![](jpg/ai-site-banner-2.jpg)
![](png/icon-football-helmet-16x16.png)
![](png/icon-ball-of-basketball-16x16.png)
![](png/cricketwidget-01-16x16.png)
![](png/icon-soccer-ball-16x16.png)
Welcome to the newest edition of Stats Perform’s Latest Trends in AI in Sport series, in which our Chief Scientist, Patrick Lucey discusses recent developments in Generative AI and their applications in the sports industry.
For those in the world of AI and sports, February’s Super Bowl served up a treat both on and off the field.
On the field and on our screens, we saw players using plays devised with the assistance of machine learning by their team analysts, and we heard announcers giving us AI-powered insights to help us understand context and meaning.
Off the field, the Super Bowl featured advertisements focusing on AI as a feature in mainstream consumer products, like Microsoft’s Copilot ad, and Google using AI to showcase its power in their Pixel phone.
In the short period before and after the Super Bowl, we’ve seen headline releases such as Google’s Gemini, Apple’s full Vision Pro launch and latterly OpenAI’s recent release of their multimodal text2video tool “Sora”, as well as rumors about the release of AI agents and increased utilization of Retrieval-Augmented Generation (RAG).
In this article we will process some of these developments in the world of Generative AI and explain RAG and multimodal LLMs. We’ll outline how they relate to us in sport (both for fans and for professional team applications), and discuss how despite the new changes, one thing remains the same: the value of the underlying data powering GenAI models.
In summary: 12+ Months on from ChatGPT
What we are seeing a year or so after ChatGPT launched is that Generative AI technology isn’t replacing humans at their jobs, as was feared in some circles, but enabling humans to do more.
ChatGPT and the subsequent LLM-based technologies that have followed have given us a range of “AI-assistive” tools that can help some of us do some of our tasks a bit better and quicker.
For example, customer service, language translation, searching and interrogating documentation (whether that is public or private) are all starting to be done much more efficiently and at greater scale than before. It can also enable create new content, whether that is textual, audio or in images and video (or a combination).
Ultimately, the progress in these areas (and what has driven Copilot, Gemini and Sora) is because GenAI maximizes our ability to extract value from data. The underlying data is still the most important element in the GenAI ecosystem.
The Third Wave of Generative AI: Differentiated Data
Vista Equity Partners Chairman and CEO, Robert Smith recently gave an excellent summary of what he saw as “The 3 Waves of Generative AI”.
In the first wave, he described that the enduring value would go to the hardware providers (e.g., NVIDIA).
In the second wave, he said that the enduring value would go to the super scalers (e.g., cloud providers).
And in the third wave, he said that the enduring value would benefit those who can utilize that capacity in the specific markets which they sell their products. In other words, if you have a key differentiation in a particular domain (such as uniquely deep and broad data in sport), generative AI can “super boost” products and services in that field.
We are seeing the first two waves play out currently.
The first is benefitting the chip manufacturers, the thirst for chips that can be used for GenAI seemingly unquenchable. This has flowed into the second wave, where cloud providers have the necessary storage and compute to scale out GenAI solutions and bundle them into their general solutions, which is also happening now.
But the long-term enduring value will lie with the third wave, where specific markets need solutions tailored towards their specific use-cases.
Why is data still so important for GenERATIVE AI aPPLICATIONS in sportS?
It’s worth remembering that to get the most accurate outputs, Generative AI systems require the most accurate, reliable and uniform data inputs. Without quality inputs, erroneous outputs or misinformation can be propagated, or the applications of the AI are severely limited, particularly when it comes to use in specific fields.
So even though it may seem that things are evolving and changing, at their core, everything is still the same, as all information and insights are predicated on reliable and accurate data.
In specific fields like sports, or medicine, with specific use cases like player analysis and performance predictions, or symptom diagnosis and treatment evaluation, you cannot get facts wrong. If you do, downstream analysis, predictions and diagnosis are also going to be wrong.
Additionally, the more detailed and consistent the data, the more specific, relevant and useful the outputs and responses.
For example, in soccer where Stats Perform has detailed Opta data for over 13,500 professional men’s and women’s teams, we ensure the same data and measurements are generated across all games (not just a select few leagues).
As such, a pass and a shot each have the same uniform definition and collection across all leagues in soccer. Every event uses the same protocols for metadata like names, timestamps and location co-ordinates.
The uniform classification and structure of the data – collected in real-time – means Generative AI models can be trained on massive amounts of data, both historical data and in-play statistics.
This improves their efficacy as well as enhancing search capabilities and making it easier to identify correlations and patterns in the dataset.
If they’re not uniformly collected and distributed, the stats and derived outputs are meaningless to both humans and GenAI models.
Generative AI helps us uncover valuable insights previously buried or unseen by the human eye alone. It helps maximize the value of the underlying data. If that data is detailed, uniform and objective, its value can be heightened even further.
The deeper the source dataset, the more insightful the patterns and more accurate the predictions produced by GenAI, and the greater the utility for specific applications and fields, like sports.
The data’s quality, depth, breadth and consistency are the key ingredients for AI systems (and specifically GenAI systems) to flourish.
On another note, the rise in media coverage of deepfakes potentially influencing elections has had me also considering the value of independent, trusted sports data to combat or counter the potential for nefarious uses of deepfake video that could threaten the integrity of sport.
Some of you may recall a deep-fake “alternative history” of the famous Michael Jordan shot over Craig Ehlo, that was posted on social media a couple of years ago. It shows Jordan missing the game-winning shot that he actually made, fooling many observers. As text2video LLMs get ever more effective, the likelihood of deepfakes infiltrating sport increases.
It’s another reason we take Stats Perform’s role as sport’s trusted record-keepers incredibly seriously, and are grateful for our amazing Data Operations team who focus night and day on the integrity of our data, because trusted, independent and verifiable sources like Opta’s data are likely to be even more critical in a deepfake future.
What is RAG?
Retrieval-Augmented Generation (aka RAG) is another relatively new development in the Generative AI field that enables “non-technical experts” to be able to retrieve and interrogate knowledge bases, without relying on an engineer or database expert to fetch that data, which is otherwise a major blocker.
If it has a foundation of solid and accurate data, RAG can then wrap that information around current LLM-based technology.
This GenAI technology can enable people to do more, and enable quicker iterations and obtain knowledge faster. In some ways it is the ultimate assistant – like having an expert permanently on your shoulder giving you the answers (if the system has access to the most accurate and up-to-date data).
Wherever you have an extensive knowledge base of information, using GenAI can enable retrieval and interrogation of sports information and RAG can make this easier for systems that need to access domain-specific knowledge.
Again, an abundance of caution is still required, however, as hallucination and out-of-date information can be propagated, meaning the quality of and trust in the underlying source data is imperative.
But if used correctly, with guardrails and oversight, RAG could be of tremendous value in personalizing content and treatments in the world of sport.
In a future article, we will do a deeper dive on this topic.
Developments using video, images and multimodal LLMs
The language of sports is both the on-ball data and off-ball movement data collected. We call this multimodal, as it captures two modes of information which are complimentary – like images or video with a text caption to describe what is occurring (which is how GPT-4, Gemini and Sora can do the text2image and text2video generation).
One way to utilize video effectively in LLMs for sports is to capture and combine both off- and on-ball player and ball data, as this enables a full reconstruction of complete performance to be measured, as well as interactive applications to be built.
In a recent advancement, we have been able to use our vast Opta database and a multimodal LLM to estimate the missing player locations that cannot be seen on the video (either partially or fully), which enables every pass and decision to be analyzed in full context of where other players are positioned and what they’re doing.
We just showcased this advanced application of Generative AI in sports at the 2024 MIT Sloan Sports Analytics Conference. But it’s very much just the beginning – from this foundation model are a great many additional applications. Expect many more updates from us in this space.
How else coulD recent Generative AI advancements relate to professional sportS analysis?
Sports-specific LLMs can also be used to create predictions which couldn’t be modeled accurately before, giving team analysts and coaches a very powerful assistant.
For example, modelling the individual player statistics of a soccer player is notoriously difficult, because it depends on which team-mates he or she is playing with, as well as which opponents, as well as the game-state for live predictions (e.g., is the team winning/losing, is a player sent-off, etc.).
The same Generative AI methods used to train the models that power ChatGPT and Gemini, (i.e., transformer neural networks) can however be used to generate predictions for every player (and the team) from the same model, which is a paradigm shift in how predictions in sports can be performed
The power of this approach is not only providing more accurate predictions, but enabling new applications such as a “virtual assistant coach,” where a user could ask “if I take player X off and bring player Y on – how could that change the outcome of the match, or lead to more or less shots?”. Or, another question could be “where should have the defender been in that situation to stop the shot?”, or “where is the player going to hit the ball next.”
This could be a new way of interacting with sports that was once only confined to a video game, and for that last example, we did that exact thing with our great partners at Tennis Australia during the last Australian Open back in January 2024. You can see it in action HERE.
As with the other AI assistants, sports analysts and coaches will be empowered to develop ever-more sophisticated strategies and tactics, which will result in even more on-field “wow-moments” for watching fans.
How do sports data and LLMs play into the advancements of Mixed Reality (i.e., Apple Vision Pro)?
Watching and interacting with sports plays a prominent part in the Apple Vision Pro release video.
In addition to watching the game from multiple angles, users are shown to have the ability to have all the data and insights they want within their viewing ecosystem, whether that is via an audio query or gesture. They can also interact with modules to consume adjacent content like highlights, other games, chatting with friends or seeing punditry (or contextual advertising).
The ambition seems to be creating a 360-degree experience that ultimately makes every detail of the game more interesting and entertaining, bonding the fan closer to the players, teams, ‘broadcaster’ and sponsors.
Quality sports data underpins many of these experiences in both visible and behind-the-scenes ways.
Among those behind-the-scenes ways would be LLMs powering predictions; curating the best content and advertising to show at the best time, including highlights; personalizing what users see, hear and feel based on game state as well as preference; powering gamification opportunities; adding incredible layers of context to make athlete skills even more memorable, and overall deepening the event’s storytelling power and hence its value to fans and sponsors.
Not to mention the tactical insights it will give to analysts and coaches, lifting the on-field entertainment to even greater heights!
Sloan Sports Analytics Conference 2024
To conclude this update, I’ve just returned from another fascinating couple of days at the Sloan Sports Analytics Conference. As mentioned earlier, we presented a paper there on how Generative AI can be leveraged for a specific application in sports, given high quality source data, being to predict accurate locations of soccer players not visually perceived in broadcasts (link below). Other panels and presentations discussed broad GenAI applications in tasks for sports organisations like customer service and search. We’ll touch on these and other specific uses in future articles.
Patrick Lucey is the Chief Scientist at Stats Perform. You can read his other opinions and insights on the role ChatGPT and large language models play in sport, and what to watch for next.
Coming soon in Patrick’s articles: he’ll discuss Stats Perform’s 2024 paper at the Sloan Sports Analytics Conference in more detail, whether Generative AI could improve the application of VAR (Video Assistant Referee) in sport, whether AI Agents will be useful in sport and applications of Generative AI to enhance sports broadcasts and sponsorships. He’ll also take a fresh look at the original goal of RoboCup, which is to enable fully autonomous humanoid robots to beat the best human soccer team in the world, on a real outdoor field, by the year 2050!