Ideas: Accelerating Foundation Models Research: AI for all

not even really thinking about, right?  So that's how we started with AFMR.

friction, that have errors, that have failures.  To say that maybe the easiest path is not the   most advantageous, but the one that you can  feel the most fulfillment or agency towards.

called Accelerating Foundation Models Research,  or AFMR for short. With me is Cesar Torres,   an assistant professor of computer science at the  University of Texas, Arlington, and the director   of a program called The Hybrid Atelier. More on  that soon. I'm also joined by Muhammed Idris, an   assistant professor of medicine at the Morehouse  School of Medicine. And finally, I welcome Evelyne   Viegas, a technical advisor at Microsoft Research.  Cesar, Muhammed, Evelyne, welcome to Ideas!

the accountants and he's like look at the amazing  work that these folks are doing. But the reality   is that a lot of what they do is hidden behind  spreadsheets and so it wasn't necessarily the   most engaging. Suffice to say I did not go into  accounting like my dad! [LAUGHTER] But then he   showed us the chemical engineer in the factory,  and he would tell me this chemical engineer holds   the secret formula to the most important  processes in the entire company. But again,  

it was this black box, right? And I got a little  bit closer when I looked at this process engineer   who was melting metal and pulling it out of  a furnace making solder and I thought wow,   that's super engaging but at the same time it's  like it was hidden behind machinery and heat   and it was just unattainable. And so finally I  saw my future career and it was a factory line   worker who was opening boxes. And the way that she  opened boxes was incredible. Every movement, every  

like shift of weight was so perfectly coordinated.  And I thought, here is the peak of human   ability. [LAUGHTER] This was a person who had just  like found a way to leverage her surroundings,   to leverage her body, the material she was working  with. And I thought, this is what I want to study.   I want to study how people acquire skills. And  I realized… that moment, I realized just how  

important the environment and visibility was to  being able to acquire skills. And so from that   moment, everything that I've done to this point  has been trying to develop technologies that could   get everybody to develop a skill in the same way  that I saw that factory line worker that day.

And I fully intended on going back. I was going  to basically be working in spreadsheets for the   rest of my life. But actually during my postdoc  at the time I was living in Montreal, I actually   had distant relatives of mine who were coming to  Montreal to apply for asylum and it was actually   in helping them navigate the process, that it  became clear to me, you know, the role, it was  

very obvious to me, the role that technology can  play in helping people help themselves. And kind   of the big idea that I realized is that, you know,  oftentimes, you know, the world kind of provides   a set of conditions, right, that strip away our  rights and our dignity and our ability to really  

fend for ourselves. But it was so amazing to see,  you know, 10-, 12-year-old kids who, just because   they had a phone, were able to help their families  navigate what shelter to go to, how to apply for   school, and more importantly, how do they actually  start the rest of their lives? And so actually at   the time, I, you know, got together a few friends,  and, you know, we started to think about, well,  

you know, all of this information is really  sitting on a bulletin board somewhere. How can   we digitize it? And so we put together a pretty,  I would say, bad-ass team, interdisciplinary team,   included developers and refugees, and we built  a prototype over a weekend. And essentially what  

happened was we built this really cool  platform called Atar. And in many ways,   I would say that it was the first real solution  that leveraged a lot of the natural language   processing capabilities that everyone is using  today to actually help people help themselves.   And it did that in three really important ways.  The first way is that people could essentially ask   what they needed help with in natural language.  And so we had some algorithms developed that  

would allow us to identify somebody's intent.  Taking that information then, we had a set of   models that would then ask you a set of questions  to understand your circumstances and determine   your eligibility for resources. And then from  that, we'd create a customized checklist for them   with everything that they needed to know, where to  go, what to bring, and who to talk to in order to   accomplish that thing. And it was amazing to see  how that very simple prototype that we developed  

over a weekend really became a lifeline for a lot  of people. And so that's really, I think, what   motivated my work in terms of trying to combine  data science, emerging technologies like AI and   machine learning, with the sort of community-based  research that I think is important for us to truly   identify applications where, in my world right  now, it's really studying health disparities.

the goal is really to help nurture a healthy  pipeline for potential big bets. What I do   is really about “subtle science and exact art”  and we discover as we do and it involves a lot   of discussions and conversations working with  our researchers here, our scientists, but of   course with the external research community. And  how I got here… well first I will say that I am   so excited to be alive in a moment where AI has  made it to industry because I've looked and worked  

in AI for as long as I can remember with very  different approaches. And actually as important,   importantly for me is really natural languages  which have enabled this big evolution. People   sometimes also talk about revolution in AI, via  the language models. Because when I started,   so I was very fortunate growing up in an  environment where my family, my extended family   spoke different languages, but then it was  interesting to see the different idioms in those  

natural languages. Just to give you an example,  in English you say, it rains cats and dogs. Well,   in France, in French it doesn't mean anything,  right? In French, actually, it rains ropes,   right? Which probably doesn't mean anything in  English. [LAUGTER] And so I was really curious   about natural languages and communication.  When I went to school, being good at math,   I ended up doing math, realizing very quickly that  I didn't want to do a career in math. You know,  

proofs all that is good in high school, doing a  full career, was not my thing, math. You know,   proofs, all that. It’s good in high school,  but doing a full career, it was not my thing,   math. But there was that class I really, really  enjoyed, which was mathematical logic. And so   little by little, I started discovering people  working in that field. And at the same time,   I was still restless with natural languages. And  so I also took some classes in linguistics on the  

humanity university in Toulouse in France. And I  stumbled on those people who were actually working   in… some in linguistics, some in computer science,  and then there was this lab doing computational   linguistics. And then that was it for me. I was  like, that's, you know, so that's how I ended up   doing my PhD in computational linguistics. And the  last aspect I'll talk about, because in my role   today, the aspect of working with a network  of people, with a global network, is still  

so important to me, and I think for science as a  whole. At the time, there was this nascent field  

of computational lexical semantics. And for me, it  was so important to bring people together because   I realized that we all had different approaches,  different theories, not even in France,   but across the world, and actually, I worked with  somebody else, and we co-edited the first book on   computational lexical semantics, where we started  exposing what it meant to do lexical semantics and  

the relationships between words within a larger  context, with a larger context of conversations,   discourse, and all those different approaches.  And that's an aspect which for me to this day   is so important and that was also really  important to keep as we develop what we're   going to talk about today, Accelerating  Foundation Models Research program.

topic of Accelerating Foundation Models Research  and unpack the big idea behind that. Evelyne,   I want to stay with you on this for a minute  because I'm curious as to how this initiative even   came to exist and what it hopes to achieve. So,  maybe start out with a breakdown of the title. It   might be confusing for some people, Accelerating  Foundation Models Research. What is it?

think of language models as just one part, right,  for those foundation models. The term was actually   coined at Stanford when people started looking  at GPTs, the generative pre-trained transformers,   this new architecture. And so that term was coined  like to go not just talk about language models,   but foundation models, because actually it's not  just language models, but there are also vision  

models. And so there are other types of models  and modalities really. And so when we started   with Accelerating Foundation Models Research and  from now on, I will say AFMR if that's okay.

to look at safety, to look at trustworthiness of  those models. And so we cannot just drink our own   Kool-Aid and so we really had to work with people  externally. And so we were already doing that. But   that was an effort which we couldn't scale really  because to scale an effort and having multiple   people that can have access to the resources,  you need more of a programmatic way to be able   to do that and rely on some platform, like for  instance, Azure, which has security and privacy,  

confidentiality which enables to scale those  type of efforts. And so what happens as we're   developing this program on the Turing model with  a small set of academic people, then there was   this ChatGPT moment in November 2022, which  was the moment like the “aha moment,” I think,  

as I mentioned, for me, it's like, wow, AI now has  made it to industry. And so for us, it became very   clear that we could not with this moment and the  amount of resources needed on the compute side,   access to actually OpenAI that new that GPT, at  the beginning of GPT-3 and then 4 and then… So how   could we build a program? First, should we, and  was there interest? And academia responded “Yes!  

Please! Of course!” right? [LAUGTER] I mean,  what are you waiting for? So AFMR is really   a program which enabled us to provide access to  foundation models, but it's also a global network   of researchers. And so for us, I think when we  started that program, it was making sure that   AI was made available to anyone and not just the  few, right? And really important to hear from our   academic colleagues, what they were discovering  and covering and what were those questions that  

we were not even really thinking about, right?  So that's how we started with the AFMR.

for you. And since you're in medicine, give us a  little bit of a “how it started, how it's going”   from your perspective, and why it's important  for you at the Morehouse School of Medicine.

architectures, edge cases that we didn't even  realize. So you could imagine when I saw that,   I said, wow, this is amazing. It's going to unlock  so many possibilities. But at the same time, this   demo was coming out, I actually saw a tweet about  the inherent biases that were baked into these   models. And I'll never forget this. I think it  was at the time he was a grad student at Stanford,   and they were able to show that if you asked  the model to complete a very simple sentence,  

a sort of joke, “Two Muslims walk into a bar…”  what is it going to finish? And it was scary.

real world. And of course, the first thing  that this made me do as a researcher was   wonder how do these emerging technologies,  how may they unintentionally lead to greater   health disparities? Maybe they do. Maybe they  don't. The reality is that we don't know.

given my time here at Morehouse School of  Medicine. And kind of what I believe is that,   you know, the likely outcome, and I  would say this is the case for really   any sort of emerging technology, but let's  specifically talk about AI, machine learning,   large language models, is that if we're not  intentional in interrogating how they perform,   then what's likely going to happen is that despite  overall improvements in health, we're going to see  

greater health disparities, right? It's almost  kind of that trickle-down economics type model,   right? And it's really this addressing of health  disparities, which is at the core of the mission   of Morehouse School of Medicine. It is literally  the reason why I came here a few years ago. Now,   the overarching goal of our program,  without getting too specific, is really   around evaluating the capabilities of foundation  models. And those, course, as Evelyne mentioned,  

are large language models. And we're specifically  working on facilitating accessible and culturally   congruent cancer-related health information.  And specifically, we need to understand that   communities that are disproportionately impacted  have specific challenges around trust. And all   of these are kind of obstacles to taking  advantage of things like cancer screenings,   which we know significantly reduce the likelihood  of mortality. And it's going very well. We have  

a pretty amazing interdisciplinary team. And I  think we've been able to develop a pretty cool   research agenda, a few papers and a few grants.  I'd be happy to share about a little bit later.

was actually surprised. I saw that, hey, they are  actually acknowledging the human element. And so   the fact that there was this call for research  that was looking at that human dimension was   really refreshing. So like what Muhammad was  saying, one of the most exciting things about   these large language models is you don't have to  be a computer scientist in order to use them. And   it reminded me to this moment in time within the  arts when digital media started getting produced.  

And we had this crisis. There was this idea that  we would lose all the skills that we have learned   from working traditionally with physical materials  and having to move into a digital canvas.

We've been doing some collaborations with  them, and we've been essentially monitoring   how they bend glass. I run an undergraduate  research program here and I’ve had undergrads   try to tackle this problem of how do you  transfer that skill of neon bending? And   the fact is that because of AFMR, here is just  kind of a way to structure that undergraduate   research process so that people feel comfortable  to ask those dumb questions exactly where they  

are. But what I think is even more exciting is  that they start to see that questions like skill   acquisition is still something that our AI is  not able to do. And so it's refreshing to see;   it's like the research problems have not all  been solved. It just means that new ones have   opened and ones that we previously thought  were unattainable now have this groundwork,   this foundation in order to be researched, to be  investigated. And so it's really fertile ground.  

And I really thank AFMR… the AFMR program for  letting us have access to those grounds.

so what do these grants involve? You've alluded  to it already, but can you give us some more   specifics on how Microsoft is uniquely positioned  to give these and what they're doing?

really had no participation from MSIs. [LAUGHTER]  And then we thought, why? It's open… it’s…. and   I think what we missed there, at the beginning,  is like we really focused on the technology and   some people who were already a part of the kind  of, this global network, started approaching us,  

but actually a lot of people didn't even know,  didn't think they could apply, right? And so   we ended up doing a more targeted call where we  provided not only access to the compute resources,   access to the APIs to be able to develop  applications or validate or expand the   work which is being done with foundation models,  but also we acknowledged that it was important,   with MSIs, to also enable the students  of the researchers like Cesar, Muhammed,  

and other professors who are part of the  program so that they could actually spend   the time working on those projects because there  are some communities where the teaching load is   really high compared to other communities or other  colleges. So we already had a good sense that one   size doesn't fit all. And I think what came  also with the MSIs and others, it's like also  

one culture doesn't fit all, right? So it's about  access. It's about access to people, access to the   resources and really co-designing so that we can  really, really make more advances together.

I'm sure are near and dear to Evelyne's heart…  the French language. Talk about that, César.

polymer fabrication, 3D printing. And so every  year I throw something new in there. And this   last year, what I threw in there was GPT and large  language models. And it has been exciting to see   how it has transformed. But speaking to this  specific project, I think the best way I can   describe bricolage is to ask you a question:  what would you do if you had a paperclip,  

duct tape, and a chewing gum wrapper? What could  you make with that, right? [LAUGHTER] And so some   of us have these MacGyver-type mentalities,  and that is what Claude Lévi-Strauss kind of   terms as the “bricoleur,” a person who is able to  improvise solutions with the materials that they   have at hand. But all too often, when we think  about bricolage, it's about the physical world.   But the reality is that we very much live in a  hybrid reality where we are behind our screens.  

And that does not mean that we cannot engage in  these bricoleur activities. And so this project   that I was looking at, it's both a vice and an  opportunity of the human psyche, and it's known   as “functional fixation.” And that is to say, for  example, if I were to give you a hammer, you would   see everything as a nail. And while this helps  kind of constrain creative thought and action  

to say, okay, if I have this tool, I'm going to  use it in this particular way. At the same time,   it limits the other potential solutions, the ways  that you could use a hammer in unexpected ways,   whether it's to weigh something down or  like jewelers to texturize a metal piece or,  

I don't know, even to use it as a pendulum...  But my point here is that this is where large   language models can come in because they can,  from a more unbiased perspective, not having   the cognitive bias of functional fixation say,  hey, here is some tool, here's some material,   here's some machine. Here are all the ways  that I know people have used it. Here are  

other ways that it could be extended. And so we  have been exploring, you know, how can we alter   the physical and virtual environment in such a way  so that this information just percolates into the   creative practitioner’s mind in that moment when  they're trying to have that creative thought?   And we've had some fun with it. I did a workshop  at an event known as OurCS here at DFW. It's  

a research weekend where we bring a couple of  undergrads and expose them to research. And we   found that it's actually the case that it's not  AI that does better, and it's also not the case   that the practitioner does better! [LAUGHTER] It's  when they hybridize that you really kind of lock  

into the full kind of creative thought that  could emerge. And so we've been steadily   moving this project forward, expanding from our  data sets, essentially, to look at the corpus of   video tutorials that people have published all  around the web to find the weird and quirky ways   that they have extended and shaped new techniques  and materials to advance creative thought. So…

fabulous. I want Muhammed's story now. I have to  also calm down. It's so much fun. [LAUGHTER]

is the whole thing about research is how do  these things come together and help? Tell us,   Muhammed, about this cultury… culturally…  Tell us about your project! [LAUGHTER]

you might not get the help that you need. And if  you think about it, a lot of like the kind of the   through line through all of these, it really  comes down to health communication at the end   of the day. It's not just what people are saying,  it's how people are saying it as well. And so our  

project focuses right now on language and text,  right? But we are, as I'll talk about in a second,   really exploring the kind of multimodal nature  of communication more broadly and so, you know,   I think another thing that's important in terms  of just background context is that for us,   these models are more than just tools, right?  We really do feel that if we're intentional   about it that they can be important facilitators  for public health more broadly. And that's where  

this idea of our project fitting under the bucket  at benefiting society as a whole. Now, you know,   the context is that over the past couple of  decades, how we've talked about cancer, how   we've shared health information has just changed  dramatically. And a lot of this has to do with   the rise, of course, of digital technologies more  broadly, social media, and now there's AI. People  

have more access to health information than ever  before. And despite all of these advancements,   of course, as I keep saying over and over again,  not everyone's benefiting equally, especially when   it comes to cancer screening. Now, breast and  cervical cancer, that's what we're focusing on  

specifically, are two of the leading causes of  cancer-related deaths in women worldwide. And   actually, black and Hispanic women in the US  are at particular risk and disproportionately   impacted by not just lower screening rates,  but later diagnoses, and of course from that,   higher mortality rates as well. Now again, an  important part of the context here is COVID-19.   I think there are, by some estimates, about 10  million cancer screenings that didn't happen. And  

this is also happening within a context of just  a massive amount of misinformation. It's actually   something that the WHO termed as an infodemic.  And so our project is trying to kind of look for   creative emerging technologies-based solutions for  this. And I think we're doing it in a few unique  

ways. Now the first way is that we're looking  at how foundation models like the GPTs but also   open-source models and those that are, let's say,  specifically fine-tuned on medical texts, how do   they perform in terms of their ability to generate  health information? How accurate are they? How   well is it written? And whether it's actually  useful for the communities that need it the   most. We developed an evaluation framework, and we  embedded within that some qualitative dimensions  

that are important to health communications. And  we just wrapped up an analysis where we compared   the general-purpose models like a ChatGPT with  medical and more science-specific domain models   and as you'd expect, the general-purpose models  kind of produced information that was easier to   understand, but that was of course at the risk  of safety and more accurate responses that the  

medically tuned models were able to produce. Now  a second aspect of our work, and I think this is   really a unique part of not what I've called,  but actually literally there's a book called   The Morehouse Model, is how is it that we could  actually integrate communities into research? And   specifically my work is thinking about how do  we integrate communities into the development   and evaluation of language models? And that's  where we get the term “culturally congruent.”  

That these models are not just accurate,  but they're also aligned with the values,   the beliefs, and even the communication styles of  the communities that they're meant to serve. One   of the things that we're thinking, you know,  quite a bit about, right, is that these are   not just tools to be published on and maybe  put in a GitHub, you know, repo somewhere,   right? That these are actually meant to drive  the sort of interventions that we need within  

community. So of course, implementation is really  key. And so for this, you know, not only do you   need to understand the context within which  these models will be deployed, the goal here   really is to activate you and prepare you with  information to be able to advocate for yourself   once you actually see your doctor, right? So that  again, I think is a good example of that. But you   also have to keep in mind Gretchen that, you know,  our goal here is, we don't want to create greater  

disparities between those who have and those who  don't, right? And so for example, thinking about   accessibility is a big thing and that's been a  part of our project as well. And so for example,   we're leveraging some of Azure API services for  speech-to-text and we're even going as far as   trying to leverage some of the text-to-image  models to develop visuals that address health   literacy barriers and try to leverage these  tools to truly, truly benefit health.

astronomy models, right? And I'm thinking, okay,  that could… maybe also we learn something which   could be potentially applied to some other domain.  So these are some of the things we are seeing.

I think it's a different way to find  information and gain some knowledge. I   think that actually when we look at that, these  are really broad tools that now actually can be   used to help customize your own curb, as it  were! So kind of the other way around.

sort of tiered structure. And I'm making a tree  diagram with my hands here! [LAUGHTER] And so we   often talk about what it's like for an outsider to  enter from outside of the community, and just how   much effort it takes to get through that gate,  to go through the different rungs, through the   different rites of passage, to finally be a part  of the inner circle, so to speak. And one of the   projects that we've been doing, we started to  examine these known communities of practice,  

where they exist. But in doing this analysis,  we realized that there's a couple of folks out   there that exist on the periphery. And by really  focusing on them, we could start to see where the   field is starting to move. And these are folks  that have said, I'm neither in this community or   another, I'm going to kind of pave my own way.  While we're still seeing those effects of that   research go through, I think being able to monitor  the communities at the fringe is a really telling  

sign of how we're advancing as a society. I  think shining some light into these fringe areas,   it's exactly how research develops, how it's  really just about expanding at some bleeding   edge. And I think sometimes we just have to  recontextualize that that bleeding edge is   sometimes the group of people that we haven't  been necessarily paying attention to.

right? The cultural context that's important to  train these models on. And the reality is that   we started off by saying, let's just have focus  groups. Let's talk to people. But of course that   takes time, it takes money, it takes effort. And  what we quickly realized actually is there are   literally generations of people who have done  these focus groups specifically on breast and  

cervical cancer screening. And so what we  actually have since done is leverage that   real world data in order to actually start  developing synthetic data sets that are…

concerns around bias that might not exist. And so  for me, that's a really like awesome thing that   we came across that I think in trying to solve  a problem for our kind of specific use case,   I think this could actually be a method  for developing more representative,   context-aware, culturally sensitive models  and I think overall this contributes to the   overall safety and reliability  of these large language models  

and hopefully can create a method for  people to be able to do it as well.

researchers are thinking about the potential  downsides and consequences of their work. So,   Evelyne, do you have any insights on  things that you've discovered along   the path that might make you take  preemptive steps to mitigate?

human beings. And I feel that it’s uh… we need to  remember that it's a proxy to humans and that we   all have some different beliefs, values, goals,  preferences. And so how do we take all that into   account? And I think that beyond the data safety,  provenance, I think there's an aspect of “data   caring.” I don't know how to say it differently,  [LAUGHTER] but it's kind of in the same way that  

we care for people, how do we care for the data as  a proxy to humans? And I'm thinking of, you know,   when we talk about like in, especially in  cases where there is no economic value,   right? [LAUGHTER] And so, but there is local value  for those communities. And I think actually there   is cultural value across countries. So just wanted  to say that there is also an aspect, I think we   need to do more research on, as data as proxies  to humans. And as complex humans we are, right?

the “what if” questions that make us excited. So  just briefly, is there anything on your horizon,   specifically Cesar and Muhammed, that you would  say, in light of this program, AFMR, that you've   had access to things that you think, boy, this  now would enable me to ask those bigger questions   or that bigger question. I don't know what it  is. Can you share anything on that line?

if something's easier, then some other thing  is harder. And then we run into this really   strange case where if everything is easy, then  we are faced with the “blank canvas syndrome,”   right? Like what do you even do if everything  is just equally weighted with ease? And so my   big idea is to actually think about tools  that are purposely making us slower…

really it becomes an infrastructure  problem. And so the more that we expose   people to these foundational models, the  more that we're going to be able to see   how we can enable these new ways of walking  through and exploring our environment.

and which drink I want. I love this, Cesar. That's  something I hadn't even expected you might say and   boom, out of the park. Muhammad, do you have any  sort of outrageous…? That was flipping it back!

access to all the care they need? What would  happen then? Would we magically all be the   healthiest version of ourselves? I'm a little bit  skeptical. I'm not going to lie, right? [LAUGHTER]  

But that's something that I'm often thinking  about. Now, bringing that back down to our   project, one of the things that I find a little  bit amusing is that I tend to ping-pong between,   this is amazing, the capabilities are just, the  possibilities are endless; and then there will be   kind of one or two small things where it's pretty  obvious that there's still a lot of research that  

needs to be done, right? So my whole, my big “what  if” actually, I want to bring that back down to   a kind of a technical thing which is, what if AI  can truly understand culture, not just language,   right? And so right now, right, an AI model can  translate a public health message. It's pretty   straightforward from English to Spanish, right?  But it doesn't inherently understand why some  

Spanish speaking countries may be more hesitant  about certain medical interventions. It doesn't   inherently appreciate the historical context that  shapes that hesitancy or what kinds of messaging   would build trust rather than skepticism, right?  So there’s literal like cultural nuances. That to  

me is what, when I say culturally congruent or  cultural context, what it is that I mean. And I   think for me, I think what programs like AFMR have  enabled us to do is really start thinking outside   the box as to how will these, or how can these,  emerging technologies revolutionize public health?  

What truly would it take for an LLM to understand  context? And really, I think for the first time,   we can truly, truly achieve personalized, if  you want to use that term, health communication.   And so that's what I would say for me is  like, what would that world look like?

of people thinking, having crazy ideas, very  creative in the research being done. And there   is some research where people are looking at what  it means, actually, when you build those language   models and how you can take into account different  language and different culture or different   languages within the same culture or between  different cultures speaking the same language,  

or… So there is very interesting research. And so  it made me think, expanding on what Muhammed and   Cesar were talking about, so this educational  rite of passage, I don't know if you're aware,   so in Europe in the 17th, 18th century, there was  this grand tour of Europe and that was reserved   to just some people who had the funds to do that  grand tour of Europe, [LAUGHTER] let's be clear!  

But it was this educational rite of passage where  actually they had to physically go to different   countries to actually get familiar and experience,  experiment, philosophy and different types of   politics, and… So that was kind of this “passage  obligé” we say in French. I don't know if there is  

a translation in English, but kind of this rite of  passage basically. And so I am like, wow, what if   actually we could have, thanks to the AI looking  at different nuances of cultures, of languages…   not just language, but in a multimodal point of  viewpoint, what if we could have this “citizen of   the world” rite of passage, where we… before  we are really citizens of the world, we need   to understand other cultures, at least be exposed  to them. So that would be my “what if?” How do we  

make AI do that? And so without… and for anyone,  right, not just people who can afford it.

that may not have seemed that way when you  started? And you can also, I'm anticipating   you people are going to flip that and say,  what wasn't important that is now? And also,   how do see yourself moving forward in light of  this experience that you've had? So Muhammed,   let's go first with you, then Cesar, and  then Evelyne, you can close the show.

concept called Ihsan. Ihsan… you know there's a  lot of definitions of it but essentially to do   more than just the bare minimum to truly strive  for excellence and I think it was interesting,   having spent time at Microsoft Research  in Redmond as part of the AFMR program,   meeting other folks who also participated in  the program that, that I started to appreciate   for myself the importance of this idea of the  responsible design, development, and deployment  

of technologies if we truly are going to achieve  the potential benefits. And I think this is one of   the things that I could kind of throw out there  as something to take away from this podcast,   it's really, don't just think of what we're  developing as tools, but also think of them as how  

will they be applied in the real world? And when  you're thinking about the context within which   something is going to be deployed, that brings up  a lot of interesting constraints, opportunities,   and just context that I think is important, again,  to not just work on an interesting technology   for the sake of an interesting technology, but  to truly achieve that benefit for society.

others to also find their own path. I was setting  up Azure from one of my classes and it took time,   it took effort, but the payoff has been incredible  in… in so much the impact that I see now of   students from my class sharing with their peers.  And I think this culture of entrepreneurship   really comes from taking ownership of where you've  been and where you can go. But it really just,  

it all comes down to infrastructure. And so AFMR  for me has been that infrastructure to kind of get   my foot out the door and also have the ability to  bring some folks along the journey with me, so…

the face here, or well, not the face, the  voice rather! [LAUGHTER] But it's so many   people who have, at Microsoft on the engineering  side, we're just talking about infrastructure,   Cesar was talking about, you know,  the pain and gain of leveraging an   industry-grade infrastructure like Azure and Azure  AI services. So, also our policy teams, of course,  

our researchers. But above all, the external  research community… so grateful to see. It’s,   as you said, I feel super blessed and fortunate  to be working on this program and really   listening what we need to do next. How can we  together do better? There is one thing for me,   I want to end on the community, right? Muhammed  talked about this, Cesar too, the human aspect,  

right? The technology is super important but also  understanding the human aspect. And I will say,   actually, my “curb cut moment” for me [LAUGHTER]  was really working with the MSIs and the cohort,   including Muhammed and Cesar, when they came to  Redmond, and really understanding some of the   needs which were going beyond the infrastructure,  beyond you know a small network, how we can put   it bigger and deployments ideas too, coming  from the community and that's something which  

actually we also try to bring to the whole of AFMR  moving forward. And I will finish on one note,   which for me is really important moving  forward. We heard from Muhammed talking about   the really importance of interdisciplinarity,  right, and let us not work in silo. And so,   and I want to see AFMR go more international,  internationality if the word exists… [LAUGHTER]