Learning Library

← Back to Library

AI Shaping Sports, Code, and Personas

Key Points

  • The episode kicks off by exploring how AI could transform major sports events like Wimbledon, the Euros, and Copa America, from performance analytics to enhancing fan experiences.
  • A new study from the *IEEE Transactions on Software Engineering* examines GPT’s ability to solve coding tasks, raising concerns about over‑reliance on AI tools for novice programmers.
  • Researchers have released a paper on generating “1 billion personas” as synthetic data, sparking discussion about whether such massive persona libraries can truly capture human diversity for training LLMs.
  • Host Tim Hong is joined by IBM experts Skyler Speakman, Kar El McGrow, and newcomer Aaron Botman to dissect these AI trends and their broader implications.
  • The conversation also touches on the practical insights gained from attending Wimbledon in person, illustrating real‑world opportunities for AI integration in sports.

Full Transcript

# AI Shaping Sports, Code, and Personas **Source:** [https://www.youtube.com/watch?v=aMcYxjZMRuA](https://www.youtube.com/watch?v=aMcYxjZMRuA) **Duration:** 00:41:24 ## Summary - The episode kicks off by exploring how AI could transform major sports events like Wimbledon, the Euros, and Copa America, from performance analytics to enhancing fan experiences. - A new study from the *IEEE Transactions on Software Engineering* examines GPT’s ability to solve coding tasks, raising concerns about over‑reliance on AI tools for novice programmers. - Researchers have released a paper on generating “1 billion personas” as synthetic data, sparking discussion about whether such massive persona libraries can truly capture human diversity for training LLMs. - Host Tim Hong is joined by IBM experts Skyler Speakman, Kar El McGrow, and newcomer Aaron Botman to dissect these AI trends and their broader implications. - The conversation also touches on the practical insights gained from attending Wimbledon in person, illustrating real‑world opportunities for AI integration in sports. ## Sections - [00:00:00](https://www.youtube.com/watch?v=aMcYxjZMRuA&t=0s) **AI’s Impact on Sports, Coding, and Synthetic Data** - In this episode Tim Hong previews discussions on AI’s role in reshaping athletic competition, a new study evaluating GPT’s coding abilities, and a paper proposing a billion‑person synthetic data framework to alleviate data scarcity. ## Full Transcript
0:00hello and happy Friday you're listening 0:01to mixture of experts I'm your host Tim 0:03Hong back from the kitchen each week 0:06mixture of experts brings together 0:07amazing group of well experts to tackle 0:10debate and explain the biggest Trends in 0:12the fast-moving world of artificial 0:14intelligence this week on the show we're 0:16going to cover three stories first with 0:18the Euros Wimbledon and the cop America 0:20all in a matter of weeks we talk about 0:22Ai and sports will AI have a role in 0:24shaping the nature of the game if so how 0:27you know they they might just see 0:28gameplay whereas we see data and an 0:30opportunity uh to derive these types of 0:33insights uh to help find that signal in 0:35the noise U to provide those 0:37serendipitous type moments that connect 0:39people to the game second a new study 0:41out in the I E's transactions on 0:43software engineering reveals new data 0:44about gpt's performance on coding tasks 0:47what's it tell us about the future of 0:48coding assistance I worry a bit about 0:51you know this uh over Reliance on AI 0:53tools for problem solving especially as 0:55you're learning in the early stages for 0:57programming and then third a fascinating 1:00paper out in archive entitled scaling 1:02synthetic data Creation with 1 billion 1:04personas from 10cent Seattle lab does it 1:07provide a way forwards for resolving 1:08data bottlenecks and what can we use 1:10personas for in the future how confident 1:12are we about the coverage of these 1 1:15billion personas and do the underlying 1:18large language models you know really 1:21understand being a a musai warrior 1:31as always I'm joined by an incredible 1:32group of panelists that will help us 1:33navigate what has been another 1:35action-packed week in AI so today we've 1:37got three panelists Skyler Speakman he's 1:39a senior research scientist at IBM Kar 1:42El mcgrow principal research scientist 1:44at IBM AI engineering AI Hardware Center 1:47and joining us for the very first time 1:48is Aaron botman IBM fellow and master 1:50inventor welcome to the show 1:53[Music] 1:56everyone first it's been a very busy 1:59season if if you're into watching sports 2:01uh Wimbledon the euros and the cop 2:03America are all happening basically like 2:05this week last week um I know we've 2:07talked largely on the show at mixture of 2:09experts on AI as kind of an internal 2:13business process um but I think you know 2:16particularly with all the sports in the 2:17air I've been kind of thinking that it 2:18might be a really good opportunity for 2:20us to talk a little bit about the ways 2:22in which AI might reshape Sports itself 2:25um and Aaron I want to start with you as 2:27kind of our our new panelists uh on the 2:29show just to pick on you a little bit 2:31you were actually at Wimbledon um and 2:33I'm curious as someone who works you 2:35know day in day out right like why did 2:37you go what do you think and as someone 2:40who works in AI all the time I'm sure if 2:41you're experienc of this kind of tennis 2:43tournament where you're like oh actually 2:44there might be a lot of ways for AI to 2:46apply so just kind of as an initial 2:47place just to to get the report uh on on 2:50uh 2:51wiblin yeah so I mean it's it's always 2:54fascinating you know to watch how we 2:56operationalize lots of these AI 2:57techniques in particular in this case 2:59with um our partner um you know um 3:02Wimbledon and I was lucky enough to go 3:05you know we've been doing this for uh 3:07with them almost 30 years now um and 3:10this year you know we focused a lot on 3:12generative AI um but also don't want to 3:14forget about classical AI right because 3:17both of them are uh very important um 3:19and we use many different uh techniques 3:22in order to ad do it but to actually be 3:24there right during the tennis and being 3:26and in sort of in the thick right of uh 3:29the space it's it's very um interesting 3:31um you know because there's there's many 3:32aspects there's a how's the technology 3:35performing how is the consumer 3:37acceptance you know of the tech um and 3:40then how's the back office um acceptance 3:42as well right um you know so it's it's 3:45always nice right to watch people around 3:48the world you know use it um you know we 3:51get billions of of users every single 3:53year you know that that use you know our 3:56systems um you know in this case for the 3:58generative AI just at the halfway mark 4:01right um people spent um 4:042,319 hours just looking and reading the 4:08generative content that we we produce 4:10well and I think if I can ask you to 4:11back up a little bit I mean our 4:12listeners won't necessarily be familiar 4:14with what you were working on we'd love 4:16to hear kind of a little bit more about 4:17like what is the technology that you 4:19were mostly focused on this year and uh 4:21what people were doing with it yeah so 4:23um so so we're um looking at um bringing 4:26the game in a personalized way to you 4:29right and so and so what we like to do 4:31is mix in different aspects you know we 4:34like to rank players we like to predict 4:37who might win a match and then we want 4:40to um create content to catch you up 4:43right so that you know if you join um in 4:46the tournament uh we want to create 4:47these digestible nuggets so that you 4:50know the consumers um around the world 4:52can you know View and understand you 4:54know what what's happening uh in the 4:56match right and it and it helps you know 4:59um the you know whereas you know they 5:01they might just see gameplay whereas we 5:03see data and an opportunity uh to derive 5:06these types of insights uh to help find 5:08that signal in the noise um to provide 5:10those serendipitous type moments that 5:12connect people to the game yeah for sure 5:14and I was talking with a friend recently 5:15about this was um so as someone who got 5:17into uh football soccer that is uh you 5:21know during my time like in the pandemic 5:23my experience of the sport has largely 5:25been like a visual experience right and 5:27it is so interesting to me that like you 5:30know having never gone to a game I'm a 5:31huge fan I watch all the time uh but my 5:33primary experience is kind of like 5:35intermediated right uh through you know 5:37social media and what I see on TV and it 5:40kind of sounds like there's been a 5:41similar exercise to try to figure out 5:42like how AI kind of plays a role in that 5:44interface right from like the fan and 5:46the viewer to kind of get more out of 5:48the game um and um I guess I'm Aon I'm 5:51curious like any Lessons Learned um 5:52things that you thought like worked 5:54really well um for this uh this work 5:56yeah um so so we used lots of different 5:58sensors around the course to gather data 6:00you know so we use like a Hawkeye system 6:01that has you know up to nine different 6:03cameras that track the ball track the 6:05players you know we get all sorts of 6:07stats that are streaming to us but 6:09there's just this Deluge of information 6:11right and it's hard for people to just 6:13comprehend it so one of the lessons 6:15learn that I think we had was to create 6:17these digestible narrations of you know 6:20pre-match postmatch about the players so 6:23they can go in and just quickly read up 6:25you know on their favorite players um 6:27and then and then some other um aspects 6:29that that we also learned is that 6:31sometimes it's nice to inject 6:32information that maybe somebody wouldn't 6:34ordinarily you know know about or read 6:36about or even think about right um so so 6:39it's nice you know to to watch that you 6:42know happen and and spread um so that's 6:45you know you know one of the pillars um 6:47I think just very quickly the other 6:49pillar uh would be on the operation side 6:51you know um that it's always great to 6:53have human and and machine and 6:55algorithms working together uh to create 6:57the symbiotic you know experience that 7:00can be um used um where whether it's 7:03mobile uh you might be on site you know 7:05as a fan um so it's uh so it's it's it's 7:08really evolving right into this sort of 7:10Moneyball 2.0 yeah I think uh the uh the 7:14power of AI uh in sports uh really is 7:17transformative and uh and AI here plays 7:20a multifaceted and transformative role 7:23uh espec especially not just on the 7:25commentary side on the user experience I 7:27think there are lots of different 7:28applications 7:30where we can see the power of AI here 7:32like things like performance analysis 7:34athlete and performance training where 7:36with using wearable technology you can 7:38have these sensors that collect data on 7:40the athletes movements the Biometrics 7:42Etc and analyze the data to provide 7:45insights of you know where the athlet 7:47can you know improve uh video analysis 7:50uh you know analyzing video footage of 7:52the training sessions and games 7:54assessing techniques identifying 7:56weaknesses game strategy also is also a 7:59a big application here health and injury 8:02prevention uh with doing things like 8:04injury 8:05diagnosis uh can algorithms also can 8:08assessing diagnosing injury injuries 8:10through image analysis uh fan engagement 8:14and experience of course is the the fun 8:15part of it uh like Aaron was talking 8:18about you know with the personalized 8:20content uh with the chatbot and virtual 8:23assistance and the augmented reality 8:25even you know you can have kind of a 8:27true immersive experience where you can 8:30enhance you know AR and VR experiences 8:32imagine watching a game and like you're 8:34there you know with the AR and VR I 8:36think that can be really a lot of fun 8:39and of course with all the game and 8:41event management you know there are also 8:43areas where you can use AI for 8:44scheduling logistic crowd management 8:47ticketing uh so lots of you know areas 8:50here where Ai and also gen AI can really 8:53play a transformative role and I I can 8:55only see this kind of uh growing that's 8:58right yeah I love the idea that in the 9:00future you'll be able to get like 9:01whatever commentator you want just 9:03generated algorithmically on the Fly you 9:05know like I want you know the uh you I 9:08want George Washington to narrate my my 9:11sports game and to have that audio 9:13generated on the fly would be really 9:14interesting I also think this kind of 9:16point about kind of like the backend is 9:18also really interesting about like all 9:19the operations it will help with and I 9:21know Skyler you were interested in 9:22particularly the idea that kind of you 9:24know maybe teams that will be able to 9:25like really manage all this data will 9:27have this huge advantage in the future 9:29um and like it'll be like a wonderful 9:31world where basically like you know 9:32someone managing a top tennis play in 9:34the future will like also be trying to 9:35get h100s to run their own fine-tuning 9:37runs so maybe actually two questions 9:40along those lines both to Aaron do you 9:43know of any of the tennis players have 9:45they used this are they looking at their 9:49narrative that was generated so you know 9:51has has it reached the player side I 9:53know we're talking about consumer facing 9:55Tech at this point but have any of the 9:56players commented uh and the second one 9:59is when is IBM going to bring this 10:01technology to Esports you know so that 10:03the data is almost already in a more 10:05usable format there but there's can be 10:07just as much hype and excitement and 10:09drama in some of these more recent um 10:12orts that are coming out and I think 10:13there's a great opportunity to bring 10:15this technology um to uh to electronic 10:18gaming that's one of my favorite 10:20pastimes so yeah Aon any comments on 10:23those yeah so so first you know um great 10:26great questions you know and suggestions 10:28um I think think that do you know 10:30players actually use some of our 10:32information and and it's really funny 10:35some of them do some of them don't some 10:37of them are very superstitious right if 10:39they were to look at one of our 10:40predictions right then it would sort of 10:43mentally affect the way in which they 10:44play the game upcoming and so some 10:46coaches do not let their players you 10:48know look at some of our uh features and 10:51then and then some properties were not 10:53even allowed to you know during a game 10:55to show you know any sort of predictions 10:57or um sometimes 10:59even gen content because it might 11:01influence right play as well but on the 11:05other side of the coin some players you 11:07know have you know used it and and they 11:09do look you know at the 11:10stats um you know that that we boil down 11:13and we we also had a project um with um 11:17the US Open um so um you know we worked 11:20with some ATP players and so on where um 11:23we we would help them train you know so 11:25they would see videos of themselves 11:26playing we would find highlights of of 11:29of they played so it was like a 11:30dashboard you know with the 11:31Developmental Center um so so so there's 11:34that that aspect and I I I was curious 11:36to do any of you play tennis or Sports 11:39and and not very well and and would you 11:41use these kind of insights 11:43or I think I would I would try to use 11:46especially if you know maybe trying to 11:49help with my performance I mean I'm not 11:50an active Sports person but I hope you 11:53know to help me maybe improve techniques 11:56and things like that but another thing 11:58is there any downside to this um and I I 12:01I worry a little bit maybe about the 12:03bias and fairness uh with with certain 12:06athletes these are always you know Flags 12:09you know red flags that we could have 12:11with the use of AI uh the AI systems can 12:14inherit biases pres present you know in 12:16the training data could this be lead to 12:18unfair treatment of the athletes for 12:20example biased scouting or algorithms 12:23might Overlook maybe some talented in 12:25individuals from under represented 12:27groups so there are some you know 12:29dangers I don't know arony this is 12:31something that you think the current 12:33algorithms are taken seriously or this 12:36is still early on and we're just 12:39evaluating the technology right now and 12:41maybe you know start looking seriously 12:43into these concerns privacy issue was 12:46bias fairness yeah fantastic topic that 12:50that could take you know hours right to 12:52talk through but um so so yes um you 12:56know fairness transparency and 12:58explainability you know where in gen 13:00might call Chain of Thought to 13:02understand what's output from the models 13:04uh but but a story real quick um you 13:07know in in tennis we used to measure um 13:10the or still do measure the excitement 13:12of videos right so we'll look at for 13:14example um signals like sound right 13:16gestures score um and we quickly found 13:20out that um somebody who's an amateur 13:23right uh who's playing golf they might 13:26have a really exciting shot but because 13:28they're not very popular player there's 13:29not a lot of people around them to make 13:31a very loud cheer right whereas there 13:34might be a you know top five ranked 13:37player who makes a routine shot it's not 13:39that exciting but has a huge cheer 13:41because there just happens to be a lot 13:43of people there right so so we take that 13:45into account and we'll debias with with 13:48postprocessors you know based on 13:50different uh restrictive uh traits that 13:52we have you know uh because yeah it's 13:55it's real right and and we work to make 13:58sure you know that we can debias these 14:00types and and and there's many debiasing 14:02ways and methods right and and the Gen 14:04AI space um you know we're I think just 14:07beginning you know um around that and um 14:11and a question for you all is you know 14:13um we try to balance creativity with 14:15factualness you know with these 14:17different generative content um you know 14:20how how do you think the field can do a 14:23better job right at doing that you know 14:25uh with respect if you think there's 14:26hallucination if you're more creative 14:28you know whereas you need fact Checkers 14:30you know um and and so on and so forth I 14:33think one thing that will become if we 14:35continue to kind of collect this data 14:37you'll be able to ask questions about 14:40how exciting was the current top star 14:44when they were just starting out you'll 14:46be able to go back in time 10 years and 14:48look at that top star when nobody was 14:50following him but he was still making 14:51the great shots we probably don't have 14:53that now because we don't have as much 14:54historic data but um if you can you know 14:57you'll really be able to watch entire 14:59uh I think careers play out over time uh 15:02at least with your ideas of um the guy 15:05who's not popular now but made a great 15:07shot um you'll be able to ask that same 15:09question all right Michael Jordan His 15:11freshman year of college didn't make the 15:13team you know that that type of that 15:15type of perspective but we're not going 15:16to be able to do that with the snapshots 15:18of data we have currently yeah and I am 15:20hoping that some of these tools will 15:22actually help teams kind of like see 15:23around corners right like I think some 15:25of the most interesting times in sports 15:27is when someone comes up with an 15:28entirely new strategy right that kind of 15:30like totally changes the nature of the 15:32game um and hopefully with data there's 15:34like a chance to kind of identify a lot 15:35of things that we might not otherwise uh 15:37in in the past so I'm going to wrap up 15:40this section I guess Aaron maybe 15:41question just to throw it back to you is 15:43you know it seems like you've done a 15:44bunch of work in tennis right so like 15:46Wimbledon US Open um and I'm kind of 15:49curious if there's like as a as an AI 15:51researcher is there kind of like a dream 15:53sport you'd really want to kind of like 15:55apply some of your techniques to or ones 15:57that like you know haven't really been 15:58investig at you know I ass seee one of 16:00the reasons for tennis is like you know 16:01it's a lot more controlled right you can 16:03like set up a bunch of cameras there's 16:04kind of a defined place where it happens 16:06but I'm I'm curious from like almost a 16:08CS Point like what's the next most 16:09exciting you know sport to get aied and 16:12and why yeah so so we focused a lot on 16:15tennis golf um we you know we're doing 16:18um some you know Fantasy Football uh 16:21which borders on e-gaming uh we we did 16:23do some e-gaming uh with with the 16:25OverWatch um which was very interesting 16:28um but I think um exploring um the 16:32intersection of gaming uh with that of a 16:35sport um you know because um I really 16:39enjoy the challenge of - gaming uh 16:41because you know um the the the physics 16:44engine can change you know you can get 16:46new skills and new abilities on the Fly 16:48you can get powerups you know so it's 16:50different um and the and your models 16:52have to adjust very quickly um and maybe 16:56quickly online learn a new her right 16:59that uh you know is transported into the 17:01game so that's that's interesting and 17:04and in a real live aspect one of my 17:06favorite sports to watch is basketball 17:08you know um um I I would love to um you 17:12know um look at that um analyze more of 17:14the team aspects um in play um and then 17:18also look at look at the Olympics um you 17:20know um um I saw an article where uh I 17:23believe it's NBC they're going to be 17:25using generative AI to recap you know 17:28some of the matches so I'm very curious 17:29about uh what they're going to do and 17:32how how that's going to um you know uh 17:34be accepted really you know by um the uh 17:37population but yeah so that's that's my 17:40answer I'm sticking with 17:43[Music] 17:47it there's been of course a lot of hype 17:50around the ability for generative AI to 17:53assist with software engineering um and 17:56a lot of excitement about the idea that 17:58at some point AI might just do the 18:00coder's job entirely from end to end um 18:03co-pilot of course um one of the most 18:05kind of successful I think products of 18:07the Gen AI era is is a great example of 18:10this um and there's a great paper that 18:12came out uh just last week in the i e 18:15transactions on software engineering 18:16entitled no need to lift a finger 18:18anymore question mark assessing the code 18:21quality of code Generation by chat GPT 18:24and basically the idea is to say okay 18:25well we know that you know these systems 18:27can can code 18:29um how good are they at doing it and so 18:31it benchmarks chbt against a number of 18:34different coding challenges to assess 18:36how well it is at generating code um and 18:38I think there's kind of two interesting 18:40findings I wanted to discuss with the 18:41group here today um you know the first 18:44one is basically that it turns out that 18:46these coding platforms have you know 18:48chat in particular has this huge 18:50variance in its ability to do coding 18:52tasks right so it turns out that for you 18:54know tasks in this Benchmark labeled 18:56hard it's only able to get it right 18:57about 40% of the time 18:59and then for easy tasks it's up to like 19:0189% and I'm kind of curious as folks on 19:05the call who all code and presumably use 19:07stuff like you know co-pilot you know I 19:09think there's been a narrative which is 19:11you know okay these coding assistants 19:13are basically just like stack exchange 19:15Plus+ plus right they just help you 19:17search the internet and get an answer 19:19for easy things um but I'm kind of 19:21curious if you all kind of buy the 19:23skepticism of the paper right which is 19:24to say for the really hard tasks we are 19:27still just not seeing you know llms or 19:30generative AI be able to kind of like 19:33really kind of Advance the 19:34state-ofthe-art or accelerate our 19:35ability to solve truly hard CS problems 19:37and um kind of curious about you know 19:39what you all think about that if that's 19:40just a temporary thing or if that is a 19:42kind of ceiling um that we are all 19:44running into and I guess Kar do you want 19:46to kind of respond to that I know you 19:48might have a view on this particularly 19:49when it kind of comes to sort of like 19:51the the coding task and it's also 19:52relationship to Hardware yeah I really 19:55enjoyed reading the paper I think it it 19:57did a very nice study 19:59uh to evaluate you know DPT for coding 20:02challenges uh with which revealed mixed 20:04performance like like you showed so 20:07influenced also by this training data 20:09cut off and the inher limitations of 20:11existing models so for simple tasks 20:14doing fantastic and I think it'll 20:16continue to do fantastic for complex 20:18things it still has limitations you know 20:21gen AI today is still struggling with 20:23understanding the broader context of a 20:24project you know which leads you know to 20:27suggestions for example that don't 20:28really fit the overall design or 20:30architecture so especially when you have 20:32you know these uh design and complex 20:34systems where we have multiple 20:35components multiple apis that need to 20:38interact with each other uh so it's 20:41still because that requires reasoning 20:44and that there are C limitations with 20:45geni when it comes to reasoning so I 20:48think the complexity of the context 20:50understanding the integration challenges 20:52you know integrating also multiple 20:55components together and then those 20:57interfaces how they communicate with 20:58each other uh so that's still I think 21:01it's a it's a bit of a challenge for for 21:03Gen AI um so I think as we're improving 21:08in terms of the contextual understanding 21:09and the accuracy of the Gen AI models we 21:11will see I think better results uh 21:14better integration and user experience 21:16with these coding challenges but I think 21:18we're still there is still a lot of 21:19research that needs to be done so um and 21:23things like best practices in software 21:25engineering just us clear system design 21:29uh and prompt engineering you know can 21:31enhance also you know these tools but I 21:33think we're still in the early stages a 21:36fun uh contextual story for this uh we 21:38just hosted about 40 high school 21:40students here at the lab uh just to kind 21:42of show them what Industrial Research 21:44like looks like uh and they were asking 21:46questions with our software engineers 21:49and they were asking questions where you 21:51know do you use a co-pilot or do you use 21:54generative AI in in your code and they 21:58they kind of a little bit but to to a 22:00person everyone down our line it was 22:03always referencing stack Overflow and so 22:06I think your your your comparison of 22:08what you kind of really want is this 22:11kind of nice integration between the 22:13tools that were really used to like a 22:15stack Overflow sitting inside your IDE 22:18allowing you to code so much smoother it 22:20was it was a great example because the 22:22high school students weren't as familiar 22:24with this thing called stack Overflow 22:26and our software Engineers are saying no 22:29exact exactly and that was this point 22:33where there was kind of the old and new 22:35hitting together and um those types of 22:38resources are so incredibly useful and 22:40it will be interesting to see if these 22:42code generators um how much how much are 22:46they really taking from stack Overflow 22:48and in that paper they made this really 22:50cool analysis they broke down the coding 22:53questions that were before I think 2018 22:57and the ones that were after 2018 um 22:59don't quote me on the date but it did 23:02very well on the old questions and very 23:04poorly on the new questions suggesting 23:07that the llm is not keeping up with the 23:11most recent content kind of experience 23:13on stack on stack Overflow so it was 23:15really cool to see that breakdown where 23:17they did the performance of the llms 23:19doing great on older established 23:21questions perhaps with answers already 23:24sitting on stack Overflow uh and not so 23:27well on the more recent coding 23:28challenges um that that came up after 23:31the training so I think we're going to 23:33see these these kind of comparisons with 23:35what exists on stack Overflow and what's 23:37been incorporated into the llms but a 23:39really cool a really cool place to see 23:41it play out wouldn't that require for 23:43maybe frequent retraining or re you know 23:46readjustments of the models yeah I think 23:49it can be a solved problem I was just 23:50giving hats off to the researchers who 23:52kind of understood that Nuance between 23:54the coding ability of the models and say 23:57wait a minute this model was trained 23:59roughly about this time let's see if we 24:01can ask it coding questions that didn't 24:03uh exist at least in the common uh the 24:06stack Overflow Universe um at that time 24:08and then they give the performance 24:10breakdowns but yes uh retraining and Co 24:12and and constantly uh taking into 24:14account new information would be a way 24:16to try to address that yeah and I think 24:18this is one of the really interesting 24:19sort of challenges that it brings up 24:20because you know like pre-training right 24:23like updating the training data is like 24:25actually kind of like a it's a it's a 24:27cost intensive task right you like have 24:28these models that don't necessarily get 24:30pre-trained you know every single day um 24:33and so there's this kind of weird thing 24:34that the paper sort of suggests which is 24:36that like if you're if you're working 24:38with older languages right you're 24:40actually going to be in trouble with 24:42these models um and kind of the best way 24:44to survive is like you'll actually see 24:46like everybody trying to migrate to 24:47avoid being automated to like there's 24:49more pressure to basically adopt new 24:50languages and then also similarly like 24:53those new languages are simultaneously 24:55ones that the model are like is not very 24:57good at assisting so it kind of imagines 24:59kind of this like bifurcated world where 25:01there's a bunch of these older systems 25:03that AIS can basically automate most of 25:05the coding for and then kind of this 25:07like Frontier of kind of code that like 25:09essentially can't get automated away um 25:12and uh and has really interesting 25:13implications for like where we see I 25:15think the impact of the technology will 25:17will be I I think this dependence on AI 25:20like if we have this over Reliance on AI 25:23tools I think maybe the danger could be 25:25uh this decline in problem solving 25:27skills would that is that going to be a 25:31problem because if especially the young 25:34generations of programmers if for all 25:36these simple tasks you know did you go 25:38ask tgp you know write a code that does 25:41this for me and how is that going to 25:43impact you know because usually you 25:45learn coding from these simple examples 25:48and then you build on top of that to go 25:50to more complex uh problems so I worry a 25:53bit about you know this uh over Reliance 25:55on AI tools for problem solving 25:57especially as you're learning in the 25:59early stages for programming and as you 26:02build of course maybe that's going to 26:04require new skills that we need to 26:05develop as programmers or coders or 26:07sorts of software developers is figuring 26:10out how to use these tools more 26:11efficiently and how to know whether 26:14these are kind of plausible Solutions or 26:17I need to change them and tweak them 26:20another thing I see is what does it mean 26:23for debugging when there are issues if I 26:26have relied heavily on these code Pilots 26:29to write code for me and when things are 26:31failing will I be able to debug things 26:33properly or should I also rely on AI to 26:36help me debug these things so it's kind 26:38of uh it's interesting to see how the 26:40interplay of all these different things 26:42will come to play and the role of humans 26:45here and coders um I think I don't know 26:49all the answers to this maybe if you 26:51have some insights but there is a 26:53downside and to this and of course lots 26:56of adventages in terms of enhancing code 26:58productivity uh but there are challenges 27:01as well we have to think about yeah yeah 27:04yeah I've seen in the field that uh many 27:05people they'll consult different types 27:08of code assistants you know because 27:10there's many different models that are 27:12specialized around different types of 27:14task and you know this agentic you know 27:16architecture where you have a mixture of 27:19experts right of which you bring 27:21together um such as many different large 27:23language models is almost like coding by 27:26crowd you know in an automated way 27:29and so now it seems like uh these 27:31developers and um scientists um and 27:34operations experts they sort of have to 27:37have analytical capability to discern 27:39about what's the best technique with 27:41these different opinions right because 27:43you're going to have many different 27:44opinions and coding Styles and perhaps 27:46even languages you know being sent to 27:48you right um and and I think one 27:51important aspect that I always try to 27:53follow is that that there's no free 27:54lunch right that um there's not a 27:56perfect algorithm suitable to solve 27:59every single problem right it depends on 28:01the context of the problem of which is 28:02at hand and so with that in mind you 28:05know I think the human really 28:07understands the context of Their 28:09audience what they're trying to build 28:11where they can deploy it uh whereas 28:13these code assistants at least today you 28:15know you know know a limited amount of 28:17the context and therefore it's important 28:19to get multiple large language model 28:21opinions as far as what they should or 28:24shouldn't do um you know and and what 28:26when one area that I have a lot of 28:29interest in is this automatic 28:31transpilation of code so say say you're 28:33running you know um an application in 28:36one language Let say python right um 28:38maybe it could be trans transpiled into 28:40rust right where maybe it would be less 28:42memory intensive you know on the Fly um 28:45and or you could have a human sort of 28:47look and say yes I agree no I don't 28:49agree you know and put sort of take that 28:51analytical approach but um but but I 28:55think it's all emerging right and um and 28:57I'm really excited about you know the 28:59future and uh what what we as IBM can 29:02also do with instruct lab right to um 29:04use skill building right um in a way to 29:08help with this as as was mentioned 29:09before but the sort of timeliness right 29:12of data of which it can understand with 29:15in context learning fine-tuning you know 29:18there's many approaches and and there's 29:19going to be many more uh in the future 29:22yeah I think one big thing that llms 29:24will probably play a big role in is um 29:26you know the average state of 29:28documentation for code is very very poor 29:30and I feel like the one enormous use 29:32case is even outside of coding 29:33assistance just taking a piece of code 29:35and make Shing it's it's like well 29:36documented will be this like huge 29:38Improvement in quality of life uh for 29:41this kind of work I I love that because 29:43documentation is always an afterthought 29:45and it 29:46is you never have time to do it so that 29:49would be a huge help yeah that's right 29:52it'll be so funny the biggest thing 29:53won't be like automated code it will 29:55just be like making sure that like 29:57someone's doing a good job documenting 30:00[Music] 30:03everything well great well I want to 30:05take us to our last topic of the day um 30:08there was a another wild paper um if 30:10you're kind of a weirdo like me uh and 30:12just like browsing archive for fun um 30:14this is one of the papers that kind of 30:16popped up recently that sort of caught 30:17my eye and you know the way it did this 30:19is because it basically is people doing 30:21SEO with their papers so the title is 30:23scaling synthetic data Creation with 1 30:25billion personas and so with a name like 30:27that you know got to click it I got to 30:28download it I got to read it um and some 30:30of us need to even print it out uh like 30:32Skyler here um and uh it's actually 30:35pretty simple idea but I think this 30:37particular group of experts would be 30:38really good to kind of tackle it um you 30:40know to just kind of give the overall 30:42background right the idea is um you know 30:45there's a need to generate synthetic 30:47data often right because collecting real 30:49data out of the real world is very 30:51expensive and comes with all these 30:52operational difficulties so people 30:54always kind of trying to come up with 30:55ways of creating data from scratch 30:58um that they can kind of just generate 30:59on the Fly and use it to effectively 31:01train their models because it kind of 31:02like relases that that bottleneck and so 31:05these researchers out of tens and 31:07Seattle lab said oh well maybe one fun 31:09way of doing this is I can kind of 31:10instantiate what they call personas 31:12which is like a personality could be 31:14like you know your job is as a dog 31:17catcher or your job is as a professional 31:20coder at IBM and their kind of 31:22observation is well we can get these 31:24different personas to do different tasks 31:26to Output data for us and these tasks 31:28will generate very different reactions 31:31right so it turns out if you ask a you 31:32know dog catcher to generate code for 31:34you it will look like you know different 31:36from the code that you have if you 31:37prompt the model to say you're an expert 31:39coder and what they kind of make the 31:41argument for is that with all these 31:43personas we have a scalable way of 31:45generating lots of training data and 31:47they do a couple experiments to show 31:48that you can use the synthetic data to 31:49train you know an llm to do math 31:52problems effectively and um and I guess 31:55you know maybe just to kind of kick it 31:56off you know particularly uh csar with 31:59you on the line you know I think the big 32:01question here is like to nowadays like 32:03how much is compute the bottleneck and 32:05how much is like data the bottleneck um 32:07because it feels like here is the world 32:09where it says okay well if you just have 32:10lots of compute you can generate all the 32:12data that you need but you know a few 32:14episodes ago we were just talking about 32:15how difficult it is to come by compute 32:17and so I'm really kind of interested in 32:19your take on like what is the what is 32:21the bottleneck right now in in the 32:22machine learning workflow yeah that's a 32:24very good question um of course you know 32:27with Gen AI compute is a big ball neck 32:29right now with all the especially the 32:31mutm multiplications that takes a huge 32:35amount uh of comput with the the current 32:38accelerators and Hardware uh data you 32:42know when it comes to data bottlenecks 32:44right now I think in certain industries 32:45I mean whether what is the bot neck 32:47today it depends in certain industries 32:50uh we don't have much data especially in 32:53Industries like industry 4.0 where you 32:55have uh you know uh machines and so on 32:59and you need to understand you know 33:00their operations sometimes there is a 33:02lot of noisy data or you know you have 33:05sensors you know and you probably 33:08haven't collected the data for these 33:10sensors for an extended period of time 33:12to be able to have enough data to train 33:14a good model to understand predict 33:16anomalies or do things like that um so 33:20in certain areas in certain industries 33:22there is this huge lack of data uh when 33:24it comes for example to texts we have an 33:26abundance of text right now online 33:29however you know that text sometimes is 33:31not properly formatted or there's a lot 33:33of noisy and redundancy in the text so I 33:37see you know both of them are 33:39bottlenecks and it depends on maybe the 33:42industry the sector the use cases so 33:44data could be a huge ball neck uh in 33:47case you know you don't have enough data 33:48or you have tons of noisy data and you 33:50need to curate the right the right data 33:53the right context Etc to build the model 33:55and in that case you know the synthetic 33:57data generation could be of huge help of 34:00course comput is still a bottleneck you 34:02know especially right now with the you 34:03know shortages of acceler the hardware 34:05shortage that we have in the 34:07accelerators and you know this R you 34:09know with these large models you know we 34:11have all these comput and of course 34:13we've talked in other episodes about you 34:15know new approaches to the m m free uh 34:18approaches or the in memory Computing 34:21approaches and neomorphic and so on 34:23trying to reduce that bottleneck on the 34:27compute 34:28um so I see both of them are bottlenecks 34:31depending on the you know the context 34:33the US the industry yeah no for sure 34:35yeah there's a reaction I had to the 34:36paper which was basically like well this 34:39just makes all the existing bottlenecks 34:40more bottl necky right it just turns out 34:42like actually you know the great way to 34:44get data is more compute it's like okay 34:46well it's just like more pressure like 34:47people want even more chips exactly it's 34:50like chicken and egg problem here yeah 34:53that's right I guess Skyler maybe I'll 34:54turn to you as someone who has printed 34:56out the paper yes uh although someone 34:59who's printed out many papers and not 35:00read them I don't want to imply that you 35:02have read them but u i mean is this have 35:04we solved the synthetic data problem 35:06like how do you like this approach what 35:07do you think about it I I do like the 35:09approach I think it is it's quite 35:11creative and they they scaled it in a 35:15way that I probably wouldn't have gone 35:17with um I've actually used chat GPT to 35:21write uh bedtime stories for our kids 35:24right there alongside them and uh case 35:26in point here is they play Minecraft and 35:28so they will basically say write a story 35:31but make it about Minecraft so now you 35:33basically have just created a Persona of 35:35a Minecraft player who's responding to 35:37the prompt so we've been doing that as 35:40kind of an individual scale and this 35:42paper has now taken it up to the billion 35:44Persona level and they're keeping track 35:47of all those generated stories in order 35:49to try to get that that diversity so 35:52very cool in that angle uh but I want to 35:55spend a bit of time talking about that 35:56very important word at the end there 35:59diversity how how how confident are we 36:02about the coverage of these 1 billion 36:05personas and do the underlying large 36:08language models you know really 36:11understand being a a massai warrior 36:14Messi is a tribe here in Kenya and so 36:17yes you can ask the large language model 36:19to take on that Persona um whether or 36:22not the generated output from the 36:24Persona of a Messi Warrior matches 36:27reality 36:28that that I don't know how well they 36:30they really covered but but hats off to 36:32the authors for having that idea of 36:34let's take a generated text from all of 36:38these different type of personalities 36:39and actually put a number behind it of a 36:41billion uh very cool um we have not 36:44solved the synthetic data question yet 36:47and I think the the the most obvious 36:48question that comes up on this 36:50is how do we know that those personas 36:53are uh well represented in the 36:56underlying model so yeah that's those 36:58are some of my thoughts on that on that 37:00piece yeah for sure and almost we're 37:02kind of in a like an interesting and I 37:03see Aon you about to come in um you know 37:06like we're almost kind of interesting 37:07place where it's like the only way we 37:08could validate whether or not these 37:10personas are accurate is to have real 37:11world data of these like you know and so 37:14there's kind of this weird chicken egg 37:15issue which is like well I don't know 37:17how validated they are but in order to 37:19validate them we might very well get the 37:20data that we we need um Aaron do you 37:23want to jump in yeah no I you know I 37:25just saw this this a really interesting 37:27you know stat where you know the average 37:29human can read about what is it a 37:31billion words or a million words right 37:33in a year um right and then these 37:35algorithms can read about six orders of 37:38magnitude more in a single month right 37:40they're just thirsty right for this for 37:42this data right and so that projects you 37:44know um I mean don't don't quote me on 37:46this but but around 2030 2032 is we're 37:49going to run out of useful data in many 37:51different domains and you know being 37:53able to synthesize data is very 37:55important uh but if you stratify the 37:58data and in so many ways like in the 38:00paper the danger is are you watering 38:03down you know all the different personas 38:05where where we could glomerate and make 38:07them less like prune them a little bit 38:09because they're not really that much 38:11different right because it's almost 38:12close to the human population on Earth 38:15right of a billion personas right so 38:18that that's that's one area um but then 38:22then I do think that um what would be 38:24interesting to hear about would be um 38:26like the notion of the turning test 2.0 38:28right you know um um how do we ensure 38:31that you know these new personas you 38:34know really do pass you know mustard 38:37right and that that that they actually 38:39do produce you know the transparency we 38:40need explainability train of thought 38:43fairness right um because we because we 38:45are going to be splitting up data in 38:47different ways and and there could be 38:49side effects about that so so I was 38:51curious what folks thought I I think the 38:53idea of um trusted personas or pruning 38:56personas Aon like that you mentioned is 38:59very important can we kind of distill 39:01all of these personas into few ones that 39:03we can trust that give us you know the 39:06best maybe accuracy and Pro in away the 39:10ones because there are you know billions 39:12here we're talking about a large number 39:13of personas another thing how do we tie 39:16this maybe in a real uh problem solving 39:19word scenarios or industrial use cases 39:23and what does that mean for example if 39:25I'm trying maybe to build a a person you 39:27know a foundation model for example for 39:30uh a factory uh you know failure 39:33diagnosis and so on does that is there 39:36Persona there uh is there for example 39:39can we maybe talk about different skills 39:41for example the uh the engineer the uh 39:45maintenance person the uh uh I don't 39:48know the uh you know the chip designer 39:51all of those could be personas here that 39:53maybe we bring together so they can 39:55bring different skills and then uh kind 39:58of collaborate together in this llm or 40:00Foundation model to solve a particular 40:02problem so it's like having multiple 40:05experts working together to solve a 40:07problem so I think the idea here of 40:10taking this maybe to the real world to 40:12solve real problems could be profound 40:14here and have lots of implications I 40:16like the idea the scaling that scholar 40:18mentioned uh the diversity aspect but 40:21this needs to be validated especially 40:23the Sol real or problems so final 40:26thoughts I mean I mean I think it's it's 40:27very exciting you know where where we 40:28are and where we're going you know and 40:30the combination of gen generative AI 40:32techniques with classical techniques is 40:35critical you know creating um I've seen 40:37the term floating around but the AI 40:39sandwich you know where you might use 40:40neuros symbolic you know pieces around 40:43you know these generative AI pieces um 40:46neural networks has been around for a 40:48long time but um but but I think um you 40:51know you know one last thought that I 40:52had is that I think uh you know Mother 40:54Nature is the ultimate teacher and we 40:56have a lot to learn 40:57from our own brains um and I'm I'm 41:00excited about what's next great thank 41:02you well as always there's more to talk 41:04about than we have time for um counter 41:06Skyler Aaron thanks for coming back on 41:08the show um and we'll hopefully Happ you 41:10for a future uh future episode so thanks 41:12for joining us if you enjoyed what you 41:14heard uh reminder as always that you can 41:15get us on Apple podcast Spotify and 41:17podcast platforms everywhere uh and 41:20thanks to you all out in radio land