Making Large Language Models Accessible | Scale AI Fireside chat with Bill MacCartney

Key Takeaways

all right thank you everyone uh thank you for the latest scale ml Meetup and I think today actually we have maybe our best speaker yet so I'm very excited to have Bill here don't oversell it um just to start off a little bit on about Bill um bill is the head of um what is actually your your title is BP of engineering but you're actually kind of head of ML and AI for for code here that's right yeah yeah um and a little bit about cohero actually cohero is really one of the the most exciting startups that's working at llm space as a startup it's working to make large language models applicable to a wide variety of use cases especially Enterprise use cases so we're Fortune 500 and other corporations are using large language models for making workflows and other aspects of what they do much more capable than what they could have done even just a couple months ago um so we're going to have a dive today into cohero's use of large image models the work I'm making llms accessible and really making them easy to use for all kinds of developers and we'll talk about challenges and opportunities and scaling these models up to longer fine training runs honestly larger data sets data sets that are larger than what we were using even just a couple of months or years ago the safety implications of delivering those fine two double elements to all these various Enterprise clients that coher works with right now Bill leads a team of 40 plus scientists and engineers in building olms to help computers understand a wide variety of language use cases and he's also a Consulting professor of computer science at Stanford but previously he was also a director of proactive intelligence at Apple he's a manager director of BlackRock and a senior research scientist at Google so my first question to you Bill Vicki back in your career you've had a very interesting career journey and you know that list of names you've moved from Academia to finance to to Big Tech I'm curious you know given your your career Arc how why and where did you decide to join cooker yeah well first thank you for having me here uh it's exciting to be here it's great to see an awesome turnout here and I'm really excited to talk about llms uh I've been a co-here for about a year and uh there were a few different factors that that attracted me to cohere the the first was the focus and Agility that comes with being at a startup before I joined cohere I had been at big companies for more than 10 years I spent five years at Google as a research scientist I spent a year at BlackRock doing quantitative Finance I spent five years at Apple building and leading a group called proactive intelligence and so you know a decade at big companies you were at Apple with me so I know you can identify with this and you were at Google as well there's there's a lot of wonderful things about big companies but there can also be a lot of frustrations at big companies and I had been involved in a couple of startups earlier in my career and I really missed the ability to move fast and the ability to change direction quickly so that was definitely part of the attraction to cohere a lot of startups can offer that and so one of the one of the factors that was maybe a little bit more specific to cohere was the core technology the technology that cohere has built around is the Transformer based neural network architecture and our founder and CEO Aiden Gomez was part of the team at Google that invented the Transformer he was a co-author of the famous paper attention is all you need and um it's no exaggeration to say that the Transformer uh is the biggest thing that's happened in Ai and machine learning in the last 10 years my field is NLP and every aspect of NLP every corner of NLP is completely different now than it was even five years ago and that's mainly because of the Transformer and now it's um transforming sorry not not only NLP but also computer vision and Robotics and protein folding and quantitative finance and and you name it so the the opportunity to play a small role in advancing this technology and in in serving as a bridge between the frontiers of technology and problems that matter to people every day was irresistible um the last thing and the thing that really sold me on it was just the people I before I joined I had the opportunity to spend some time with the founders and other members of the team and I was just super super impressed um not only with their intelligence and their accomplishments which kind of goes without saying but above all with their um warmth and generosity and shared values and um humility and and uh empathy and I was just like this is the kind of people that I want to be around this is the kind of people I want to build with yeah that's great always super important when you join New York um you are maybe underselling the impact of this paper attention is all you need I would say it's maybe one of the most you know kind of groundbreaking insights of the last couple years in this field can you maybe give that elevator pitch for the paper and the key insights that that you cleaned from it yeah I mean the paper unlocked this new approach to building large language well to building all kinds of models but I'm going to focus on large language models and um large language models are [Music] incredibly exciting for a couple of reasons the the first thing for me is just um their amazing ability to capture the Nuance of meaning and to um capture the impact of context on meaning in a way that previous approaches we just couldn't couldn't remotely do and I'll give a concrete example of this um so imagine you have a movie review that says something like uh the plot was thin but the star-studded cast delivered a scintillating performance and the visuals were spectacular and um then there's a question which is what did the reviewer think about the story so this is the task of faceted sentiment analysis you're not asking about a sentiment overall but you're asking about sentiment toward a particular aspect of a particular facet of the product and it's a really hard task because uh well for several reasons one is that the overall sentiment can easily drown out the sentiment toward a particular facet because you've got all these positive adjectives like star-studded and scintillating and spectacular and it that's it it's easy for that to drown out the plot was thin um then a second challenge is uh the question was about the story but the reviewer said the plot is thin so you have to connect story to plot that part is not maybe not that hard the third challenge which is the hardest part of all I think is that the meaning of sin or the sentiment expressed by Finn is really dependent on context a thin plot is not good but Finn isn't always bad if you're talking about a phone or a laptop or something thin is good so this is a genuinely really hard task and I know this firsthand because many years ago 15 years ago when I was a grad student I did a Consulting project which was focused on faceted sentiment analysis and at that time we used kind of classic NLP techniques we used a bag of words and logistic regression and hand engineered feature representations that capture things about the syntactic parts of the sentence and stuff like that it was really hard and the results were extremely mediocre and so one of the things that I'm astonished by with large language models is how almost effortlessly llms can solve this problem and with really high accuracy capture all of that nuance and all of that kind of context dependence um what's even more impressive about llms for me is that the very same model without any task specific training can not only do faceted sentiment analysis but can also summarize news articles and uh extract movie titles from Reddit posts and generate marketing copy and identify toxic tweets and correct transcription errors in transcribed speech it's really amazing how one model can solve all these problems and even solve problems that you didn't even have in mind when you train the model things like I don't know um uh decoding uh internet acronyms or uh frivolous examples like translating ordinary speech into pirate speech um just kind of magically comes out of of these models um when I first started learning about NLP almost 20 years ago the whole field was balkanized into these separate sub-communities so you had the the syntactic parsing people and the named entity recognition people and the um the the sentiment analysis people and the Machine translation people and each of these subfields and sub communities uh made different models they used different data sets they had different approaches to evaluation different model architectures different feature representations and today you're seeing these great conversions where all those walls kind of fall down and um and um one one model can can solve all these tasks and the same convergence is actually happening at an even higher level across modalities with models that can handle natural language but also images and video and so on so it's just an incredibly exciting time to be working in Ai and machine learning yeah yeah Kevin that you worked in NLP you know before this current Revolution by which I mean you've been in the field longer than two years or three years um and and the variety of techniques that you brought to bear back then backwards and logistic regression do you feel that a researcher starting now do you understand exactly which techniques that they would go to start with and how they would get started in this field and is it totally different from what you did back when you were a graduate student and you were first getting interested yeah it's pretty different um I mean uh when I was a grad student it was important to know a lot about Linguistics to do NLP um that's one one big difference and these days it's much less important to understand you know to be able to read a parse tree and know what the different syntactic categories are and and stuff like that so I think that kind of prerequisite has completely melted away um and I think that's in many ways to the benefit of the field it's enabled a kind of cross-pollination between very different disciplines and machine learning you see lots of people who were trained in Vision now like coming in and doing great NLP work and vice versa people moving in the opposite direction yeah so I think broadly that's been really healthy for the field to have this kind of convergence of different different fields of machine learning that's great one interesting thing about your background is you've also worked at some of these big tech companies Apple's one of them and one interesting thing about the field of ml is there's a perception that you need a lot of really expensive investment in infrastructure in order to to work in this field and to train new models I'm curious how you think about the role that infra plays in this field you know especially working at a startup where maybe you have different thoughts on where you can most sufficiently spend your resources and your energy and infrastructure yeah um building these models is indeed very expensive and requires lots lots of resources um and um it's it's uh it's both data and compute um the best large language models have tens of billions or hundreds of billions of parameters they're typically trained on something like a trillion tokens of carefully cleaned and curated text so they're they are indeed very very expensive to to train [Music] um the this is um it's coupled with one of the really fun things about working at cohere which is that we have an honest to God AI supercomputer to work with um and for this we've partnered closely with with Google we uh about a year ago we announced uh a multi-year partnership with Google Cloud to provide compute infrastructure and to help us bring our products to Market and Google has been a fantastic partner for us um one aspect of that is that uh we are now the single largest user of tpus outside of Google itself and on a typical day we're using something like one and a half to two exaflops of compute so I mean that's just a staggering amount and exaflop is a million million million multiplications per second so it just boggles the mind and for a computer scientist to be able to work with this machine it's like um it's like a kid who builds model rockets getting the chance to launch the Saturn V I mean it's just incredibly fun um it's also a lot of energy and one of the things that we really appreciate about working with Google is their commitment to renewable energy all of the energy that Google uses to power our AI supercomputer is produced from renewable sources and mostly wind power and we we really appreciate that it's um you know aligned strongly with our with our values um also on infra on in terms of software we use Jacks we use xla we use pjit this is another area where we've had a great collaboration with Google we have a close relationship with the Jax team and um have gotten lots of great support great advice from them my My Hope Is that we've reciprocated by helping them to develop the the Jax ecosystem by providing uh you know kind of constructive product pressure back into them uh and then data is another really important uh input to this process so it's becoming ever more clear that um the both the accuracy and the safety of large language models are driven more than anything else by data by both the quantity and the quality of data that's used for pre-training so we train on more than a million tokens of data we collect huge amounts of data from the web and from from other sources we invest a lot in ensuring that the quality of that data and so that means things like very careful HTML parsing smart deduping of documents model driven filtering and and so on and we we measure carefully the impact that data improvements have on model quality and model safety yum that's great I also I used to work at YouTube at Google and I'd have to request resources through Borge I think if I had tried to request an axle plot with resources I probably would have been fired on the spot but so that's pretty amazing you're able to do it in your partnership um can you talk a little bit about the difference between the work that goes into training some of these underlying models you know generating a new model from scratch versus the work that might go into What's called fine-tuning the models making the models applicable to an Enterprise use case for a specific use case which of those two areas do you spend time on and what what drives your decision to focus on fine tuning as opposed to training your own model it really depends on the the customer the use case the application our main focus and most of our attention and most of our resources goes into training the base models that's that's the really expensive work so we have two different families of models we have a family of generative models which um are you know similar to gpt3 they take a piece of text as a prompt they give you back more text we also have a family of embedding models or representation models that take a piece of text and give you back a vector representation of that text then we build on top of those models to provide API API endpoints that adapt those capabilities to specific use cases so things like text classification or summarization and so on and then for customers who have very specific tasks and who have data sets that are kind of Representative of that task we offer the ability to fine-tune the base model to perform that that specific task now you know whether that's appropriate for a specific customer really depends on what application they have in mind on whether they have data a lot of our customers kind of don't really have large amounts of of data they have just a few examples of data where it makes much more sense to use a few shot Paradise few shot prompting to to get the the result that they want but for customers that do have large labeled data sets um fine-tuning Can yield much better performance on whatever that whatever their task is in computational terms it's comparatively inexpensive and we're looking for way but constantly looking for ways to make that ever cheaper so we're investing a lot in techniques for parameter efficient fine-tuning where you can fine tune us a quite small proportion of the parameters in the overall model and still get very good results we a lot of our engineering effort goes into ensuring kind of reliability and speed yeah a fine tuning uh just to make that capability much more usable for our customers that's great and t-shot for you means does it mean a couple million examples a couple thousand two examples for fine-tuning it would typically be uh upwards of I mean ideally at least thousands some customers come to us with millions of examples and that's that's great um by fuchsia I actually meant something different which is where um if a customer only has you know maybe five examples of the task that they want to solve there's an approach to taking those those five examples and actually incorporating it incorporating them into the prompt that you provide to a generative model and basically saying do do more like this and another another one of the Miracles of of large language models is that that actually works without without doing any task specific training you can provide just a few examples of the task in the prompt and um get get get uh you know get get it to solve more examples like that yeah yeah it that seems like especially over the last few months to be this incredibly powerful technique that everyone's just wrapping their heads around yeah it's like black magic yeah how do you go about thinking about that you know something that right now seems almost more like an art than a science to understanding what's capable with prop construction um how do you think about how you're gonna systematically approach this space or do you even systematically approach how you think about building problems um we haven't made a great investment in um trying to systematize or automate prompt engineering but I think it's a really interesting area there's there's there is some academic research that has that flavor um I it's not something that we've made a great investment in I think a lot of our customers do come to us with enough examples uh that fine-tuning is the best past that best path forward for them yeah um we have lots of you know um we have lots of experience of kind of playing with prompt engineering ourselves to try to elicit the the best performance you know I mentioned earlier the example of um trying to uh convert ordinary speech into pirate speech and this is something that works astonishingly well just using a few shop prompt engineering you can like give it five examples of rephrasing something in Pirate speech and then just ask it to do more and it just kind of works it's amazingly well right and it's really fun to play with things like that yeah my favorite example I saw around the office recently was a pretty complicated classification task I think it was like Insider training or a class like that yeah and you can you take the prompt and it does a pretty good job to start you change your prompt just slightly by saying I am a smart computer at the beginning of the prompt and suddenly there's like a marginal interest in the performance but enough that's significant yeah there's there's just crazy techniques that work like that that I think we're just uncovering yeah how do you think about for these Enterprise use cases you know a pirate speech is obviously a case where maybe there's some sensitivity around it but but not so much as some of the other Enterprise use cases you think um and you've talked about doing metrics driven development measuring your performance and improving over time what's the right way to think about adapting these large long Hood models to Enterprise use cases where there is objectively a criteria that you need to to hit and you need to make sure that you're hitting certain Precision or recall or other metrics that are really important to the customer yeah this kind of brings in the the larger question of evaluation like basically how do you measure how do you know that your model is good and um and this is particularly important when we're talking about uh a specific customer specific application is it good for this customer um broadly we have kind of I mean evaluation is something that we're kind of obsessive about um and I I think that's just part of good machine learning methodology is start by defining success quantifying it giving you know creating a way to measure it and then relentlessly Hill Climb against that metric broadly we have three different Avenues to evaluation and I think of them all as kind of complementary uh the First Avenue is essentially academic benchmarks so we evaluate our models on a wide variety of academic benchmarks including some benchmarks that are themselves collections of very diverse tasks like big bench light for example and academic benchmarks are great in in the sense that you can run them automatically and frequently during development and so they're an easy way to evaluate a new approach to tune hyper parameters to compare to llms from other groups but many of them are a bit artificial and good performance on a benchmark doesn't necessarily correlate with good performance on a task that a customer cares about so the Second Avenue to evaluation is evaluation on customer specific tasks and as I mentioned many of our customers come to us with labeled data sets that kind of capture what they're looking for we can eval we can use those not only for fine tuning but also for evaluation and this is great because then we know that we're measuring what the customer actually cares about what really matters for for their use case but of course those those data sets are not public we can't report those results um and then the the Third Avenue to evaluation is human evaluation so here we partner with folks like scale on um getting human assessments of the quality of our model outputs so regenerative model we'll ask humans to assess attributes of the generative output like fluency and coherence and relevance to the prompt and things like that and this is great because it kind of represents the gold standard of whether the output is is really good um but um compared to automatic benchmarks human evaluations are they're comparatively slow and expensive so we don't use them kind of routinely during development instead we tend to use them uh to validate model release candidates when we think we've got something good can you give a good example of that and you know it is interesting because a lot of benchmarks that we use against machine learning models are just subjective dense parts that you can represent through you know a video game and how you're actually performing as you're playing a video game and this is maybe an example where you do need human beings in order to do some evaluation it's a good example of something where you'd use human judgment in order to inform your release candidate yeah I think uh one of the areas where it's been important for us is is long-form Generations um a lot of the academic benchmarks that are kind of aimed at generative models are um not good at evaluating the quality of long Generations they're based on they use data sets where the the generative output is relatively short one of the problems that we saw at one point during the development of our models was a tendency to wander off topic after during a long generation we've also seen tendency to become repetitive after long generations and existing academic benchmarks weren't really helping us to to identify those problems or to um you know to to evaluate strategies for for reducing that and so um it's it's really important I think not to rely too heavily on any one Benchmark or even any one of the these three Avenues to evaluation but to kind of see them them all as as complementary broadly they correlate with each other but the places where they diverge can be really instructive yeah that's very interesting one area that sometimes gets a lot of attention in the press that is is you know just an interesting area to explore and to try to understand is areas where the llm May produce an output that has some sort of societally bad or really negative outcome you know examples that folks have given have been hate speech or misinformation or other ways where the llm can be induced to produce a result that you really don't want how do you think about those scenarios that cohere and how do you start to address them before llm is put into an Enterprise use case or a mission critical use case where that outcome could potentially be really really negative for a customer yeah I'm really glad you asked about this because responsible use of AI is something that really matters to me and uh it really matters to cohere um in fact I said at the top that one of the things that attracted me to cohere was alignment of values and this is this is one of the things I was thinking about I think at cohere we're keenly aware that this large language model technology is extremely powerful and um you know as as Uncle Ben taught us with great power comes great responsibility and so responsible use is a core value for cohere for our Founders for everyone we've hired from the beginning and um something we we strive to excel at um this past summer uh we partnered with some of the uh leading llm providers we partnered with openai and AI 21 labs to Define and publish a set of print principles for responsible llm development so we articulated principles in three broad areas they were prohibiting misuse mitigating unintentional Harms and thoughtfully collaborating with stakeholders and I'm really proud of cohere and our partners for taking a strong stand on this and for not dismissing concerns about responsible use of course it's one thing to articulate principles what's more important is following through by implementing specific measures to to mitigate harm yeah yeah that's interesting and do you evolve along with customers as they're thinking about this space and they find their new stakeholders with an organization that need to be brought to bear to these sort of problems yeah uh this is this is something that varies by customers I mean uh some of our customers don't really care we we still care but some of our customers are not really focused on this issue for other customers it's really important so you can imagine um uh for example um if you're if you're working with a retail a retailer that wants to provide customer support through a chatbot it's not okay for that chatbot to start spewing noxious stuff so they care passionately because they want that bot to speak in their brand voice and they want their brand voice to align well with with their company values and and their intention in creating it so this is one of the one of the areas that we think most about how do we ensure that the output of our of our generative models is um is not toxic isn't does not have you know profane or obscene content does not have hate speech does not have bias and things like that um there are multiple places in the in the pipeline where we can intervene to to try to mitigate those problems one is right at the beginning in the data that we use for pre-training the model we can try to filter toxicity out of the pre-training data so we're voracious consumers of data we you know ingest most of the internet and Reddit and all kinds of other stuff and you can imagine if we did that carelessly we would be hoovering up all kinds of noxious stuff what we can do is use both deterministic and model driven filtering methods to try to reduce the prevalence of that toxic stuff in the pre-training data and that makes it much less likely that we'll wind up parroting that stuff back in our generative output uh a second remedy is we can um fine-tune our base models to promote outputs which are better aligned with human values and and human intentions uh a third thing we can do is prompt filtering so this basically means trying to identify prompts which seem to be like specifically designed to provoke the model into into producing toxic outputs to to go to the model into a into a into some kind of response and we can simply refuse to play along um and then we can do output filtering so we can use a toxicity classifier to um look for potentially toxic outputs and fall back instead to innocuous outputs even if they're less probable according to the model so it's kind of a I think you kind of need defense in depth it's a complex problem that none of these none of these measures are perfect um but by having defense at multiple points along the way you can greatly diminish the the likelihood that that something bad is going to happen yeah that's great yeah that sounds like a wide range of techniques that you learn in a particular situation which one is most applicable to yeah finding this content it's great I'll ask one last question before we open up to the floor for questions from the audience um and this is maybe that the Longview question although the space this space is like evolving so quickly and we're talking about capabilities and problems we wouldn't have talked about six months ago so it feels awkward to even ask about what's gonna happen five years from now but you know thinking about the next five years and thinking about how large language models are evolving what do you think the Field's gonna look like and what gets you most excited about working in this field yeah I'll mention a few things maybe I'll um maybe I'll talk about a couple of things that I I think are a little bit closer in and and then maybe some some uh more open-ended long-term things um so there's a couple of opportunities that I think are pretty exciting over the next couple of years one is uh neural search um I mentioned that one of the amazing things about llms is their ability to capture nuance and meaning and and meaning in context and one place where this can really help is search because search is about understanding the meaning of documents and understanding the meaning of search queries and connecting the two together and in fact within Google search was one of the places where the Transformer and Transformer based llms first had big impact um I don't Envision us going head to head with Google in web search but I do think that Enterprise search represents a big opportunity big companies have millions of documents internally that often capture a huge a huge proportion of the institutional knowledge of the company and that they want to make available to their people and today Enterprise search by and large sucks it tends to use you know very simplistic keyword matching techniques that don't capture the Nuance of meaning at all um so we think this is a big opportunity uh over the next couple of years another one is um conversation dialogue chat um personally I don't get excited about social chat Bots which I think are mainly entertainment but I do get excited about dialogue agents that help you find information and help you get things done we've been very inspired by Google's Lambda which learns to incorporate search into dialogue in a very flexible and Powerful way so Lambda understands when uh moving the conversation forward requires retrieving some information from the web it understands how to formulate a search query that will help and it understands how to incorporate the information retrieved into the conversation in a really fluid way so that's a great first step I think we can go much further and I think the big opportunity is to build dialogue agents that can in the in the course of the conversation learn to invoke arbitrary apis not only to retrieve information but also to take action um so uh one example of this might be like a flight booking scenario where you need to retrieve information about available flights but then you also need to take action to book the flight or to choose a seat or to record your frequent flyer number or something like that or you can um you can imagine managing your calendar or choosing a health insurance policy or um building a 3D model of a new car or something like that I think language enabled this kind of language enabled complex interaction or tool use represents a really big opportunity for the coming years I'll mention one more thing which has maybe this is a little bit more wacky but um one one thing that gets me really excited about large language models is what you might call um AI augmented creativity um I think there's a there's a scene in one of the um one of the Iron Man movies where Tony Stark is talking to his his virtual assistant um Jarvis Jarvis thank you and he's like he's like let's all right let's game this out and they start talking about scenarios for I don't know something or other and that's what I'm talking about I'm talking about like having an AI help you come up with ideas brainstorm ideas a large language model can actually help you come up with ideas for product names or for um recipes or even for movie scripts and here's a Whimsical example a guy on my team created a demo that was a date idea generator so like you give it a a location and a season and it gives you back a suggestion for a date activity so he put in Toronto and summer and it suggested roll down the street together in a giant bubble which I mean I think that's an awesome date activity and he did it again and it said um get a collection of really bad jokes take turns telling them to each other and laugh like crazy people and like I I mean it's a frivolous example but it those those date ideas are way better than anything I would have come up with those are some killer data yeah and I I just think it points to a really exciting potential yeah that's interesting you can also Imagine you know I I love the idea of like AI is like an augmentation of ourselves but it's almost like bringing out the best version of ourselves yeah in a way that's very hard for us to do something it's great to think of that yeah well great I think we'll open it up to the audience that folks have questions we have a whole bunch of questions so the tough parts to figure out who goes first his name a company is called the Pegasus AI and initially we have a product that delivers highlight reels faster than any human is capable of editing to athletes so more like computer vision yeah yeah but um my question was primarily um what would you say if you have to estimate the ratio of expense for training these large numbers models depending these massive amounts of data you're suggesting to um how expensive it is to host these cloud apis per call yeah so so basically the the I think the question is about the relative expense of training versus serving yeah the models um they're and they're they're both really expensive I mean I talked a little bit about the expensive training serving is also really expensive and and not to be underestimated um part of that is just that these models are enormous that like I said the best models are tens of billions or hundreds of billions of parameters um huge memory footprint the biggest models won't even fit on an 80 gigabyte a100 that means that serving you have to figure out ways to distribute inference over multiple gpus in an efficient way that's not easy um and um inference winds up being really expensive um so we work really hard to to optimize those those caught to bring those those costs down um I don't have a number for you on like what the relative ratio is I I can say that we we spend a lot more on training than on inference but we spend a lot on inference as well um I I don't have a precise number for what the for what the ratio is Yeah you mentioned your partnership on with Google and Jax which is another really important framework for trying to see things up but that's right yeah hi a great discussion by the way um so hi I'm Alan I'm from trailers consulting company that helps more traditional companies navigate their AI Journey from the first steps to putting models in production so this is over relevant um so aware scale so I have to ask our scale we we thought a couple of years ago that escaping the models and the number of parameters was like the most important thing that you could do it was kind of the the way that open AI was showing everybody but recently within Gmail paper we kind of learned that scaling data is much more important than scaling model size so there has been like a kind of parallel shift and some people are kind of fearing are we going to run out of training data to train these large models in the near future and how's it going to look like when probably most of the data generated on the Internet is going to be either generated by these models or highly affected by these models let's say so what's what do you think is going to happen the chinchilla paper is a paper that looked at whether data and the size of these data sets was a bottleneck or whether the number of parameters of the bottleneck it had a provocative question post of whether we're going to run out of data on the entire internet in order to effectively train these models yeah I I think this is a great question um I have thoughts about some some parts of that um the last part of it is a real conundrum but let me talk about scaling a little bit over the last year maybe two years since gpt3 was released um there's been a few different Avenues of research in the research community that have all kind of had this common theme of we got results just as good as gpt3 or even better with a far smaller model by doing X Y or Z different and so chinchilla was one it's like just trained for much longer basically like you didn't find the right uh you didn't you didn't find the compute optimal trade-off between training and um and data um or between sorry parameter size and and data that was a really important insight and one that we've Incorporated um uh another one I think with us with a similar flavor was instruct TPT which is look we can get as good even better performance with it kind of a different training Paradigm using different training data and um and different kind of different kind of supervision another great Insight I think of retro as representing this so retro is a paper from deepmind which is basically about moving a lot of the the burden of knowledge capture outside of the language model itself and making it an external research resource that the model can learn to consult as you go and I think of it at a very high level as having a similar kind of flavor of we can do more with less um these are all amazing insights for the research community my take on scaling is you can incorporate all of these insights and it remains true though that adding more scale is still going to help as far as we can tell we have not yet even after incorporating all of these insights we have not yet seen the the you know the the trade-off between scale and performance flatten out it looks like there are still gains to be had by building larger and larger models the question I think becomes is that actually the best thing for the applications that you're trying to build a lot of the scaling lot papers that you see and the the graphs that you see the x-axis is on a log scale and so you know I'm going to oversimplify a little bit but basically yup you can eek out more gains if you increase the model size the number of parameters by another factor of two or five or ten you can get another point in accuracy on on some Metric but like does that actually matter that much for whatever application is and does it matter like we're very commercially focused so we're thinking about our customers does that matter for our customers does the customer really care whether we get like 81 or 82 on hella swag not really probably what they often do care very much about is latency depending on the application it matters much more for some applications than others but many of them care very much about latency and they would be very happy to sacrifice one point of accuracy to get a 10x speed up in latency so those are the kind of questions that I think become really important when when thinking about scaling and depending on the specific application the specific use case building the biggest possible and and best in the sense of accuracy possible model that you can may not may not be the right thing the last thing you mentioned I don't have an answer for and I think it's really fascinating what do we do when all the data on the Internet is generated by machines and we just we get into this like uroburos where we're pre-training on data that was produced by another large language model um I don't know yeah yeah thanks a lot for the talk er um my question is how you think about access to these models evolving um specifically you know like open AI at the start of the idea that these walls but then they also discovered that as you scale of those models it's a very International speed training and so you have to kind of tap into capitals of uh so excuse you know like especially as we talk about scale how do you think about access to these models uh I think this is a great question and I'm going to talk about it from two different perspectives one is um what's efficient and one is what's safe so let me talk first about the efficiency um these models are incredibly expensive to to train and um it's the computational cost which is you know it's often millions of dollars to to train one of these it's the it's the data um that you need to collect and curate it's also I mean it model development depends on highly specialized expertise which is locked up inside a small number of brains um it's not practical for most of the organizations that could benefit from large language models and most of people who could benefit from large language models to train their own so um in a way kind of the big idea for cohere is to centralize those costs and share the benefit widely that's in a way our kind of the central thesis of cohere is that those costs should be centralized and then the benefits should be shared widely and we see an analogy here to electric power generation it's not practical or efficient for every Factory and every supermarket and every dry cleaner to generate their own power because of the very high Capital costs so as a society what we do is we centralize those those uh the the capital cost of power production and then we distribute the benefits widely through the grid and modern AI with these giant models we think is a lot like that we see it as our role to amortize that very high cost of developments development over a huge number of of of beneficiaries of the amazing power that it that it unlocks so that's kind of the efficiency aspect of it the other angle that you might have in mind is the the safety aspect of it I mean uh dpt2 initially was um not released because of concerns about danger and about the potential for for harm for misuse it eventually was and then tpt3 was released as well um well it's available it's not open source it's available for use but it's not open source but we do have lots of other models now that are that are available in open source you can download all the model parameters and if you've got the hardware you can fire them up and do what you want with them um this is a really tricky question and I I don't think I have um an authoritative answer to what's the right thing for society um on the one hand as I mentioned responsible use is a core value for coher and so we think it's important that we maintain a role in um trying to understand how our models are being used and make sure that they're being used for good and not for harm in society and it's hard to do that if models are open sourced and you you completely lose control over how they're used on the other hand there is an argument that like essentially the horses out of the barn and um it's uh it's less you know it's too late let the chips fall where they may hopefully it all works out um maybe I I am uncomfortable with that um I feel like that's a Cavalier attitude towards responsible use of AI and I get worried about all the things that could go wrong um but maybe it'll work out all right um and there is a point that there are many technologies that can be used for good and for ill that the telephone can be used to do horrible things and yet we all have telephones so um maybe there's something to that argument I've never been able to convince myself to to be really reassured by it great thank you Bill it's refreshing to hear a really nuanced thoughtful considered takeout obviously a question that's going to be really pressing to all of us over the next couple years yeah appreciate that with that thank you so much Bill thanks for joining us it's been great to hear about your workout go here and my pleasure this was a lot of fun and we will be following the developments very closely