Note: This is written entirely by me, and I did not use ChatGPT or any other form of AI in creating what’s below. Of course, how would you know if I did or didn’t, right? I’ll tell you how: AI, by definition, is “intelligent.” What I write… well, that’s up to you to decide how intelligent it is!
I admit that I know very little about the technical details of how AI works, but I did play the part of an AI expert on TV once… and I stayed in a Holiday Inn Express last night (one of those things is far more likely to be true than the other!).
But over the past ten months, I’ve read a lot about ChatGPT, LLMs, and all sorts of other things that are often referred to by TLAs. If you’re not familiar with TLA, it means Three Letter Acronym, and it seems that the tech world competes with the military to see who can use them the most.
And if you’re not familiar with LLM, here’s what Wikipedia has to say about them:
“A large language model (LLM) is a language model characterized by emergent properties enabled by its large size. They are built with artificial neural networks, (pre-)trained using self-supervised learning and semi-supervised learning, typically containing tens of millions to billions of weights.”
Okay….
What does this have to do with driving and motorsport? That’s what I asked myself back in December of 2022, just after I’d been exposed to ChatGPT during a dinner with some friends in Indianapolis. I came home from Indy wondering if I had a future in our sport!
Being a non-recovering learning addict, I decided to do what I usually do when introduced to something new: read. One of the best books I read is Impromptu: Amplifying Our Humanity Through AI, by Reid Hoffman (co-founder of LinkedIn, and investor in OpenAI) and his co-author, ChatGPT. Yes, Hoffman used ChatGPT to help write a book about ChatGPT. Kinda cool, kinda creepy, kinda meta. But I have to tell you, when someone knows how to use ChatGPT (for good, like Hoffman does), it really does a fantastic job. And no, it’s not going to replace human authors, but it can help them with research and even idea generation.
To address the idea of AI replacing all of us, the best quote I read was, “AI will not replace people, but people using AI will.” In other words, the message is, you better embrace it or else you might get left behind.
I know that major race teams are using AI in ways that most of us will never know, or understand, so let’s keep this discussion focused on the grassroots level of motorsport.
As soon as I found out about ChatGPT, I started asking it almost every question I could think of about high-performance and race driving. In most cases, the responses are, at best, okay. Often, ChatGPT would come back with six or so bullet points; four of them would be accurate and useful; one of them would be questionable or irrelevant; and one of the bullet points would be completely wrong.
So, 67% accurate and/or useful. Hmmm…
The more I used ChatGPT, the more I “trained” it. After months of playing with, training, and learning how best to prompt it, I felt like I had gotten it to around 75-80% accurate. And that was after hundreds of prompts and questions to improve the responses it gave me. With that, I knew that I still had a place in the world. The problem, as I found out, is that ChatGPT mostly sources information from whatever was out on the internet as of September, 2021 (and yes, everything in that world is changing at least at the speed of an Indy car heading into Turn One at the Speedway, so it’s possible that itʻs different now from when I researched this a couple of weeks ago!). So, not terribly up-to-date. What I also learned is that not all the information thatʻs on the internet is accurate. I know, what a surprise, eh?!!
One day, out of the blue, I received an email from a driver, Jesse, who participates in track days and club racing. As part of his introduction to me, he said that he had gotten tired of having to dig through the index of my Speed Secrets books, or through other content I’d created to find the answer to a question he had to help him with his driving. And since he is a “tech guy” who develops apps, websites, and who knows what else, he had started to use AI to search the content that I’d written. That led to a conversation where we decided to see if we could create a Speed Secrets version of ChatGPT. Well, Jesse said he could do that, as I’m a bit (a LOT) lacking in the software development department.
First off, SpeedSecrets.ai, as we call it, would only source from information that I had created, and not from random sources on the internet. Fortunately, I’ve written over five million words about high-performance and race driving, so what we were developing easily had enough content.
Second, I didn’t want SpeedSecrets.ai to go public until I had tested it as much as I had ChatGPT (as well as Google Bard, by that point) and felt that it was generating responses that were at least 95% accurate. It took some time, but that’s where it is now.
As you may or may not know, I publish a Q&A column every week on my SpeedSecrets.com website called Ask Ross. Rarely does a week go by that I don’t receive at least one email with a question about driving, so years ago I started publishing my responses online. The downside for those who ask is that it might take months before my response shows up on my website. While I’m still currently publishing the Ask Ross Q&A column, SpeedSecrets.ai provides an answer – as if I was there, answering it myself – within seconds. And, I’ve tested it: it responds with 95% accuracy with how I’d respond.
Over the past four decades of coaching, I’ve worked with just about every type and level of driver, from young kids who have gone on to win Le Mans, the Indy 500, and NASCAR races to novice participants in HPDE (High Performance Driver Education) events. I find as much challenge and enjoyment in helping a group of drivers at a car club event improve as I do when I’m coaching at Le Mans. When I’m coaching one driver competing in pro racing, well, I’m coaching just one driver. When I’m writing, conducting webinars, creating videos, publishing podcasts, I’m helping hundreds and possibly thousands of drivers. As someone who really enjoys helping people learn, improve, and get the same thrill out of driving and racing that I have, the more the merrier.
The use of AI is one way of helping more drivers. I wish I could be at a track, coaching one-on-one with every driver who has ever wanted to drive on a track, but that’s not possible… unless I use technology, like AI.
I believe that within the next few years the use of AI will be just like when mobile phones came along. Do you remember when the iPhone was launched? Take a guess at what year that was? Yes, 2007. Sixteen years ago. And yet, doesn’t it seem like we’ve always had it in our lives (hey, no answers from you teenagers!)? How about the internet? It was available to you and me in 1993, just thirty years ago. AI will be no different. There will come a time in the very near future where just about everyone is using it, in some form. My guess is that it will take less time to become ubiquitous than it took cell phones.
My hope is that drivers all over the world will have access to something like SpeedSecrets.ai, an artificial coach who is there to answer questions, but more importantly, to give drivers specific instructions, practice strategies, and objectives prior to them going on track. That would result in more drivers learning more in less time.
That would make me very happy.
Amazing to me where this will lead...
As someone who's followed Ross for years and taught Machine Learning at the graduate level, I think SpeedSecrets.ai is a great way to use this new technology to make Ross more accessible to new generations of drivers. Meeting your target market via their medium of choice helps make this knowledge more relevant and accessible. I do raise a general concern about how outputs of LLMs are often interpreted by users as "real" knowledge - LLMs are not knowledge bases; they simply provide sequences of words based on a statistical analysis of the prompt against the data it's been fed. The reason ChatGPT's output seems so real is because it's giving us word sequences that follow the patterns of what we expect to read. It's actually pretty good at correcting grammar and summarizing larger blocks of text, but ChatGPT is notorious for making things up and presenting them as "facts," something that it seems Ross discovered. This is why so much effort needs to be focused on devising prompts which plumb the depths of what the LLM has been fed. Kudos to Ross and Jesse for your appraoch in setting up SpeedSecrets.ai using the SpeedSecrets books as primary knowledge base; I hope that you'll also work towards providing guidelines on how to make best use of the tool via the careful construction of useful prompts.