going to say about like really like making sure that our talents are actually inventing the technology is here and staying here in the state well um look I can brag about title town and sports all right I don't think there's a better place in America as seen by last night the quality of life here right now Massachusetts just for everyone to know we have been ranked number one in education number one in Innovation number one in healthc care we're the number one place to have a baby to raise a family the number one place to live if you're a woman and we are among the safest states in the entire country Boston is the safest city this all speaks it's it's quality of life and also you know this we believe in knowledge we believe in science we're the home to the first public school and public library right we invented the computer you told me we invented AI somebody did way back here yes I remember that so like look um you know there's freedoms that we protect and that's what I as Governor I'm going to continue to focus on with our team protecting those things things that make us great now working on because and I think those things that make us great attract Talent you know it's why we have students and grad students coming from all over the world that's why we have companies looking to grow here we have challenges high cost of housing so I've said about fixing that we just I just passed signed into law A5 billion doll housing bill that's going to generate more housing all over the state so that we can lower housing costs to keep people here yes public transportation infrastructure okay those of you on the red line you can knock the red line but tell you what um for the first time in eight years when I started as Governor I put an engineer with Transportation experience in charge of that red line and we have already seen improvements we are a state that is making those infrastructure improvements in transportation I put together a team we've brought back 8 billion from the federal government the Biden Harris Administration already and just 22 months $8 billion going into Transportation infrastructure housing all these things I think are going to make it easier for people like you to grow and scale here for our um scientists and researchers and graduate students and teaching fellows to actually stay here and for businesses to look to expand here and you are just going to be so integral to that because as I say you're able to do something twice as fast with half the energy and we're going to sync up and make sure that we are supporting you because I know it's going to be great things for economic development for our state amazing thank you let's talk about AI regulations what is your approach was this the state's approach compared to The Californians and and and other nations I embrace AI um I embrace AI some of you may or may not know this I was the Attorney General here I actually work with kamla Harris when she was attorney General in California and I'm quite familiar with laws and regulations and all of that serves a purpose but I think we've what we've done in Massachusetts in AI is different one of the things that we did and I thank secretary how secretary Jason Snider who is our secretary of Technology we set up an AI task force and some of you I'm looking out are on that task force you know what that task force is about it's not about regulation it's about Innovation it's about how do we work together to seize this moment to make sure that Massachusetts is the global hub for Applied AI I think we can lead in Innovation we can scale this we can figure out what we need to do um while also of course being mindful of doing so in a way that is that is thoughtful and and doesn't do an injustice uh to to to people or to populations we can do both at the same time absolutely so so let me conclude by asking you the last question so I want to ask you about the you know like AI is expensive AI needs investment you know how are we in the state making sure that we get the right amount of investment in place for AI Ventures and AI entrepreneurs to actually become Innovative and contribute more and become more and more competitive well I'll start with100 million that's what we'll give you 100 million dollars for Applied AI um and um and I think what we can do is we can leverage that with existing Venture money that's here I mean we are fortunate in Massachusetts because I think you know per GDP we we have um probably the the the the lead when it comes to uh Venture funding and so much of it is going in this space so I think if we work together as a team um as Avon says and she'll give you all these bracelets team Massachusetts and we're proud that liquid AI is on team Massachusetts it is about teamwork working together how we leverage State dollars private investment and the great work of our colleges and universities and research institutions to really blow this open and get this done I'm so excited about what's possible as a state that invented the rocket ship and we did um I just can't wait to take to see this all just take off amazing give it up for the governor of Massachusetts Mara Hil thank Youk awesome all right awesome it's going to be great okay bye guys please help me in welcoming ranit Bawa CTO deoe good to see you that's a tough act to follow all right let's get into it so we thought that let's talk a little bit about future of work okay as someone that has like visibility into a lot of things that goes goes on in the world of AI and these technological Evolution what I was thinking is that how are we thinking about AI replacing jobs is that something that we have studies on what are the you know what are the impacts of this technology right now in the future for yeah for sure first of it's good to see you rine it's been a while you know we first met in Central Park I biked over we were looking a little more disheveled than today so I'm glad we both cleaned up but um AI certainly it's top of mind there's no cxo or board conversation we have today that AI is not you know part of the agenda there are many studies that have happened based on the Persona work the different levels of optimization Etc we certainly think for most of our clients and we serve 95% of the Fortune 500 there's is going to be a significant impact anywhere from 20 to 40% depending on the persona but that also means it's going to open up all kinds of new opportunities that folks are not focused on today even for our business while we may find that we will get to a point where we'll have a 10x auditor a 10x analyst or a 10x tax professional there's so many more things that need to get done that will unlock through AI so I'm more optimistic than most in terms of the real opportunity had of AI in terms of doing what we do today better cheaper but then opening up all kinds of other avenues amazing and then given the speed of AI advancement you know what can businesses do to Future proof their operations because now we we are having concerns like people are losing jobs or you know how are we thinking about these processes and how are we making sure that operations are getting done right from the Enterprise perspective and from the people's perspective yeah I mean I don't know if there's any such thing as future proofing the business I think we're in this stage where we're going to constantly change and be ready to embrace that change I think what you guys do at liquid is a case in point for many of us right just in terms of constantly challenging orthodoxies you know pushing the the envelope on rewriting the rules and things of that nature I think organizations that need to be resilient need to be similarly configured to be constantly challenging orthodoxies urban legends to build structural agility so they can pivot more quickly as has no new shifts come whether they are technological or regulatory or geopolitical and being able to do that is the best way to be able to Future proof and then constantly rethinking their products and services and business models and we do a lot of that and you know we're we don't take that responsibility lightly because that's a great opportunity for our clients to also rethink their business constantly and you know where we're partnering together on a number of those as well yeah let's uh talk about that a little bit so how are we like tell us a little bit about the partnership that we trying to build yeah so you know like the governor said very eloquently we are in the business also of Applied AI applied engineering applied Tech applied AI so we we our partnership is quite perfect in that regard because we very much focus on domains Industries functional domains Etc and so if you have to talk about um anti-money laundering and financial services or investment advice for you know investment manager which is your background as well or disease pathology we can bring some of that deep Insight working with our clients and with liquid AI to bring that to life both in terms of poc's but then scaling rapidly across the Enterprise and as you know the Enterprises in general are slower to move and navigating them and being able to have them Embrace this and be able to leverage and show results quickly is critical and I think that's where you know we naturally come together quite effectively so that's one of the values that we unlock is like one of the products that actually we put out is private machine learning so can you give us some examples of how we can use that yeah most certainly right you know we serve in large number of regulated Industries in fact we are also partly regulated because of our audit and Independence business and a lot of um our clients just like us are not going to at any point be willing to put our data out there it's sensitive it's our IP it's critical to our you know mission and vision so as a result in a future world we absolutely Envision a hybrid world where there's going to be private llms and you know private lfms as well as in the public domain and depending on the workload the sensitive of the data the regulation the geopolitical environment it's operating in we're going to be able to have to navigate those you know seamlessly and hence the opportunity to create a single pan of glass be able to create operational and you know trustworthy view into that world and make navigate that effectively is is quite key amazing so tell us also a little bit about your uh view on the challenges of AI right right now what kind of challenges are we seeing that the Genera obviously everybody wants to use this technology everybody is like you know hungry to getting access to this gen General from a technology perspective I can tell you that what we see is like people have to get technical you know this technology is deeply technical you really have to understand it but what are the challenges do we see from the adoption point few you know and what can we do to to facilitate that yeah I that's a great question so you know over the last couple years at least there's a general sense of realism around what AI can do and how to adopt and apply it so if I was to maybe look at the three things that bubble to the top the first one is really a focus on real Roi as in you know are they be getting the value for the investment and what is the line of sight to that value and the fact that you fundamentally rethought the model to get to a more efficient model and a step function change in the price performance curve makes a huge difference just in that conversation the second big challenge that you know our clients continue to face and us as well is still trustworthiness trustworthiness observability the white box con construct that we've talked about as well is quite critical because for regulated and unregulated businesses being able to understand explain Ai and the explainability is quite key and part of that is also the financial governance around it observability and the third which I think is really Critical with all such change is the people dimension of this how do we drive and embrace change in a way that people are not threatened by AI but can embrace it and do other things and be able to uplift their own Focus Etc so we're focused on all those three things and just like you know other shifts that have come in the past we expect this to also be embraced is just coming at it us much faster so some of the existing paradigms have to be challenged and rethought and we're in the midst of that you know all the time amazing so do you think in the future we're going to have ai CEOs and CTO you know never say never right I think I'm quite easily replaceable maybe already U but I think there certainly are you know you know ways certainly already we you know it's not already people have you know CTO co-pilots and CEO co-pilots and the like which certainly make them smart in every conversation we use it all the time you know we have a big team here number of II leaders are in the room number of them are from Boston and from Massachusetts in fact our AI leader Jim is based here so even the governor's comments about making this in AI Hub is quite key and we are very focused on creating the right personas and agents for all of those personas right a tax analyst a CEO a CTO so you know it won't be long and maybe we should do that together y replacing ourselves that's right okay so and then um last question for you as a business executive what are your applications of gen are you using AI every day yeah we're using AI every day we have been for the last couple of years now obviously that you know more than anything else to one set the tone across the firm so our leadership team our CEO down you know is a big embracer of AI so you know that also helps us better understand what's working and what's not working we're using it in collaboration plan platforms using it to prepare for meetings for research reports for insights in various conversations Etc as well but we're also pushing it down through the rank and file obviously you know our developers are doing a number of autonomous coding platforms today as well as our implementation work our tax and audit professionals as well so very much and like you alluded to you have to set the tone at the top otherwise you know the change just becomes harder and you know takes longer absolutely thank you so much ranit give it up for ranit B good to see you thank you thank you please help me welcome Raymond Lea managing director Samsung [Music] next hey long time no see Raymond oh yes okay so let's get into it very good from your point of view why do we need General on device from a technical and from a business perspective yeah so this morning I was still in Orlando attending gner it Symposium all right the one B that across the whole Symposium is AI agents okay this is one of the largest gathering of cios okay so now from business angle we say okay what's going to happen with the generative AI which is two years old so far most of the activities has has been on the infrastructure layer but the key question is what's going to be the application the sense is that it's going to be AI agent empowered Services endtoend Services because now you have ai you can automate a lot of things you can deliver end to and services to user and where where would the request being generated and the service be consumed by most of the users it's likely going to be on your smartphone or future even smaller devices you wear regardless whether this consumer or business so that's why from our perspective gener generative AI is actually big opportunity for a device vendor like Samsung okay my personal opinion is probably bigger than the mobile Cloud era and the techn technology wise it makes a lot of sense because when you think about put the user first what's the user experience can they tolerate the latency can they tolerate sending all their private data to the cloud or from the service provider perspective can then tolerate in burning all those dollars on inference serving all those hundreds of million of users so there are many reasons technically speaking we need to have more compute and more data stay at where they are you move the compute to where the data are those are the personalized data personalized experience that belong to the user stay at the where they are what makes it challenging like to bring AI on device the um what do we see already right that many I would say Flagship apps they use many models okay it's not just one model power one application that use many models now you say you have in my phone you have maybe tens of this kind of applications then think about how many models I have to maintain and manage that lifetime left cycle in my phone so it's very challenging and I have a power budget this thing should not join too much power I have to make sure that patch right away if there's a bug in there there a vulnerability happen I need to patch them right and you guys already showed it right it's actually still hard to shrink a 70 billion parameter model to 1 billion but it's non trivial this require a lot of personalized information you understand us as contact text uh and also their data right it's long Tri and it's not saying that I can have just one small model marvelous marvelously can serve all the users yes so for both Technical and the business perspective is non trivial what do you think uh the liquid models you've studied them pretty well I think you've read all of our papers already and I wanted to know what's your perspective on what what is fundamental and unique about the liquid foundation models the way you saw it so you see fit in this space yeah maybe I di back a little bit what does Samsung next do right we are an open Innovation on an adventure investment primarily for Samsung Electronics okay we make one new investment a week we are very early how early we go to Major AC academic conferences where we meet you know Ram as his whole team right we read their papers what's standing out honestly is that one I see from the early days of nuclear neuronet to when he tells me oh I'm merging with a stat based model I can do 100 of million parameter model now then I say okay that's very different now I get even more excited when they start telling me I have this mechanistic design methodology the diff Kate that's a big deal okay that's really a big deal I use the analogy it's like in the semiconductor industry before defate everyone has to design R their Circuit by by hand okay every design silicon design is B spoke it's a dark magic based on experience today's L is like that you need some people with special talent and they canot tell you why they know how to tune the hyper parameter of the model and make the model work but this defate it's like you have a Ed tool for second design so now I think our R&D team have a really powerful methodology and tool to manage the nmare they are facing managing all those hundreds of models on a phone because there all the different tradeoffs and all those techniques are already built in inside defk plus they also have this IP library right I call this IP Library they have their own Foundation models they did a great job just put as just same thing like as your IP core inside your deficate people can license it okay that's really a big deal no one has done that so far thank you so Raymond what why do you think oems like like Samsung should work with us well there's no other choice right from what I just said right from business Direction this is what we want to go after um we want to work with the the greatest Builder out there right uh I see this team every month there's something new literally every month there's something new if you check their history how the three co-founders founded at MIT it's not about this right every month they invent new stuff impressive so this is basically showing the strength of a really creative team and also shows the you know the management how they can attract talents okay and how they can do business partner with companies big like Samsung like deid and attract also heavy amount of know uh Venture uh Investments right you need all of these not every entrepreneur can do this so so far we are very excited to see luid AI being know progressing so well so fast on the global stage amazing so I want to say the final word but um you know like I might be a little bit biased but Raymond is one of the smartest investors that we have and um and and we really enjoy conversations together yes and U yeah thank you all right let's wrap it up ladies and gentlemen thank you Raymond this company came very fast it just sort of happened we were writing papers and all of a sudden we said okay we're going to bring this to the world Ramen called and said hey we we started liquid how about you join us and we do some some something together and then I dropped everything and and joined I believed in their ability I believed in my ability and so it was a very easy decision for me to say yes last fall our our Tech team was basically three people in a shared slack channel to now like you know we've grown quite a bit um but still you know we managed to keep that like very special feeling I feel like like what we built is so exceptional and it's just such a such an as I mentioned like said humbling experience to really interact with our with our people what I love the most about liquid is really the team that we've built and the people but I think what what's also differentiating or driving us to to to what we do is that we that that there's an ambition to do something outstanding that truly to change the game we don't really care about all of the additional things that make things look good on the outside but really aren't good we really care about the core attributes of what people are building we wake up every day and we're solving problems that the rest of the world has not solved yet we're approaching this by not taking other Technologies and trying to wrap them and package them into a solution we're approaching it actually foundationally my biggest inspiration was really answering unanswered questions and discovering solutions that people have spent years decades trying to solve I mean in some sense this notion that you can actually uh create a brand new machine learning model that has something in common uh with small creatures in the natural world is kind of at the intersection of Science Fiction and science not every day you know like Haley's comment that something comes around you're like holy [ __ ] I mean this ain't a lemonade stand this isn't a science experiment this is something that Can Go the Distance we see a lot of the field is driven in hype and flashy demos but we actually hold value in something completely different we don't want to build a company driven based on hype we build our company based on foundational technology that works and is practically guided to deliver value if you scale things things are become like new kind of uh new Behavior get unlocked you know you can build general purpose systems we've seen you know the the power that AI can bring us and it opened up our imagination you know like what is possible with this kind of AI system being in this kind of world where AI can become smarter than the collective knowledge of the humanity that's like the most exciting thing and it's the one of the most important thing that happens in the world and I feel like like Technologies like ours is going to just be absolutely necessary in this world [Music] I need him oh you do how long have you been doing this like five minutes okay do one more five minutes ladies and Gentlemen please welcome Danielle aruse director mitc sale and co-founder liquid AI Joshua Bach Chief AI strategist at liquid AI male Perkin CTO Shopify and Sebastian bck VP of AI at Microsoft hello everyone I hope you're having a a rich morning I hope you like um what you hear and hello Seb it's been a while hi um and in particular yes yes it's been a while uh since we have seen the Sparks of uh AGI so uh can you tell us a little bit more about what has happened happened in the field since that momentous visit you made to Cale MIT in March of 2023 yeah I I remember fondly when you introduced me for that talk which was you know a very important uh moment for me so I think in the last 18 months or so since that presentation basically the world has come around and hopefully now everybody agrees that these were really the Sparks of maybe not AGI but at least something fundamentally new and just to to to give some perspective people like teren sto now you know the famous Fields medalist he's saying that the latest generations of AI basically would qualify as a bad PhD student so of course not a great one but still you know PhD student level this is just you know absolutely astounding so um that's that's really extraordinary isn't it um we have a bad PhD student that's built on terrific AI models uh that work uh based on scaling and so I wonder what do you all think about scaling is there a limits to scaling can we get from a bad PhD student to an excellent PhD uh student by scaling Alone um questions for all of you maybe mik why don't you start well that's that's a very interesting question you know on one hand if you look at the log log plot of all the models and how this quality improves with the scaling so far all the models are on very close to straight line and they're this seems to be no end to scaling at the same time we have existence proof that scaling is not everything you know human brain and you could see that both in terms of power consumption and overall complexity we will be exceeding it soon and not not really capturing what brain does so I think uh and this is where I'm very happy with liquid because I have I have a little bit of a physicist background and I've seen you know finally we have differential equations coming into machine learning rather than just just matrix multiplication and so here uh if you look at say diffusion models there we see constraints and a lot of really uh ability to extract more from the same scale with L Dynamics and and physics inspired uh sort of ideas brought into the ideas of inference so here I'm thinking that uh there got to be again we have existence proof that there got to be something Beyond scale and we I'm very happy that liquid is exploring that josha what's your take scaling sealing or not sealing I think that scaling is in a way a cheat it's a way to deal with um a world in which you cannot play yourself in which you cannot make experiments yourself if you can do selfplay like you can do with chess and go you don't need to scale scale in the same way as you do with say llms and Foundation models and uh in the meantime there is a deeper question the universality hypothesis uh behind the scaling hypothesis that is based on the observation that if you take different uh newal architectures and machine learning architectures you give them the same data you give them enough data you give them enough compute they end up with similar model capabilities with a similar structure in the model and there is the question does this also translate to Minds so if we were to uh train a model on the input and output of biological organisms of human beings for long enough with enough data would it turn into something like a human mind and at the moment this is not clear it's an open hypothesis I think we can see that scaling makes models better we we don't really know if we run into a plateau with scaling we also don't know how long it can continue but um as Mel said the human brain is an existence proof that you can get to intelligence not by scaling but by being smarter about it so Sebastian what are some of the Technologies we should be aware of uh as we think about achieving models with better and better um function yeah so maybe let me relate this question to your previous one about scaling so I think you know scaling is is only one part of the puzzle and and in terms of new things that came out in in the last you know year or so I would say that there are essentially three Trends one of them is still of course the scaling up and we want to make bigger models and more more capabilities are going to be unlocked that way but another Direction which you know uh liquid is is also pionering is the opposite to make the model smaller and by making the model smaller you also discover new capabilities for example just the cost efficiency the speed which allows you to do all kinds of things that would not be possible with the very very large language model so these are two direction scating up and scating down and then there is a new Direction which I'm personally extremely excited about and I think unlocks a completely new Direction which is to think harder so not only you know have a bigger model but have a model that tries to do more cycle before it gives you an answer so of course the one demonstration uh right now of this new paradigm is is the o1 model by open Ai and this is the one that Teran St was referring to as a bad PhD student so I think this new this is a new technological breakthrough this uh thinking harder and I'm very optimistic that we haven't seen you know the end of scaling in that direction so Sebastian you say think harder by doing cycles that sounds like feedback but Transformers don't have feedback uh they feed forward mechanisms do you care to comment on that yeah so right now you know of course so first of all I don't know exactly what was done in one uh and even if I knew I don't think I could say but I I don't know but what I can tell you is that you can try to basically mimic the feedback mechanism on the transcript of the Transformer so of course this is you know maybe not the ideal way there could potentially be better way to incorporate this feedback but there is at least an easy a kind of hack to do it on the transcript of a transformer and the amazing thing is maybe this is not the best way to do it but it already gives you an incredible model so you know can you imagine what will happen once somebody unlocks the right way to do it I mean this is going to be probably very spectacular Joshua what do you think about uh feedback is it important and how should we achieve it if it is I think if you have a feed forward model that is acting on memory then it is in some sense recurrent so there is some feedback but uh in the way in which you mean it I think there more intelligent way to give feedback for instance when we have a conversation after the conversation we often rethink what happened just now and uh we create a very differentiated reinforcement learning signal for ourselves that leads to us updating ourselves and in principle it would be possible to also have AIS that continuously learn with every interaction that they have with the world and with themselves and in this way creatively transform themselves to get better and better and we are just getting started in this line of research and I think that the Chain of Thought that we see in um 01 preview is going in this direction and it's also something that I think most AI companies are right now exploring and their research but uh we feel that we are just beginning to tap into what we observe humans being to be doing when they learn so I I'm watching the time we only have time for a quick question and so let me ask around Robin question and this is about the impact of the various aspects uh of the total per performance of a model of a large language model so what is the Divide between data curation pre-training post-training and model architecture and are there other factors what do you all think let's start with mik uh yeah that's a good question I mean to me honestly it sounds a little bit like what's more important for human life oxygen water or food I'm like well technically oxygen but you're not going to be alive without either of them so uh I think uh I think we learned recently relearned the new appreciation got new appreciation for data uh I know Sebastian got a lot of mileage of just thinking very carefully about the data uh I think liquid is showing that there is quite a bit to be gained in architecture itself which for a while honestly like in my circles people were like architecture everybody is the same like you you don't need to think about it and then like secret comes in I think uh the only one immutable thing that I think stays for everyone is just the amount of compute and we know this little dirty secret of the industry is that uh compute trumps everything and you know this is for us to solve still Seb how do you see things I don't have a much better answer than what mik just said uh I mean except maybe I would add that I think we are only at the very beginning in terms of the potential in terms of creativity for data because as you know you know I I I'm very very excited about synthetic data and once you get into synthetic data well you have basically full control on what you're going to teach your AI and and you know your imagination is the only limit at that point so I think we can go a lot lot lot further in that direction but then again everything that mik said is is correct so you know architecture comput all of those things do matter Josh are you're the closer I think we can take a page from human beings we have several faces when we learn a childhood in which we need a safe environment to boot up our capabilities and then we have a combination of online learning and we also have offline learning dreaming where basically fine tune in synthetic data and it seems that all these are important but uh machines are not limited to this they can exchange data between models in artificial formats and so on so uh there is a way in which we see biology does it that is inspiring us but I think we can also move Beyond it in some regards well thank you all so much this is all we have time for there is so much we have to do in science in order to better understand Ai and there is such a great interplay between understanding ourselves and understanding Ai and we have many questions we need to answer sooner rather than later so thank you thank you very much let me introduce again the CEO of liquid AI Ramin Hasani and Ichiro sug chairman uchu CTC [Music] right okay so we have sugan chairman of CC one of the largest system integrators in Japan so we recently built a relationship together yep tell us a little bit about that okay uh just today uh the payment has been done so then now we are the official partner in Japan and we will start the strong partner parip in in Japan from now on Amazing may I speak in Japanese no you can yeah can liid way help help translate no not yet not yet okay not today all right we'll do that in the next one okay so sugan how how are Japanese Enterprises thinking about AI uh yeah they they are Mo most of the most of them uh thinking that they like to use AI for better productivity of course but at the same time they they're concerning about the privacy or security things so so then uh in that sense I think uh liquid liquid AI would be the suitable tool for for for them absolutely and do you think you know AI is an important matter to be integrated in companies are they traditional or are they thinking about uh revolutionizing their Industries with AI how is that process uh the before talking about Enterprises uh we J Japan has s suffered uh from uh uh deflation economy for a long time and now it is changing and uh we are change shifting to our growth economy uh most of the uh executive Executives remember that we we need we need to make investment for growth uh finally we remembered and then uh uh and one more thing uh in Japan the the population I I I mean okay uh working age population is decreasing so then sometime in 80s U it was about uh 80 87 million people uh the working working age population so now maybe about 70 million and then uh in 15 years or 20 years uh L than 60 milon almost you know 30% less than from the peak so then we need something to to drive the growth in Japan and I I think uh most of the Enterprise will use liid AI because so many so many areas uh it can be applied so but I personally expect uh uh some good Solutions in agriculture agriculture things because again in in Japan uh self how to say it food self-sufficiency rate yeah is quite low uh it's about 388 8% it's the uh calorie base but if you take out the the rice from there other the other very you know low so then uh to secure the food uh not only for me uh but it it might be uh the good good you know uh tools for for them the uh the agriculture is the one one of the area where the real management is necessary in in Japan and also technology absolutely so we recently opened up um branch in Japan uh this is liquid AI Japan and we are very excited to actually take the technology and together with um CTC to deliver this technology like to worldwide like in in that kind of uh um center of the world but what I want to um mention as well is that like we developed a like an example of the stuff that you saw in the previous hour was a Japanese Foundation model that we built together we're very excited about that that model is a very capable language model and we want to adopt that um you know like as one of those examples that how we are working with Enterprises like CTC to actually bring value and unlock value so zasan what are the Enterprises like the sectors that you feel like this technology can be integrated you mentioned agriculture oh yeah but what else uh I think uh so so many but uh maybe customer service area and the financial uh Institute and the manufacturing th those are main you know stre I think absolutely all right and what do you think is the rate of adoption right now in Japan of of AI oh very few very few it is said that the market market right now this year is a about uh uh 300 million Us doll a year but it said it will be ER more than 35 times in five or six years I don't I don't I don't know but uh it is said so just just start it so then uh NE next next year we will see the totally different you know landscape there you saw the governor of Massachusetts spoke about the adoption of AI and how pumped up the the state is about AI how is the situation there with the government uh as I mentioned uh the Japan Japanese government uh remember that we need a new investment so then they they're think thinking very seriously about this and just they started to to support uh not only AI but Advan advanced technology things um main the two year from two years ago they they have started to do that great and do you think the Japanese approach right now is build versus buy are they building do they want to build in-house or buy from outside ah Case by Case Case by case some of the big Enterprise are doing uh by themselves but but uh it's not uh sufficient by only only those you know we need to combine and or hybrid things 100% yeah all right thank you so much ladies and gentlemen thank you thank [Applause] you ladies and gentlemen a special announcement we have a special guest at the end of the show so don't leave your seats no matter how hungry you are now I would like to introduce Keith Williams CTO of Cap Gemini [Applause] engineering good to see you good to see you good to see you right we see we turn a little bit this way so you can see the so we have something to show here yeah we do so Keith tell us the story how how have you got to know chair so um it started about four years ago when we signed a research collaboration with um MIT uh with Danielle's group on trusted AI because that was really important to us um and then about and that's been great that's been a big success and then about a year ago we found out you guys um were spinning off so we join that party if you like through a um an investment through capture our Ventures and we've now got a global Partnerships which been a lot it's been a long a long relationship we've had yeah and it's great to see you at this stage really brilant it's brilliant the first Enterprise that we ever interacted with was Cap Gemini because it was like through MIT connection like we were doing research together and then yeah we got into more commercial aspects of things so tell us a little bit about augmented think we need to click the slide do we oh yes we do beautiful okay it's not what I was expecting Ah that's that's one one yeah we'll go here ask me the question it'll work yeah okay so augmenting engineering yes so we we tend we talk about um Cap Gemini the reason we we got involved with you guys is um we think about um I'm the C for engineering so I'm concerned about products how people build products whether there be aircraft CT scanners cars Etc and and how you build those products and how you put AI inside of those products and we're pretty you know we're convinced that AI um is transforming both the way you build engineered products and also the products themselves right but we see it very much as a as an augmentation of that activity it's not it's not some kind of magic dust that you sprinkle on engineering and all your problems go away which is some of the narrative that you almost get um so we tend to talk about augmented Engineering in terms of using AI tell us about the you know when you look at AI applicability you know to engineering projects tell us examples I me you know any anyone involved in in um product engineering will know that it's dominated by by two things I think correctness okay and constraints so you know if you're building an aircraft if you're if you're building a car or whatever it is okay um a good guess is not really what you're after okay um you actually want to know whether it's correct or not so we've generally been looking at Technologies around AI that give you help you solve that problem of correctness and the other aspect is constraints and those constra constraints are twofold right you've got product constraints which you've touched on a lot today things like memory footprint you know realtime um performance you know you can't tow you can't tow a server behind your car right um and um the other constraint actually is the process you know people have mentioned regulation a lot of our Industries are are dominated by standards and regulation Etc so any technology that addresses the correctness question and helps us address the constraints whether they be memory or regulation okay that's why the explainability is so important is really important from from an engineering Viewpoint and how is Capt and deploying AI for customers so we've got I think we've got a slide we'll try there we go um uh part of it anyway so one example that we've done I talk you through a couple of examples I and they use the different they use the different products that you've that you've got okay one example here is for um um uh a car handbook this is for you know you got you've got an 800 page owner manual if if you buy a car these days and no one re really read no one reads it and you need to be able to interact with that with that manual um effectively and nlm is a really nice way to do that but of course you know from an automotive Viewpoint you're limited with the compute okay you cannot assume that you've got you've got connectivity so we managed to deploy A really lovely owner's manual type interaction on on on just a conventional Hardware that you would find in in today's automobile okay which is a really great demonstration of a practical application of llms when when they size correctly the other example that I think I'm really not yeah now we okay hold on whoa you should have Ah that's the one I want that's the one I want hold on this one this one um we did a another one that we've done is a is a project around optimization of telecom's networks right and if you know this but you use a huge amount of energy and cost in in mobile networks and we've done a great project where we've been modeling switching an antennas off okay in order to save energy and cost which you can do at certain times okay if if the the operational conditions allow and we've been using your technology we've been benchmarking against you know other um state-of-the-art um deep learning technology and basically it's more accurate okay it it more accurately predicts the number of users that you're going to have it it accurately predicts the packet loss that you're going to have okay and that enables you before you do take that decision to switch an antenna off to know that it's not going to affect the quality service of the of the user right um so we really see those kind you know a whole range of those kind of Time series type problems is is is a real strength of your technology amazing I don't think we played the previous demo like can we can we see it I I don't let me maybe you can work in AR beond let me see okay so this one let's go on back okay just go play I'm play one more time and one more time there we go all right so you got the car manual over there and a liquid model is actually offline is actually taking um you know information from that car manual so you can imagine like uploading a massive amount of documents in there right finally we'll show you the audio mode in this mode users can ask their questions using their voice the speech is transcribed and automatically submitted the AI resp responds by writing the answer and then reading it aloud watch the demonstration okay great and obviously for for a lot of clients this ability to deploy an nlm like that in something you know this just a power of a normal PC is just you know hugely valuable for for all a lot of our customers in a lot of different use cases yeah looking in the future where do you think AI is going um for you for capture but I think we're I think from a from an from talking from an engineering Viewpoint I think what we see is what we would call hybrid AI is the real when you look at it from an engineering domain we think you've got to combine um different kinds of AI Technologies but also with other types of engineering mathematical models okay that help you solve the correctness problem and obviously your technology helps with that right and I think where you put that that knowledge physics where you put that knowledge of the system whether that's in the AI or it's in a different model of some description I think is is will be interesting to see but I really think it's going to be a hybrid approach of different kinds of model within the context of of engineering do you believe in small models or larger models do I believe in small models or larger models as an engineer I would say that would depend on the context obviously a smaller model is better by definition but it depends on the quality metrics so um for that so I think I think it's I think it horses for courses I think it depends but but I I definitely believe in models that are tuned to specific problem that you're trying to solve absolutely absolutely being big or small it depends what the problem is you're trying to solve and reliability matters reliability can H fantastic thank you so much for joining us thank you Ladi and gentlemen thank you Keith Willam thank you so much hey folks more information on this special guest it's a surprise that I'm going to give away right now because at the end of the program Steven paga co-owner of the championship Boston Celtics and former chairman of Bane capital and liquid AI will come on stays with John Warner and have a conversation so stay tuned but now I'd like to introduce Alexander amini CSO liquid AI Stuart shriber CEO of arena bioworks in a conversation for you to enjoy thank [Music] you incredible thank you Stuart for joining us today hi everyone I'd like to start by first of all maybe uh allowing you to introduce Arena bioworks which is a a biotech right here in Cambridge maybe you could share a little bit about Arena what the vision is what the mission of the company is and then we'll take it from there sure um first of all uh it's been a great pleasure working with you Alexander and your colleagues uh I hope we'll get a chance to talk a little bit about that and congratulations so uh happy to say a word about Arena bioworks uh we are a human biology based Center uh that whose mission is to understand the mechanism of disease in the context of human physiology and to translate those insights into new medicines through the formation of Biotech companies within Arena now we're able to do that because we are privately funded and so we study disease and translate to medicines we say Under One Roof and if anyone has seen our logo you'll see the little roof and so private funding turned out to be important amazing can I say I I on the way in I I heard Steve puka mention yes um it it really helps when he he's you may all be surprised Eve is intersected yet once again um we have an amazing group of of visionary investors and Steve is one of them he's actually our executive chair um but their alignment with the scientist of tackling big unmet medical needs and pulling in modern technology um to get there as quickly as possible is kind of the magic of Arena amazing yeah I completely agree and I think one of the the fascinating things about the biotech field and maybe you could speak to this a little bit is how multidisiplinary it is is and I think the the ability for AI and geni in particular to really start to accelerate a lot of the initiatives that you have going on and the problems that the the company is tackling is really transformational maybe you could share a little bit about the the problems because I think as you said it starts with those but also your vision for how gen fits into this space sure Alexander um Let me let me pick one or two examples actually I'll pick two and I'll pick one that happens to be what we've been working on together um and our collaboration our partnership uh but I also want to share one that I think is going to be a little bit more challenging maybe take a little bit more time but ultimately will have tremendous impact on human health so so the first one um actually I think it was Alexander in your presentation you you talked about design proteins and it's we all know it's kind of amazing what what generative AI has done in predicting protein structure and designing protein structure that doesn't mean all the problems have been solved and one of the ones that we've been tackling jointly is that these beautiful structures we see in designed proteins are largely static now we know that in the body proteins move around they're highly dynamic and we know that it's essential for their function you know when an enzyme a protein enzyme catalyzes a reaction there's all kinds of side chain motions that are critical along that lowering the overall energy of the the potential energy curve of this reaction um when a drug binds a protein it's rare that the pre-organized structure of the drug and the pre-organized structure of the protein simply fit together by lock and key and yet that is where we are today it's largely the structures that we create or static I'm optimistic that this is a problem that we can solve I think there's a lot of different approaches and we've focused on on one um I'll point to my my personal experience in the early 1990s uh when I met Martin carplus and Martin um um who who's a a giant in the field came to me and and and actually talked about the same question of dynamic properties of proteins and I remember he he in his case he wanted to integrate what was called molecular mechanics and quantum mechanics and um his work was uh recognized in a Nobel Prize years later and it really put on the map that proteins are highly Dynamic and there are ways to compute th those Dynamic there is very limited in what you can do but I think this is the time to rethink that problem from a generative AI perspective and couldn't be more excited about doing this jointly with liquid absolutely I think actually so one of the things that you mentioned that got me really excited originally was that this this Vision that you just described is very true to the the original vision of liquid as well the world is not static in these snapshots the world is liquid it's very FL it's adapting it's changing it's Dynamic over time and that's actually the roots of the architecture that we had built in the very beginning days so I think it's a it's a perfect problem statement I also want to tie that into something that you and I were talking about backstage which is the importance of basically problem first thinking because this is also something that we hold very dearly in the company that we see a lot of AI companies out in the world really based on hype that go after just new types of technology for the sake of building new technologies and then Plug and Play into existing uh you know problems that without innovating on the problem itself could you speak a little bit about from your point of view especially in a field like biotechnology the importance of problem first thinking yes I'm very happy to I was listening to these talks it was very very interesting but I I I confess that I I feel very strongly about the notion in in biom medicine in the human body it's exceedingly complex and it really helps to start with the problem to know what is the problem we're trying to solve because the alternative of creating massive data um using generative AI will will very likely provide useful insights but not necessarily to solve the problem you care about um I said I was going to give a second example I'll try to tie this in I going to sort of make up a little things as I go along to to try to illustrate the concept but um I think that the human body as I said very complex there's there's a lot of things we could measure happily there are things today and there we at Arena are working very hard with partnering including with Sovereign Nations to um have access to large number of participants large number of uh samples from the body longitudinal but still what are you going to do with all of this so um I'll give you an example of the human brain because it feels appropriate with liquid um we are we have major unmet medical needs and neuros psychiatric disease neurod degenerative disease um it feels like some progress is being made we read about new medicines but but those of us who study this really carefully I think the lack of progress is more notable for the the substance of progress and so I think you know if we tackle those problems and start with the fundamentals like what is a what is a brain what is it composed of and I think one of the things you know we all intrinsically know our brain is kind of squishy I think I heard reference to that with the the liquid model there's a reason for it it's a sack full of what we call lipids in Biochemistry yes and these lipids are very Dynamic and they're changing and I think that if we um focused in on neuros psychiatric schizophrenia or neurodegenerative Alzheimer's and got away from the biases of what we think is going on you know in Alzheimer's we know that there are these ameloid deposits but are they are they clausal or are they correlate and just take a step back gather the relevant data and the relevant generative AI I strongly suspect that whole new Concepts will emerge and they will provide a blueprint for new directions in creating the medicines we so desperately need so I I picked it as a kind of hypothetical it's not entirely hypothetical we are doing work in this area and and what I've outlined is sort of a summary of what what we've seen thus far amazing amazing I think I mean especially just building off of that this Loop of basically sensing and actuating right we want to sense the body but then actually build back and act on it and build new things that we would want to sense from that Loop is a very iterative process that I think it really also comes at the core of the the mission of liquid so maybe I just want to end on on one more one more question because you mentioned earlier that Arena and liquid will be partnering in this space of biotechnology and gen at the intersection of the two maybe just a quick comment for the the audience on why you believe this relationship is so important for you and for the field yes Alexander it's a really great question and it's another thing that listening to the conversations it keeps coming back to me you know we looked at a lot of groups in this area and to be honest many of them sort of felt familiar that because you know I think we at Arena we have basic knowledge of what generative AI can do um machine learning can do neural Nets can what the contributions and we've we found a lot of sort of like-minded people and we were we would debate one problem over another I think with liquid uh we found almost we found two groups that have deep domain knowledge one in in medicine the other in generative Ai and so we had to work hard to get across the language barrier but we were all super excited about the fact that this was potently synergistic and not you know know redundant in any way so I think the Deep domain knowledge and expertise of liquid um and an interest in listening to us think about some of the big UNM medical needs uh was kind of magical absolutely Stuart thank you so much for joining us today thank you everyone [Music] please help me help me in welcoming back to the stage Daniela ruse director MIT Cale and co-founder liquid Ai and Ralph whittig head of R&D at [Music] AMD hello again so uh Ralph I'm so excited you are here with us today um MIT and AMD have been friends uh we've worked on so many things together and we're so excited uh to think about a partnership with you in the future um so um please um uh welcome yeah thank you Dan you know it's a pleasure to be here and thank you for hosting us over here today to the Fantastic launch event quite excited to see uh Team like liquid come together and uh and and really make the jump from from Research into uh the startup era and uh we've got a lot of good things things we can work on together um uh AMD research and MIT has been you know a strong partnership over many uh decades already and uh uh we're very excited uh to collaborate now on the AI uh research as well um we also have a lot of MIT grads coming to AMD there's actually a big facility right up the road here in boxboro AMD still does uh uh a significant uh processor design uh right here in in the Bost area so all our big CPU are designed over here the cloud CPUs um and so we really could not be uh filling this Talent pipeline without the strong collaboration between MIT and AMD so um Ralph can you share with us uh the range of products AMD is focused on as one of the leaders uh that is building the substrate for AI yeah no that's a really good question yeah so AMD is actually a full port portfolio we go anywhere from uh the cloud uh high-end gpus that are really used to train AI models uh CPUs uh so the Instinct line is the GPU line the Epic line is the CPU line uh they really power all the data centers around the world and they also sit in supercomputers and National Labs um we also have the um gaming gpus there's a lot of AI going into J gaming now um and on the uh client side so all the laptops that you you're you're running those are our um uh apus those are s so's that are mixing of CPUs gpus and also npus and then we also have the embedded product line uh coming in from the xlink side uh where we get into uh uh close to Edge close to sensor applications really highly efficient and uh and it's really ubiquitous in all kinds of applications right whether that's uh factories uh or uh cars uh or healthare it's actually quite diverse right so AMD really has uh uh processors for all market segments uh that are that are uh powered by a AI now so all the way from the clouds to the edge uh kind of like how we think about uh AI all the way from Big models to to small models yeah and it's really it's interesting right now right so we're in you know the models are being trained on on these big gpus the big gpus are um quite power hungry right so the city around a kilowatt per GPU um and we trained scaling to hundreds if not thousands of gpus right so these big training data centers are sitting in the megawatts of power right and uh we heard about the talk about scaling right how far can we go before we run out of or hit a plateau um the scaling continues right and we're going to see uh data centers uh approaching the gigawatt uh power usage right so we're getting into training clusters in the tens if not hundreds hundreds thousands of of gpus right that scaling is definitely alive and well um in the end it's going to be Energy Efficiency that's going to enable uh further scaling over there that's a real problem yeah I mean in fact Energy Efficiency uh is also one of the drivers of the work at liquid uh uh traditional AI models cost so much in energy and they have a huge environmental footprint uh we are very interested in uh more energy efficient approaches so can you give us some statistics on the energy cost of AI for for development and deployment yeah that's also why we're excited to be uh partnering with liquid the on the cloud side as I said it's a th000 watts per GPU scaling to hundreds of thousands of gpus you're getting you know megawatts going to to gigawatts right on the client side your laptop uh that can't burn more than tens of Watts right so 50 watts is already fairly power hungry right the actual processor inside the laptops sits at 10 watts or below right so now you got a single GPU at 1,000 watts in a cloud and you got an APU sitting there at give and take 10 watt power consumption and the actual npu inside your laptop the neural processing unit is is uh is is in a single uh digit Watts unit right so it's quite a range we're dealing with over there now uh the Innovations we contribute is process technology processor architect we're building specialized uh uh processing fabrics for training we're building specialized processing fabrics for inference right but the contributions uh the Team Liquid is making is are tremendous because we also have to uh simplify the models uh shape them make them efficient um this is all very important right so anywhere from actually training the foundational models is important the contributions you make are really is is really great but then also the the broad Market is taking the models and is fine-tuning them for their for their uses right so the the models have to fine tune efficiently as well and then we have to slim them for a very low efficiency Energy Efficiency uh uh high energy efficiency low energy consumption on the client side right that's so exciting um listen I I just want to move through a few other topics uh before our slot um uh comes to an end uh can you tell us more about uh client and embedded products and how they impact your revenues yeah so the uh uh now of course I got to be careful what I say I'm not uh uh the financial uh uh representative of the company but if you look at public uh uh uh uh quarterly reports you'll see that we are sort of balanced between the cloud side the client side and the embedded side right now the current fastest growing segment is the cloud because we're pushing into really more training and deployment the actual uh models are trained on on a GPU they're deploy deployed on the same GPU right right then they get optimized for client what we're really seeing right now is we're in the in in the cloud wave uh when a human uh uh client a human actor interacts with a with a with a with a virtual agent in the cloud right that's the that's the wave we're going through right now tremendous growth we're all seeing it right you see what the the gpds of the world can do you see what the uh liquid foundation models can do right it's qu quite amazing right um the second wave is really starting it's coming together it's the personal AI getting onto the client side its laptops its tablets it's your phones so that wave is starting to uh take momentum but we think that the um the third wave when we get into the physical world and we get into humans interacting with physically powered AI agents and also physical AI agents interacting with other physical AI agents that we uh an equally big wave right so currently we're seeing that as balanced there's a trend to it's a cloud right now where we think the uh the uh client side as well as the edge embeded side are equally important um that's just so exciting I I can really imagine so many applications uh at uh on on the edge in addition to all the many applications in the cloud but um let me ask you this as the AI models grow in size um are there innovations that are needed in chips to cope with that yeah no certainly so we we're seeing the um the AI models growing in size really needs a lot of communication technology if you look at our high highend gpus there are about um just as much about the compute as it about the the the iOS the 30ies so we have uh uh more than 100 uh highspeed cies on a single GPU and we really Network them with other gpus so it's all about networking technology and that enables that scaling versus on the client side you know battery powered environments need to be extremely energy efficient and over there we are actually uh building a neural processing unit that is just focused on Energy Efficiency right and then you can take it all the way to the edge side we're sitting there with our programmable logic Fabric and programmable logic allows you to completely customize your newal network model you co-design with the hardware itself um and uh we're really bringing um very low Precision we're talking about binary chary neural networks onto onto Edge application s with extremely high forms of sparity right the human brain is somewhere between five to six orders of magnitude more energy efficient than what the cloud Hardware we're building that has to come from somewhere right so there's a lot more Innovation we can do together not on the liquid side but it's also on the MIT side on both sides on the science side and on the entrepreneurship side so we have time for only one question and I have to ask you about liquid um so what is the core uh reason you think uh liquid Technologies and products are attractive yeah so again I I I think as was stated many times today right it's we need to bring the efficiency up we've got we've got the scaling problems Transformers in the uh uh uh the fil phase or fundamentally memory BWI bound right driving that efficiency is is really important uh bringing you know uh scale making its scale differently is is really important and I just will enable so much more pervasiveness of AI not just in the cloud where we have the high bandwidth where we have the the the high-speed iOS to scale to many devices we have to make it efficient for bringing into everyday use of AI into the personal space and also into the physical world right and that's really where uh liquids focus on Energy Efficiency in the models is really important for us the interplay between AI body and AI brain is so important to AMD uh to liquid and to advancement in AI thank you so much for sharing your thoughts with us um let's give Ralph a a big round of applause thank you Dan thank you that was great thank you thank and now I'd like to invite back on stage our CEO uh liquid CEO Ramin Hasani Ramin help me in welcoming Ramin Hasani CEO liquid Ai and Steve Pagliuca co-owner of the world champion Boston Celtics and former chairman of Bane capital and a liquid AI [Music] investor okay this was um a surprise so thank you Steve for joining us congratulations on the B Boston Celtics it's great to be here thank you I'm going to keep it very short for all of you so let me just ask right into it why did you invest in liquid well I need a couple days really to talk about that but I I I'll keep it short you know I go way back and and starting out at the Bane and Company in 1982 and uh we were we advising tech companies at the time and I go so far back it's hard to believe my first client was McCormack and Dodge it was a company of 128 and they had invented packaged software for County and back in those days um you had to a company like GE would spend 25 to $50 million to build a proprietary accounting system system on a Mainframe and Frank Dodge had the idea of why don't we package that which will lower the cost dramatically and we could sell those packages at the time for two and a half or3 million and that was a venture startup company I was advising um so I got really interested in technology and the changes I've seen and the dramatic things that can happen I.E reducing the cost from 50 million to two and a half million through that through package software and then you fast forward uh we started a a technology fund at Bane capital and actually helped start the Gartner Group which was referenced today is now I I think we invested uh 6 million equity and now the company's worth 35 billion equity which I hope will happen with liquid AI as well um at some point in the future um but as i' I've looked at the whole landscape and all this Tech investing uh we we've seen major disruption you know one of my unfortunately also early clients was aund million dollar time sharing business so we went from mainframes you know to time sharing sh Ing and then this company bought the time sharing company for $100 million and and six months later one year later was worth nothing because we went back to distributed computing then the internet came in I would step back and say why why did we invest in in liquid AI I think liquid AI is a transformative technology transformational technology um it is going to have the same kind of impact in AI that we we saw from the internet on business and commerce and and and exploration and knowledge and if you if you look at what are the problems with AI the big problems are the enormous power cost uh I think close to 500 million in gpus alone to load chat GPT for um and that number is growing you know you know all the time secondly uh these models are hard to explain you know you know the hallucinations and people talked about it today and thirdly um you know liquid AI offers flexibility so so really the key factors are sustainability lot less power use um explainability and then flexibility as we talked about you can do you can do smaller models large models the whole nine yards and then you combine in the the ability to do time series which I think is really important in financial services and many of the industries we were very excited about that and I think it's going to revolutionize Ai and actually make AI accessible to many companies when you think about companies that want to be proprietary you don't want to put your all your information of the world on and and take some outside information put it on a public cloud or put it in a public uh uh piece of software this and you can't load a model no companies can't spend $500 million to load their own proprietary model but with liquid AI companies can have flexible models that you can load for the for the five to 10 million five million 10 million uh some of these applications uh that we talked about with with Arena um biotechnology pro protein analysis and folding uh incredible things that you can do that you really can't do if it cost you 500 you think about it if every company had to spend $500 million to load a proprietary model we'd have a very slow adoption of AI so what is the reason we invested we invested because it's a fantastic team out of out of MIT first of all congratulations thank you uh second because we think it's a revolutionary technology that'll make AI ubiquitous both on a vertical and a horizontal basis in vertical applications and uh hor horizontal large language models and it has the explainability the flexibility and and sustainability that people will be looking for so in a nutshell uh that's why we're very excited to to to back the company and and uh we're looking forward to you you know really seeing it take off I also think it's great for Massachusetts as the government was here today in in 2009 I wrote an editorial that we shouldn't lose the biotech sector and the tech sector like we lost it off 128 30 years ago and and then develop Patrick came in they committed 2 billion to biotech they've committed to to AI now and I think we can we can really want to protect that ecosystem in Boston so I think it has has social and Massachusetts implications as well so we're very excited to be a a big investor in liquid AI all right that's the perfect wrap for our day thank you so much Steve so and ramin why don't you uh put put on the championship ring for the final actor oh my god wow amazing amazing and are we going to have the team the team coming out here is the team coming out here yes yes so thank you very much everyone for joining us today this is wrap we are very excited we are early on in our journey but we're very excited to actually get you all access to the technology get in touch with us and before finishing I just want to mention that this is all a team effort so whatever you've seen today it's basically the magic of people of liquid so I want to ask the people of liquid whoever works at liquid you got come come on stage and please give a round of applause for the liquid AI team [Applause] I have one I have one final surprise I I have a uh a uh custom Lenny kravis Edition Dom Peron for the liquid team and they're going to they're going to store this and and they'll we'll drink it when they reach a 100 million in sales let's do it thank you so [Applause] much don't don't get used to it all right all right thanks everyone [Applause] [Music] [Music] e e e [Music]