Shownotes

François background

03:31 -

• I’m François Mondave, a professor at Wake Forest University in the Department of Pediatrics. Despite my department, I’m not a paediatrician—I’m trained in mathematics and AI, which explains my role here.
• I’m part of two centres: one focused on remote health monitoring, where we work with wearables and devices to collect health data and explore ways to improve health, and the other dedicated to AI research. I’ve been working in AI since the early to mid-90s, and that’s where most of my expertise lies.
• I’ve lived in the U.S. for about 24 to 25 years now, with stints in Texas, Florida, and currently, North Carolina. My first triathlon dates back to 1993 in Quebec, where Peter Reid won the race—a name many in the sport would recognise.
• The event was called the Saint-Jean Triathlon, and I’ve been hooked ever since. While I’ve had to step away from racing for the past few years due to having young children, I’ve been immersed in the triathlon world for over 30 years.
• Before the triathlon, I did a fair bit of running and swimming, and I was very active as a child, partly because I was a bit hyperactive.
• One of my most memorable triathlon highlights was in the same year that Lothar Leder broke the eight-hour barrier—a huge milestone at the time—and Peter Reid placed fourth. I vividly remember climbing Solarer Berg, which, if you’re familiar with the race, is not a long hill, maybe half a mile or so.
• The atmosphere was electric, and I had goosebumps all the way up. When I reached the top, I went to shift gears and realised I’d already climbed in the big chainring without noticing because I was so caught up in the excitement.
• It was one of my most incredible moments in the sport, and Roth remains one of my favourite races.
• I’ve also raced in Nice a few times before it became a full-distance race. The last time I was there, Luc Van Lierde won the event. These aren’t recent races, but they are among the most fun and memorable experiences I’ve had in triathlon.

Definition of Artificial Intelligence

09:05 -

• When we talk about artificial intelligence, it’s essential to first define what intelligence really is. At its core, intelligence is the ability to collect, analyse, and use information to learn, make decisions, and solve problems more effectively and efficiently.
• This broad definition helps frame the discussion around AI, as we are essentially building systems that replicate these abilities in some capacity.
• However, human intelligence is nuanced, and we don’t yet fully understand or agree on a formal definition of reasoning. Intelligence also has a trainable component—much like VO2 max in athletes. The brain has plasticity, which means aspects of intelligence can be developed through experience and learning.
• When we discuss AI, we are talking about building computing systems that process data and help us make decisions or automate tasks that would typically require some level of human intelligence.
• The systems differ depending on the type of data they process and the training they receive, similar to how human intelligence varies across different domains like reasoning with numbers, images, or language.
• AI’s origins go back to John McCarthy coining the term in 1956. However, the underlying principles of reasoning can be traced as far back as ancient Greece, over 2,000 years ago, when early forms of logic and reasoning were defined.
• The first practical AI tools were expert systems, which emerged decades later. Expert systems rely on knowledge bases containing factual information and inference mechanisms—using Boolean logic (and, or, implication, equivalence) to deduce new knowledge from known facts. These systems, while relatively simple, laid the foundation for modern AI.
• In AI, we typically discuss four paradigms: thinking like humans, acting like humans, thinking logically, and acting logically. Not all of these overlap, as humans don’t always reason logically. However, modern AI focuses largely on optimisation—mimicking processes to achieve specific outcomes.
• Machine learning, a major subfield of AI, is divided into different categories based on how the system is trained. Supervised learning involves labelled data, where the system learns to associate data with its correct label, such as classifying an image or predicting a category.
• Unsupervised learning, on the other hand, deals with unlabelled data and looks for patterns or clusters of similar data points in multi-dimensional space. There is also semi-supervised learning, where some data is labelled while other data is not. This hybrid method is used when labels are expensive or time-consuming to generate.
• Other strategies include reinforcement learning, where systems learn through trial and error based on rewards, and deep learning, which is a key technology behind many of today’s advanced AI systems. Deep learning powers large language models (LLMs), which represent a pinnacle of modern AI technology.
• These models are built using deep neural networks and variations of deep learning techniques, enabling them to understand and generate human-like text. While the technical details can get complex, the fundamental idea is that these systems are designed to learn from vast amounts of data and continuously improve their predictions or outputs.

Training AI models

18:39 -

• One of the oldest machine learning models is the artificial neural network (ANN), which mimics how the brain works—or at least how we initially believed it did.
• The concept dates back to 1949 when a paper on the perceptron, the simplest form of an ANN, was published. The perceptron works by taking labelled data and using it to learn how to classify new data.
• Let’s walk through a simplified example of how a neural network can be used for binary classification, such as predicting outcomes for patients after surgery.
• Imagine we have data from 20,000 patients who underwent surgery—10,000 with good outcomes and 10,000 with poor outcomes. Additionally, we set aside data from another 5,000 patients as a testing set that we’ll use later to evaluate our model.
• To train the ANN, we feed it data about each patient. The data may include variables such as age, sex, BMI, and the severity of the condition.
• The network adjusts its internal connections—or “weights”—to learn how to map these inputs to the correct outcome (good or bad).
• Training involves progressively adjusting the weights until the model correctly classifies as many patients as possible. Once the model is trained on the initial 20,000 patients, we can then test its performance on the 5,000 patients it hasn’t seen before.
• This is a critical step because using independent data helps ensure that the model hasn’t simply memorized the training data, a problem known as overfitting.
• Overfitting occurs when the model becomes too tailored to the training data, which causes it to make poor predictions on new data.
• By testing with unseen data, we can better assess whether the model has learned generalizable patterns that will work in real-world scenarios.
• In the past, limited computational resources meant we had to be careful with how much data and how many variables were included in a model.
• Today, with vast improvements in storage and processing power, we can use much larger datasets without worrying about running out of resources. This opens up more opportunities to consider additional variables and improve the overall accuracy of the predictions.

Strengths and weaknesses of AI models

23:57 -

• The strengths and weaknesses of AI depend greatly on the field of application. In healthcare—particularly in areas relevant to endurance sports and triathlon—AI excels at processing large, heterogeneous datasets.
• This capability enables it to analyze vast amounts of data from sources like electronic health records (EHRs), medical images (MRIs, X-rays), and even ECGs. For example, AI tools can help detect atrial fibrillation (AFib), a common issue among endurance athletes, by analyzing short ECG readings and predicting when irregular heart rhythms might develop.
• This predictive power allows for better care and proactive treatment. Similarly, AI can predict surgical success rates or identify the likelihood of conditions like sepsis by analyzing hundreds of thousands of variables across large patient populations.
• However, deep learning models—while powerful—bring significant challenges. One of the biggest is interpretability.
• In healthcare, it’s critical to explain why an AI system made a specific recommendation, especially when it informs decisions like whether a patient should undergo surgery. With deep neural networks, which have many hidden layers of interconnected neurons, this explainability diminishes as the network’s complexity increases.
• Physicians need to understand and justify these AI-driven recommendations to patients, but deep learning’s “black box” nature makes that difficult. Without clear reasoning behind a decision, trust in the system and its recommendations erodes.
• This issue is not unique to healthcare. A well-known example is AlphaGo, the deep learning system that became the best player of the ancient board game Go.
• AlphaGo didn’t just play better than humans—it developed new strategies that no one understood. While impressive, it failed as a teaching tool because its decisions were not explainable, making it difficult to learn from or apply its strategies. 
• The same challenge exists in other fields where decision-making is critical: if an AI system performs perfectly but no one understands how or why, it can’t be trusted in high-stakes scenarios.
• Another challenge is generalizability. AI systems are often trained on institution-specific or geographically isolated datasets.
• When these models are used in a different institution or on a different patient population, performance can drop because the model wasn’t exposed to enough diversity in its training data. 
• This makes it difficult to replicate the same success across different settings, even though it’s a common issue in traditional statistical methods as well.
• The balance between accuracy and interpretability is critical. If AI could consistently provide perfect predictions, explainability might be less important.
• However, AI models are not perfect, and when outcomes are uncertain or controversial, explainability becomes essential—especially when decisions could result in severe consequences or side effects.
• This also ties into broader ethical concerns around AI, which will undoubtedly be a growing part of the discussion moving forward.

AI applied to triathlon

32:43 -

• I’m currently working with a coach to adapt my training, but I don’t really use AI tools. The only exposure I have is Strava’s beta-testing AI-generated workout comments.
• While I like Strava as a platform, the feedback from this AI has been hit-or-miss. Sometimes it does a decent job, but the data window it pulls from seems narrow, and the comments depend heavily on how I describe my workout. 
• It’s essentially a closed box—I don’t know how the comments are generated or what factors they weigh. Others may have better experiences, but it’s clear Strava is still in its experimental phase with this feature.
• When it comes to AI in training, I haven’t found any predictive modelling tools that seem truly useful. There are platforms that allow you to generate customised large-language models, like using subsets of ChatGPT, which is intriguing.
• Still, I haven’t seen anything that effectively uses predictive AI for endurance training. Academic environments often share their data and models for validation, but industry settings typically don’t reveal how their products work internally. You just get a black-box product that completes a task.
• In endurance sports, especially triathlon, data collection has advanced considerably. There’s now the ability to measure things like core temperature, sweat rates, heart rates, and lactate more accurately than ever before.
• The Norwegians are a prime example of using extensive data in their training, and many in the triathlon community know about their approach.
• While there’s now an unprecedented amount of data being collected, I haven’t seen AI tools that can interpret it comprehensively. There’s no tool that can assess all your data and provide actionable feedback—like recommending a shift from threshold training to anaerobic capacity if your goal is to increase race speed.
• This type of predictive, decision-making AI tool for triathlon training doesn’t exist yet, at least not in a form that’s widely accessible or reliable.

Challenges for applying AI for training

38:17 -

• In athletic performance improvement, especially for experienced athletes, the goal is typically small, incremental gains—like a 2% improvement.
• This contrasts with healthcare, where interventions are often about significant changes, such as curing illness or injury. The challenge with advanced athletes is that the margin for improvement is very narrow, and the methods used for average athletes won’t necessarily work for those at the elite level.
• In healthcare, AI models are being developed to predict conditions like sepsis or differentiate between diseases such as endometriosis and adenomyosis.
• However, despite these advancements, clinical adoption of AI is still minimal. Many doctors prefer traditional methods because AI models aren’t always practical in critical care settings.
• For instance, sepsis prediction models may exist, but critical care doctors still rely on antibiotics and fluids.
• When it comes to athletes, the data challenges are significant. Wearable devices, often not of clinical quality, produce data that is sometimes unreliable.
• Chest strap heart rate monitors, for example, can produce artefacts depending on how tightly they’re worn. Wrist-based heart rate monitors can produce inaccurate data if the device moves too much.
• Since models are only as good as the data they’re trained on, poor-quality data will produce inaccurate results.
• AI models for athletes also need to consider the type of athlete being targeted. Average age-group athletes can benefit from generalised models, but elite athletes at the top of their sport need more tailored tools.
• Unfortunately, many tools are proprietary, making it difficult to assess how they are built or trained.
• At present, there are no reliable AI models that can precisely guide an athlete’s training for a specific event. However, companies developing these tools often have good intentions and aim to build effective models.
• The challenge is that industry tools are not held to the same rigorous standards as academic research. In the US, there is growing interest in combining the strengths of academic research with the agility of industry to create better-quality models. This collaboration could result in high-quality data and models that meaningfully improve athletic training.
• The rapid advancements in AI offer hope. New wearable devices regularly emerge, capturing more types of data, such as single-lead ECGs capable of detecting conditions like AFib and Parkinson’s.
• Given the speed of AI development, it seems likely that, in the near future, we’ll have tools that can gather quality data and make accurate endurance training predictions, such as how to improve VO2 max or threshold performance.

Building a model for triathlon training

47:50 -

• If someone is interested in collaborating on this, I’d be eager to get involved. There are several key aspects to consider. First, on the technical and hardware side, it's essential to develop tools that are reliable and valid—devices that collect high-quality data, accurately measure what they’re designed to measure, and ensure consistency across every data collection session.
• In an ideal world, I’d have a diverse team evaluating various devices, including those that track sweat rates, core temperature, and lactate levels.
• Tracking lactates, for instance, isn’t a new concept—Jan Olbrecht worked on this back in the 1990s, and even before that, studies from the mid-1980s explored lactate monitoring. What’s revolutionary now is how easy it has become to collect and interpret these data points.
• We could potentially track running economy and oxygen consumption in real-time, much like how we already monitor power output with cycling power meters.
• There are also tools to monitor body position on the bike in real-time, which can help adjust for aerodynamics and reduce wind resistance.
• My vision would be to build a robust “athletic data warehouse,” much like a clinical data warehouse in the medical world, where we consolidate diverse and heterogeneous data sets.
• These would include data from a wide range of athletes—from beginners and competitive age-group athletes to elites. The goal would be to create adaptive models that cater to various athlete types and goals.
• These models would dynamically adjust training programs daily. For example, if your plan originally calls for a VO2 max session tomorrow, but today’s performance data indicates you didn’t recover well, the system would automatically modify the plan to suggest a lighter, easier session instead.
• It’s about creating a system that constantly tweaks your training to optimise performance day by day.

How can a model predict which training methodology will bring more improvements for a specific athlete

51:46 -

• There are parallels between healthcare and athletic performance, particularly with how data from large datasets can be used to make specific decisions for individual patients or athletes.
• There are real-time tools, such as ECG monitoring during workouts, as an example of how technology could inform immediate training adjustments based on the body's responses.
• For instance, if an athlete's heart rate or vagal tone seems off, AI could signal the need to modify or stop a workout to prevent strain.
• Drawing from personal experience, I recall training at a sports centre in Australia, where they had the opportunity to watch world-class athletes, such as Alex Popov, train. His coach noticed a swimmer's stroke breaking down, despite it being undetectable to the untrained eye.
• This prompted the coach to adjust the swimmer's routine, something that could be assisted by AI if sensors could track water pressure and stroke quality in real-time. In the coming years, such technology will allow athletes to tweak their performance mid-session based on sensor data.
• However, not everything can be measured. For example, the mental aspects of training, which I personally struggle with, cannot yet be captured by wearables.
• Mark Allen’s famous advice to ignore data and race with intuition, illustrates how sometimes athletes need to disconnect from constant data monitoring to focus on their internal cues. 

AI and the complexity of the sport

59:40 -

• Triathlon, particularly at the elite level, is often seen as an individual sport, excluding certain team-based events such as Olympic-distance races.
• While the sport itself may not appear overly complex, there is a growing level of sophistication, especially as athletes strive to gain marginal gains—those tiny 0.2% to 0.3% improvements that can make a significant difference at the top level.
• This drive for better performance is evident in long-distance triathlons, where the intensity and speed have drastically increased in recent years.
• Nutrition also plays a crucial role in triathlon performance. However, when it comes to AI applications in nutrition, it's recognized as a very challenging field. Despite advancements in various sectors, including healthcare, improving people's eating habits has been notoriously difficult. While triathlon may not be as complicated as some other sports in terms of data, if we have access to high-quality data, there is still room for more efficiency and insight.
• Drawing parallels with American football, the interviewee highlights the complexity of that sport due to real-time data analysis. In football, cloud services offer predictive analytics on the likelihood of successful passes, player actions, and other game factors.
• The dynamic nature of football—players constantly changing depending on the situation—means that real-time insights are vital.
• Even though triathlon doesn't involve capturing real-time metrics such as lactate or heart rate in the same way, there is still room for similar data collection, such as tracking speed or changes in direction to inform performance improvements.
• However, triathlon, being an individual sport, is inherently simpler than team-based sports in this respect. Despite this, the data analysis tools available for triathlons could soon improve, potentially adding a layer of strategy similar to that in football.

Rapid fire questions about AI in endurance sports

1:02:30 -

Training Planning (Short-term and Long-term):

• Rating: 5
• AI has a high potential to revolutionise training planning with enough resources. It can adapt training schedules to an athlete's individual needs, making it highly effective for both short-term and long-term planning.

Strategic Planning (Long-term Athlete Development, Race Selection):

• Rating: 3
• While AI could contribute to strategic planning, it is a more complex area. The variation in athlete levels and training approaches makes it challenging to rely solely on AI for long-term decisions. The timeframes for planning are often narrow, adding another layer of complexity.

Race Planning:

• Rating: 5
• AI can be extremely useful for race planning, particularly by analysing past performances, weather, course characteristics, and other variables. It has the potential to provide insights that guide both immediate and future race-day decisions.

Injury Prevention:

• Rating: 2-3
• While AI holds promise in injury prevention, it’s still a long way from being able to predict injuries accurately. The complexity of the data required, like biomechanics and training load, makes this a challenging area.

Nutrition and Hydration:

• Rating: 3 for Nutrition, 4 for Hydration
• AI could be very useful for hydration strategies, considering real-time data and environmental factors. For nutrition, the complexity of individual needs, preferences, and metabolic responses makes it more difficult to rely on AI with the same level of confidence.

Mental Strategies and Psychology:

• Rating: 1-2
• AI has limited potential in this area. While it may offer some basic tools or suggestions, the complexities of mental health and psychological strategies are difficult for algorithms to address, as they involve emotional and personal nuances that AI cannot easily quantify.

Equipment Choices and Optimization:

• Rating: 5
• AI is already highly effective in optimising equipment choices. Analysing large datasets can help tailor gear, such as bikes, shoes, and clothing, for maximum performance based on an athlete’s specific needs and characteristics.

Ethical and Privacy Considerations:

• It’s important to remember the ethical concerns surrounding AI in sports, particularly regarding the privacy of data, such as GPS data from training sessions. Athletes should be mindful of when and how they share this data to ensure it is protected and used responsibly.

Risks of implementing AI tools in their planning

1:06:03 -

• The biggest risk, in my opinion, is the lack of a clear understanding of data quality. This is not a unique issue to sports or even to technology, but it’s something we see in healthcare as well.
• In healthcare, we have clinical-grade tools, but not all of the data we collect meets that standard of quality, often due to issues with workflows or other factors in the clinical environment.
• This difference in data quality is something that also concerns me in the sports field.
• Another significant risk is becoming too reliant on technology, particularly AI but also in a broader sense. I think we’ve all encountered moments where our equipment fails.
• For example, as a coach, you’ll know how frustrating it is when you get on the bike and realise your power meter isn’t charged. In a race like Ironman, having a clear plan in mind is important, but if your data is gone, it can create a lot of uncertainty.
• Many beginner athletes would panic in such a situation, fearing they could end up overbiking—something that’s a real issue, especially when you consider that overbiking is one of the fastest ways to ruin an Ironman race. It’s a reminder that while technology can be helpful, there has to be a balance.
• Too much reliance on it can lead to serious issues, and that’s something I always stress to athletes: to not let technology dictate everything but to use it as a helpful tool, not a crutch.

How self-coached athletes can use AI

1:08:01 -

• I believe the future of training platforms lies in centralising data and embedding predictive tools. While some current platforms share insights on training load, the data isn’t always science-based, raising questions about how performance curves are generated.
• However, soon we’ll see tools that analyse progress in real-time—metrics like running economy and how it interacts with VO2 max or high VO2 max trade-offs for long-distance events.
• This "multi-objective optimisation" will help optimise various performance pathways on a daily basis, enabling seamless information sharing between athletes and coaches through cloud platforms.
• When it comes to sports science, AI will play a transformative role. Current research is often limited by small sample sizes—typically small cohorts of around 12 athletes per group in strength or running economy studies.
• AI will allow us to analyse longitudinal data across much larger athlete populations. This will not only improve models but also filter out ineffective strategies. I also see great potential for AI to be integrated into injury prevention.
• The cycle of injury and recovery often limits progress, and it’s hard to gauge whether athletes are progressing too quickly. AI could help mitigate these risks, offering better insights to avoid overtraining or injury.
• AI tools are already being used to some extent, especially in studies leveraging large data sets, such as Strava data for runners.
• With access to data on a much larger scale, I believe AI can revolutionise performance monitoring and injury prevention strategies in sports.

Ethical considerations regarding the use of AI

1:12:13 -

• In the US healthcare system, data use in triathlon and other sports is very different from healthcare. The healthcare system in the US is profit-driven, especially at the insurance level, which creates ethical concerns.
• For example, imagine a platform like Strava collecting real-time ECG data. If that data flags an irregularity and the insurance company uses it to deny care, that’s a real possibility here. This scenario might shock European audiences, but in the US, we face these challenges regularly.
• Some insurance companies are already using AI to determine whether treatments should be covered, which is alarming because many of these tools aren’t sufficiently advanced to make such decisions.
• Physicians often share stories on social media about their patients being denied care by insurance companies based on AI decisions, even though the doctor confirmed treatment was necessary. While this issue is more pressing in the US due to its healthcare structure, it’s something we must navigate.
• From a community standpoint, the primary concerns revolve around privacy, such as revealing someone’s location. However, there’s also the question of liability when using AI tools in coaching.
• Suppose I create an AI tool for coaches, and after six months, you report that three of your athletes developed Achilles tendonitis, which they’d never experienced before.
• This raises concerns about whether the AI ramped up their training volume too quickly, potentially leading to injuries.
• In Europe, litigation isn’t as prevalent as in the US, but in the US, such a situation could lead to lawsuits. Questions then arise about who is liable: the tool developer or the coach who used the tool?
• Additionally, there’s the risk of AI systems missing critical information, such as a significant ECG irregularity, which could have severe consequences.
• These ethical and legal questions need careful consideration, especially as AI use continues to grow in these contexts.

AI bias

1:15:40 -

• When considering socio-economic status and gender, there are clear differences in how men and women respond to training.
• Generally, women are easier to motivate and tend to find internal, broader motivations more compelling, while men often seek more specific, comparative goals.
• For example, many men actively engage with Strava, focusing on likes and comparing their performances. In contrast, my wife, who runs, doesn’t care about tracking achievements like the queen of the mountain titles on Strava at all.
• Bias plays a significant role in this dynamic. One issue with wearable devices is that wrist-based heart rate measurements tend to be less accurate for athletes with darker skin tones compared to lighter skin. This discrepancy is rooted in how these tools are designed and the populations used in data training.
• If AI tools are not trained on diverse populations representative of their actual users, predictive models will generate unsuitable recommendations. This mirrors issues we see in healthcare, where a lack of inclusivity in data leads to disparities in treatment and outcomes.

AI business monopolization

1:18:07 -

• Cloud services dominate the digital ecosystem, with Amazon Web Services (AWS) at its core. While many people have recently expressed dissatisfaction with Amazon and are cancelling services like Amazon Prime, they may not realise how widespread AWS’s influence is.
• It powers platforms like Netflix, Disney Plus, and even services from other companies, including Google. AWS is the backbone for many of these platforms, creating a near-monopoly in cloud hosting, particularly with concerns around privacy and potential power plays.
• In terms of devices, there are companies that produce excellent hardware, but some fall short with their web platforms.
• Although I believe technology is largely democratised, hosting services are a different matter, with one company holding an undeniable dominance.
• While I am passionate about the scientific side of things—how data is used, how models are built, and how improvements are made—the business side is not my area of expertise. I can’t predict the market’s capacity or future challenges like buyouts or barriers to entry.
• For those insights, I would defer to my colleagues who are more experienced in navigating those aspects. My focus remains on advancing the science, not managing the business side of innovation.

General trends in AI

1:21:21 -

• Large language models have revolutionised how we approach healthcare data, particularly with clinical notes. Roughly 80% of patient information in electronic health records comes from these unstructured notes that doctors input during consultations.
• Although some areas, such as pathology, may have more structured data, most of it is just plain text.
• Thanks to the rapid evolution of large language models over the past few years, we now have powerful tools to analyse this unstructured information effectively.
• This capability opens exciting possibilities for healthcare. We can now extract meaningful insights from notes written in natural language, offering a deeper understanding of patient conditions and enhancing clinical decision-making.
• Another fascinating development involves protein folding. Understanding how proteins fold has been a critical challenge in drug development.
• Proteins fold naturally to minimise energy, but predicting these 3D structures from amino acid sequences has historically been computationally intensive. A few groups, like one at the University of Washington, made strides in this area, but deep learning has now transformed these predictions.
• With incredible accuracy, we can now anticipate protein folding, which has enormous implications for developing new drugs and drug delivery mechanisms.
• Although we’re still a few years away from fully leveraging these protein-folding breakthroughs, the future of drug development looks incredibly promising.
• Combining these advancements with more integrated healthcare data and training platforms could lead to a more personalised and effective approach to patient care.
• Existing platforms tend to be quantitative, but large language models could unlock the qualitative insights hidden within clinical notes and comments, potentially transforming how care is delivered and improved over time.

AI resources

1:24:48 -

• If your maths knowledge isn’t too rusty, there are some excellent online resources. One blog I recommend is Towards Data Science—it offers great information, though it can be quite maths-heavy. However, if you have no maths background at all, it’s a lot more challenging to find suitable materials.
• Honestly, I’m not the best person to suggest resources for that level, as nothing comes to mind. That said, Crash Course AI on YouTube is a solid option. It’s quite basic and covers the foundational concepts pretty well.

Pieces of advice for the use of AI

1:26:10 -

• AI is already revolutionising clinical care, though not in the scary way some might imagine. Instead of replacing doctors, it will likely enhance how we do our jobs.
• In the near future, I see AI helping physicians focus more on patient interaction rather than getting bogged down by analysing all the available patient data.
• It will allow us to hone in on what makes us human, such as empathy and observation, while AI handles the heavy lifting of data analysis.
• Similarly, for triathlon coaches, AI could take over the quantitative analysis of an athlete’s performance. This would free coaches to focus on the subtle cues that data might miss—things like a slight change in expression that hints at a mental block or fatigue.
• These nuanced observations are inherently human and cannot be replicated by AI.
• That said, I believe we’ve overhyped AI somewhat. Many issues still need addressing, such as the quality of the data used and the transparency of the models in industries like healthcare. In the business world, openness is more limited compared to research settings.
• While AI is exciting, we should remain cautiously optimistic rather than blindly embracing it. Its real value lies in its ability to manage vast amounts of data and allow humans to concentrate on tasks AI cannot replicate.
• In coaching, this shift could take us back to an era where coaches focused less on spreadsheets and more on their athletes. Having AI handle the metrics would allow coaches to build deeper, more personal connections with their athletes.
• Similarly, a famous medical adage advises doctors not to just understand what disease their patient has but to understand what patient has the disease. The same principle applies to athletics: using AI as a tool grants more time and freedom to focus on the person behind the performance.

LINKS AND RESOURCES:

• François' LinkedIn, ResearchGate and Instagram
• Towards Data Science - "The world’s leading publication for data science, AI, and ML professionals."
• Crash Course AI - very simple explanations of AI for laymen without prerequisite knowledge