Open-source innovation is the future of AI. In this episode of Smart Talks with IBM, Malcolm Gladwell and Tim Harford discuss the open-source AI community with Jeff Boudier, head of product and growth at Hugging Face. They chat about the history and future of open-source AI, its critical importance to AI progress, the IBM watsonx partnership with Hugging Face, and how businesses can leverage open-source AI for their specific needs.
Visit us at ibm.com/smarttalks
Learn more about the Hugging Face partnership: https://newsroom.ibm.com/2023-08-24-IBM-to-Participate-in-235M-Series-D-Funding-Round-of-Hugging-Face
This is a paid advertisement from IBM.
See omnystudio.com/listener for privacy information.
This is an unofficial transcript meant for reference. Accuracy is not guaranteed.
You're welcome to textile production,
From my heart radio
Today we are witness to one of those rare moments in history: the rise of innovative technology, with the potential to radically transformed business in society forever. That technology, of course, is artificially
elegance and it's the central focus for this new season of smart talks with ibm join hosts from your favorite pushkin bod guests, as they talk with industry experts and leaders to explore how businesses can integrate a I into their workflows and help drive real change in this new era of aid.
And, of course, host malcolm gladwell will be there to guide you through the season and throw in his to sense as well look out for new episodes of smart talks with ibm every other
big on the iheartradio apple podcast. Wherever you get your pod
tests and learn more at ibm, dot com, slash smart talks alone,
Welcome to smart talks with ibm, a podcast from Pushkin industries. I heart radio and idea I malcolm Gladwell,
season were continuing our conversation with new creators, visionaries, who are creatively applying technology in business to drive change, but with a focus on the transformative power of artificial
televisions and what it means to leverage ay. I said
changing multiplier for your business, our guest today is Jeff booty. I had a product and growth, but hugging face the leading, open source and open science. Artificial.
elegance platform, an engineer by background. He has a self professed obsession with the business of technology.
Recently ibm and hugging face announced a collaboration bringing together hugging faces repositories of open source, a models with ivy aims, watson ex platform
it's a move that gives businesses even more access to ai, while staying true to ibm, longstanding philosophy of supporting opensource technology with open.
Most businesses can build better ai models that suit their specific needs using their own proprietary data, while browsing a ready catalog of pre trained models. In today's episode, you'll hear why open source is so crucial to the advancement of ai. How I b m's watson, X,
interacts with open source, ai and Jeff's thoughts, and why the singular omnipotent a model is a myth. Jeff spoke with TIM, harvard hosted pushkin podcast caution.
tales along time, columnists at the financial times where he writes the undercover economist tim is also a bbc broadcaster, with his show more or less. Ok, let's get to the interview. I am a Jeff voodoo you product director at
face. So I'm immediately intrigue hugging faces this movie or something else
It is not, and it made me
You not obvious to a listener, but hugging face is the name of that cute emerge
you know the one that's smiling with his two hands extended like that together,
I that's hugging face. So basically we named the company after an emergency, and it is
I saw your website and it is a very friendly emoji. So that's that's nice. So tell us a little bit about hugging face and and about what you do that of course
I can face the leading open platform for a builders.
and its staff that place that's all of the air we searchers used to.
Sure their work.
There are new a models and collaborate around them. It's the place where
the data scientists, go and find those pre train models and access them and use them and work with them and increasingly it's the place where developer.
are coming to turn all of these a models and said that assets into their own applications down features. So it's it's like them. Another facebook group or the read it over the twitter for for people who were in,
is in particular, generative language, I'll all kinds of artificial intelligence, all kinds of a I really and, of course, generated
He eyes this new wave that has caught the world's by storm, but on a haggard face he can find any kind of model. The news for transformers model
To do anything from translation, or if you wanted to transport
what I'm saying into texts, then you would use a transfer
former model she wanted to then take that text and make a summary.
That would be another transformer model. Is she wanted to
it's a nice little thumbnail for these pot podcast by typing a sentence that would be another type of model
so all these models you can find is actually three hundred thousand stead our free and publicly accessible can find them on our website at hugging, phased out.
and use them using our open source, libraries, and so this is this is fascinating. So so there are three hundred thousand models. Now
when you say model, I'm thinking in my head. Oh it's kind of come like a computer program. There were three hundred thousand computer programs is that is that roughly right or not really,
the general idea a model is
giants set of numbers- that our
working together to sift through some inputs that you're going to give it so think of it? Have a big blog box filled with numbers, and you give it as a as in inputs may be some text
maybe a prompt, so you're asking you giving an instruction to the model, or maybe you give its an image as an input, and then it will sift through that information
thanks to all of these and numbers which we call in the field parameters and it will produce an output,
when I told you hey, weaken transcribed this conversation into text, the input would have been the conversation in an odious file and then the output would have been the text of the transcription. If you want to create a thumbnail for these podcast episode, then the input would be what we called the prompt
she's, really attacks description like a frenchman in San francisco, talking about a machine learning and the output would be completely original image. So that's how I think about
sweat and a model is- and I think what we are
starting to realise is that this is because
mean do a new way of building technology in the world if it has been for the field of
stealing understanding generating text for quite some time? But now it's sort of moving across every field of technology. We have models to create images.
I said, but also to generate new proteins to make predictions on numerical data. So every kind of
field of machine learning is now using this new type of models. But, what's interesting is that
if you're say a product manager at a tech company- and you say hey, I want to build a feature that does this a few years ago. The approach would have been to ask a software developer to write a thousand lines of
code in order to build a prototype and a new way of doing things today is to go look for an off the shelf. Pre train model that does a pretty good job at solving exactly that problem, so you can create a prototype of that feature fast. So it's a new approach of building a tech, I'm not a programmer, but I am aware that there was this idea of open source code, and now we have,
source models. So what does it mean for them to be open source open source? A I actually means a lot of different specific things. It's the open source implementation of the model. So if you use the
hanging face, transformers library to use them,
Oh you're, using an open source code library to use that model just to just went out and that their transformers. These are these kind of ways of turning our picture of a dog into the text out. But the says hey, this is a picture of a dog, or this is a french text and with the transform is helping, you turn it into english text. Always doing all of these things that you ve been describing. That's the the transformer is the kind of the engineer at heart
that, yes exactly and we call them transformers because they correspond to this new way of building machine learning models that was introduced by
Google, actually with a very important paper, called attack
is all you need, and that was pollution. Two thousand seventeen by researchers out of Google deep mind. Well ass, an that's the six years. That's it! So
It is very new and ever seen, stared at the piece of innovation, of like new model architectures has
you really accelerated, I, but you really started from this inflection point that came from the
paper and its implementation in what is now called transformer models transformer the has conquered every area of machine learning since ok, so so so it up so say you ve, got the livelihood of transport models and that open source, and that means that means what anyone can use them for free or the
oh, that anybody can implement them for free. What does it mean saw again? There's like last. I go into it, but the most important thing is for the.
the model itself to be available so that at that scientists or an engineer can download them and use them, and also
There are a lot of considerations about how you make them accessible and a very important one.
is whether or not you give access to the training.
all the information that went into training,
model and teaching it's to do. What, when its trained to do so tat, I might have fed millions of words into into a languid transformer. I might have fed millions of photographs into her into a pig's transformer here.
Yes, and now it's trillions in the end, the accessibility of their training data is very, very important. What's the delay,
shit between the the hugging face, libraries and github, which I find some github correctly. It's this the repository of open source code
lots and lots of lines of code and routines and programmes that are shared, updated and
back and all available on on github, which sound similar
what you're doing with hugging face for I saw it, but what is the interaction or the relationship that yeah,
you're. U nail on the head, the so hugging phase is too I would get hurt is to code right is the central platform,
where a builders can go, find and collaborate around a? I aren't you
facts, which are models and does it so it's quite different than software, but we played is central,
rule in the committee to share in and collaborate in an access all those artifacts for a I like it had offers for code among community must be incredibly important to me in the open source is nothing if you don't have a community of people working on it, so I will have you been able to too far.
Staring nurture that community. I think he goes to work through the air or regions of the tree
former model and and hugging face role into that. So when the first sort of open model came out, it was called dirt and he came out of Google. The only way you could access it was to use
a tool called tensor flow, but it happened that most of the a I community,
was using a different tool. Called pie. Torch in something that's hugging face did is to make that new model bert
accessible to all pay torch user and they did it in open source. I was a project called birds, pre trained pictorial, bird pight, what free trade,
so this is like being able to play my zelda game on an xbox or play station right, or am I
Why not really understanding? What's going on? No that's exactly what it is, and the thing is
Nobody was using the gameboy, and so it became very popular and from there the community's would have
gather to make all of the other models that were then published by ai researchers available through that library, which was quickly renamed from bert pre trained by Doj into transformers to welcome, like all of these different new models and to dame that
open source library transformers, is: what's all I builders are using
they want to access, does model see how they work and build upon them? What striking him at his feet,
is that its changing so fast its improving so quickly, so how to open source models keep up without how do they get iterated and improved? Actually is not so much that open source is keeping up with it. It's actually open source that is driving. This is driving this pace of change, and that's because with open source and open research data scientists, richard researchers can build upon each other's work. They can reproduce each others.
they can access each other's work, using our open source, libraries, etc. So in a sense, is not really that open source a eyes.
New idea, it's rather the opposite. There's been a blip of time in which
closed source a I seemed to be the dominance way
but it's really a blip. In fact, you know
None of the incredible advances that we're marvel about today would
possible without open source, we're standing upon his shoulders of fifty years of research and open source software. So I think that that's really important it if it wasn't for that.
who probably be fifty years away from having these amazing experience is like Chad, jpg or stable diffusion, etc. So is really open source that is fuelling this pace of change or all these new models. All these
You can't abilities to giving an example saw meat I released,
lemme large language model. Just a few months ago,
and ever since this been this cambrian explosion of variations in improvements upon the original models and today
over a thousands of them that we host and track and evaluate so yeah open source is really there.
the gas in the engine for four that just just made it clear that it is open source, not closed, had set the pace for I innovation. If that's,
true, then forward thinking, businesses shouldn't shy from leveraging opensource ay I to solve their own proprietary challenges, but how businesses can face serious obstacles when trying to adopt open source technologies like complying with government regulation or making sure their customers data stays protected in the next part of their conversation, Jeff and tim discuss. How ibm collaboration with hugging face empowers businesses to tap into the open source a community and how the watson ex platform can enable them to customize those ay. I models to their needs. Just wants to ask about the partnership between hugging face
and an I b m. But how did that come about? Well, he came through a conversation and a conversation between a our ceo clement along and a bill Higgins, I b m who's, really really
It goes to all the amazing research work in an open source work, that's happening at ibm and that conversation sort of sparked the
evidence that we needed to do something together. We share a lot of,
values in terms of the importance of open source, which is fundamental to us. With the importance of
doing things in an ethics first way to enable the community to incorporate ethical considerations in how there are building a I and we sort of how the difference audience to start
which is all the more I builders use hugging face today to access all the models, we talked
how to use them using our open source and build with them
and I b has this incredible history of working with enterprise companies and
an enabling them to make use of that technology in a way that compliant with everything that an prize requires
so being able to marry these two things together is an amazing opportunity and now
and enable the largest corporations that have sort of complex requirements in order to-
Deploy machine learning systems and ghip give them.
and these experience to take advantage of all the latest and greatest that air has to offer through our platform, is talk about
this idea of a single model or a variety of models, because
what have been hearing. You say even
other lots of models. There are hundreds of thousands of models.
available on hugging face, but you also said, there's a single thing, but transformer and they're all transformers. So so, if they will basically the same thing, why can't you just build one super clever model
do everything. That's that's a really interesting,
idea and very much a new idea. The reason
We have over a million repositories. Three hundred thousand free inaccessible models on a haggard face platform is that morals are typically trained to do one thing and their typically trained to do one thing with specific types of
and what's became new in evidence in the research that came out over the last couple years? Is that if you train a big enough model with enough
data, then those models start to have a sort of general capabilities. You can ask them to do different things. You can even train them to respond twins,
actions so with the same model, you can say hey summarised. Is paragraph translate this into english start a conversation in french and pivot. Germ
and so these are general sort of language capabilities, and I think when
Chad, jpg came online and the world should have discovered these new capabilities. There was, at least for a short period, this sort of idea, the sort of man
then the end game of all this is me.
be one or a handful of models
So much better than anything else than exists today.
And do anything that we can ask them to do, and that's the only model that we will need, and I, for one thing it is a mess. I don't
it is practical. For a variety of reasons, say your
writing an email, and you have like this grey suggestion of text to sort of complete your sentence. Well, that's a lie. That's that's a lodging which model that's a transformer model. That does that. So there are a ton of existing you
cases like this, and these use cases are powered by specific models that have been trained to do one thing well into do it fast.
If you wanted to apply these sort of all knowing powerful or a coal type of model,
are you would not be able to serve millions of customers through a search engine
would not be able to complete people sentences because
the amount of money that you would need the number of computers that you would need to run such a service? It does is just
deeds? What is available on on the planet? So why?
the reason for which it's not a prayer
The scenario is that is just very expensive to run those very
the large models. What I'm hearing is it's like look? If you want to screw in a a screw, you need a screwdriver, you don't want an entire tool shed full of tools. If the task is to screw in a screwdriver and short, you could bring the tool shed.
they're all the tools there's a screwdriver there, but it's not
sorry it's incredibly expensive, it's incredibly cumbersome and that cost exists, even though, maybe as the user, whose just typing it into a prompt box visa may not see it, but it still very real, that's right and then another one is performing. So
taking their screwdriver example so and by the way, like we're, not quite there them at this
moment where we have this all knowing powerful oracle. That is still sort of a safe. I scenario, but we have
screwdrivers, but we also have the leather men right there, the multitude swiss army knife and that sort of the moment that we are in today, but now, if I, if I'm trying to open up my computer, turns out that's it requires of specific kind of screw like these tiny little talk screws,
and having a talk. Screwdriver will give me much farther than trying to use my leather men where maybe I'll get the knife blade and ill. It will message disk
then maybe eventually I'll get to what I need. But my point is that if you take a very specifically trained model for a particular problem, it will work much better. It will give you better
results then a very, very generally stick big model that can do a lot of things and so forth
like search engines or things
translation for things that are very specific com.
These are much better off using smaller and more efficient models that produce better results. That's really interesting, and presumably then
being able to know which muddled use will be able to
Oh, who to ask which model to use becomes a very important capability? Yes, and that's what we try to make easy throughout through our platform said. Tell me about how this works. With I b m's watson ex platform. How do you see hugging faces customers benefiting from that d? The end goal is to make it really easy for watson x, customers to make use of all the great models
libraries that we talked about other they heard the three hundred thousand models are today on hugging face, and to do this, we need to really collaborate deeply with the eye
began teams that build the watson ex platform, so that
Our libraries are open source. Our models are well integrated into the platform
if you are a single users, you are that our science students and you want
He was a model. Is we make it super easy right? We have our open source library and download the model on your computer and run with it
but in enterprises there is a vast complexity of infrastructure.
And rules around where people can do and how the data can be accessed and all these complexity is sort of solved by the what's on x platform. This season of the smart talks. Podcast features what we call in new creators. Do you see yourself as being a creative
First. I think I think it's a requirement for the job I mean we're in such a new and rapidly evolving industry. That's that you have to be creative in order.
or to invent the bill.
Models to use cases of tomorrow. My role within the company is really to creates the business around all the great work of our science and open source
and product team and by and large, the business model of a lie within the whole ecosystem is still something that's. Companies are trying to add to figure out,
as the creativity is really important, to really have the conversation with com
Please understand whether trying to do in and build the right kind of solution. So that's like work. Creativity comes into play and one of the things that you ve you been,
talking about his justice. Growing number of models is growing number of capability.
This growing number of use cases enormously exciting, but also, I think, completely bewildering for most people,
you are trying to navigate their way through this
maze of possibilities, that is, that is growing fast of making even learn about it. So how are you helping people navigate and make choices?
not environmental, and how does the partnership with ibm help with that?
I said I'd. Our vision is that ai machine learning is becoming the default way of creating technology
I mean like every product up service that you're going to be using, is going to
using ai to do whatever it is better faster, and I guess there are two competing visions of the world. Coming from that there is this vision of the oracle, all
a pair of all the model that can do. Everything in our vision is different. Our vision
is that every single company will be able to create their own models that they own and that they can use that they control and that stir the division that were trying to bring to life through our open source tools that make these work easier through. Our
platform, where you can find all those between models are shared by the committee. So we really want to empower companies to build their own stuff, not to outsource. All the intelligence to a third party and de what's annex platform from ibm gives those tools to enterprise companies so that you can use d up and source models that hiding face offers. Them
you can improve them with your own data without sharing that data to a third party, and then you could do every all of this work
in compliance with whatever governance requirements. Are you have for your company, maybe
finance services, company and you have a specific set of rules. Maybe your healthcare company and you have very strong privacy requirement,
for patients data. Maybe your take
Benny and you have your your customer
your users, personal information, so you need to be able to do this work respecting all of that Jeffrey. I thank you very much. Thanks understands fun to create the air models of the future, we're going to need open source. That means as a place for business in the open source community to harness the game. Changing potential of a I innovation. Like just said, businesses face unique challenges. They need to solve at scale without proper support systems, tapping into open source ay. I add enterprise level is daunting. Finding the right size model for the job, fine tuning, its
this all, while addressing governance acquirements around data privacy and ethics, so for businesses, ibm collaboration with hugging face is a mark of progress because it signifies had business can tap into opensource a I, while preserving enterprise level integrity. Businesses should embrace the open source community,
a future much like hugging face and its emotion namesake suggests Malcolm global. This is a paid advertisements from ibm. Smart
talks with I b m is produced by Matt Romano David jaw, niche of n cat and roast in reserve with Jacob goldstein were edited by lydia jean
Our engineers are jason, gambro, cerebral air and then tell the day theme song, but grammar school.
special thanks to colleague, liore and kelly, Cathy Callaghan and the
far and ibm teams, as well as the Pushkin marketing team. Smart talks with ibm is a production of pushkin industries and ruby studio at I heart media to find more pushkin, podcast, listen on the eye, radio, radio, apple apple podcast or wherever you listen to podcast.
Transcript generated on 2023-12-12.