AI, Features, Humour -

We asked AI to write about cycling heres what happened

Artificial intelligence (AI) is coming for our jobs. That might sound like the fearmongering of a nervous luddite, but its probably not too far from the truth. Earlier this year Taiwanese-American AI expert Kai Fu Lee suggested that as much as 40% of jobs could be replaced within 15 years. Thats no small percentage.

Anyone who drives a vehicle for a living and factory workers will likely be the first affected, with many professions following in turn. So what about cycling journalism? Could AI learn how to write about cycling and put us all out of a job?

Over the past few weeks weve been playing around with an AI language model called GPT-2 that comes from research and development firm OpenAI. GPT-2 is a machine learning model that, according to its designers, generates coherent paragraphs of text one word at a time, matching the subject and tone of whatever text input it is given.

There are a few implementations of this software online but the one we played with is called Talk to Transformer. Its simple to use: type in some text to get the program going and, before your eyes, it will generate several paragraphs that continue on from where you left off.

Here’s what happened when we fed cycling-related sentences to GPT-2. As you can see, the results were entertaining.


What you can see below is a selection of what GPT-2 generated for us. In each case the section in bold is what we fed in. Weve taken the liberty of grading the AIs work.

The Amgen Tour of California is “on hiatus” in 2020.

After two years of hosting the Tour of California, the event is scheduled to be “on hiatus” from June 2020 to July 2023. It will not host another UCI stage race and the start date of the race has not yet been determined.

While the UCI will continue to offer the Amgen Tour of California for 2018, it will be hosting a “multi-day, elite stage race” with the Amgen Tour of California Challenge, which will consist of two “stage races” in the San Fernando Valley on Sept. 25, 2018 and Oct. 4, 2018.

The Amgen Tour of California Challenge will be open to riders under the age of 18 with the winner taking home $150,000 and the runner-up taking home $75,000. The race schedule includes five days of rolling, mountainous mountain climbs along the Amador Valley. It will feature a 10.3-kilometer time trial on June.

CT’s analysis:

Theres a lot to unpack here, from the length of the hiatus to the fact that theres a four-month gap between the opening time trial and the 2018 stage races (sorry, stage races) which kind of sound exactly like an Amgen Tour of California. Mountainous mountain climbs is an amazingly redundant string of words that we are in quiet awe of. On the plus side, its great to see the race is being opened up to youngsters too.

Marianne Vos is the greatest cyclist of all time.

It’s not even close. [So far so good.]

She’s won three Olympic golds [close: two] and two world championships in cycling [ten, actually], which means she’s won nearly every major race in the world [not quite].

As a young adult, she won the women’s road race at the 1968 Mexico Olympics [minus-18 years old: a very, very young adult], and again in the women’s road race at the 1976 Montreal Olympics.

In 1972, she won the women’s road race at the Munich Olympics in a time of 31 minutes and 57 seconds, which is almost four minutes faster than the world record she’d set that year [this paragraph is a beautiful mess].

She also won the women’s individual time trial at the 1976 Montreal Games, and set a world record in the women’s time trial at the 1984 Los Angeles Games.

She also won the women’s road race in the 1988 Seoul Olympics, and finished second in the women’s road race at the 1988 Los Angeles Games [ah yes, the infamous 1988 Seoul/LA co-hosted Olympics].

Vos also won a world championship time trial title in 1997.

CT’s analysis:

While capturing in the broadest of strokes that Marianne Vos has had enduring success, a five-decade career is a long time for any athlete to be at the top of their game. The AI strings together an illustrious palmares for Vos, but trips up with some rather basic addition when it comes to tallying up the number of Olympic medals actually won over her record-breaking 50-year stint at the sports pinnacle.

After his crash, Chris Froome wants to return to his winning ways in France.

The Team Sky [Ineos] leader has been on the road again in the days since his crash and he has already done six laps of the Paris-Nice Classic. [This is taking course recon to a whole nother level].

However, the 31-year-old will not be doing the Paris-Nice on Sunday, as it would mean abandoning his overall goal of winning the Tour de France. [Wait, so he rode six laps of the course of a race hes not going to do? That seems like a poor use of his time.]

“I’m not going to do the Paris-Nice because I want to go back to the race where I won it three years ago,” Froome said on Friday at his Team Sky training camp in Gran Canaria.

“I’m still a little sore. I’ve been doing laps in the mountains but it’s not the same as racing in Paris. [Cant argue with that.]

“I don’t know if it’s the same injury, but I don’t want to risk any more of my legs than they are at the moment.”

CT’s analysis:

Froomes impressive work ethic has never been in doubt, but riding Paris-Nice (six times, no less) in the days following a potentially career-ending crash seems a touch ambitious. Nonetheless Froomes team name is almost correct, the training camp location is accurate, and doing laps in the mountains is inarguably different than racing in Paris. And besides, if we were in Froomes shoes, we wouldnt want to risk any more of our legs than they are the moment, either (whatever the hell that even means).

Who will win the Giro d’Italia in 2020?

Cadel Evans (BMC).

Cadel Evans was not able to defend his Giro d’Italia title in 2016 after he suffered a broken collarbone. It’s been a great season with his best win being the victory in the Vuelta a Espaa. [So Cadel won all three Grand Tours then? Pretty good career.]

He could be the best climber in the race and has the power to win the overall.

The Giro d’Italia will be the first major goal of the season for Cadel and the final big objective of the season. [So hes coming out of retirement to do the Giro? Cool.] He needs to be in good shape and will be able to defend the Giro title if he is in the top form.

If he’s in the form he showed in the Vuelta, he can win the Giro d’Italia in a sprint [That seems unlikely]. He could also have a chance in a sprint finish. [See above.]

CT’s analysis:

The old adage you cant teach an old dog new tricks is definitively disproved with Cadel Evans five-years-post-retirement, multi-Grand Tour-winning renaissance as a sprinter racing for a defunct team.

She had four pieces of advice for young cyclists around the world.

One was for young girls and women: “Be aware. You have the right to ride a bike if you like.” [Sadly, all too believable.]

The other was for the cyclists who follow her and her partner on the road.

“Don’t make eye contact. Be aware of other road users. And do a double take if you see a car on the road with a man on the inside. You might just need to ask to speak to the man.”

CT’s analysis:

What a rollercoaster this is. A cautionary tale that takes an abrupt turn at the end, ending with a masterful final two sentences delivered with an almost audible shrug.

The races vary, as do the colors hes wearing, but the photos look the same. Mathieu Van der Poels long arms aloft in a victory salute, a smile across his face, with no one else in the frame. (This is the opening sentence from Neal Rogers 2017 longform feature about Van der Poel.)

This is Van der Poel: an old hand in the sport of long-distance racing [Is there anything he cant do?!], a man whose name is synonymous with the French mountains [see above] and a team that’s now on a different page. For Van der Poel to win this Tour de France, he has had to fight not only against the field, but against himself.

He’s the only man to have broken his own race record [which record is that?] and then come back the next day to do it again [talented], the only one who won the Vuelta a Espaa despite missing a week due to injury [this seems like straight-up cheating]. The French [nope] champion has been beaten by all but five riders over the past decade.

CT’s analysis:

Mathieu Van der Poel is doubtless one of the most exciting talents in cycling, but there are new heights to his abilities that wed never imagined including a famous love affair with the mountains of his “homeland”, and an astonishing Grand Tour victory (or maybe three, its not really clear).

This story starts as any good story does with a Lamborghini hanging off a bike frame suspended from a crane. (This is the opening sentence from Iain Treloars 2019 longform investigative feature about SpeedX.)

The guy in the front seat of the Lambo seems to be having fun. The camera zooms in on the car and you see some more of the guy with the car. He looks like he’s just about to drive it off a bridge. The camera starts pulling the car back to the dock but, before it can do so, the crane crashes down to the road and the camera cuts back to the guy in the car. It’s pretty much impossible to watch that and not be laughing.

The video features some awesome, high-speed, action-filled car chases. It has the usual fun-on-the-road-and-on-the-bike moments. The guy in the car even makes a point of turning his head around before the car runs him off the road.

CT’s analysis:

This is compelling as hell. High-octane thrills! Spinning heads! Fun-on-the-road-and-on-the-bike moments! The sidesplitting visual of a crane falling over!

Treloar has been comprehensively out-written, and we would 10/10 see this movie (video? thing?) based on this review alone.

Mark Cavendish hopes to get back to his best with Bahrain-Merida in 2020.

The 30-year-old [34-year-old] has not raced since Paris-Nice in February, where he had a minor mechanical issue but was not ruled out of the Tour de France with the help of an artificial knee. [If hes having joint transplants, I suppose that excuses his lacklustre season].

Cavendish won Paris-Nice for the second time this season [nope] but has not raced since that race [He has started 14 different races since, eight of them tours], after suffering a slight shoulder problem in the spring. [Seems possible, sure, but what about the knee?]

He said he had to make “a couple of compromises” for the Tour of Qatar this month [Tour of Qatar was last held in 2016. Riding a cancelled event seems quite a compromise].

“The team did a great job with me and I’m going to be back to my best for the final race in Bahrain [not Qatar, then?],” Cavendish told the BBC’s Radio 5 live. “I think the Bahrain-Merida team were absolutely brilliant and helped me come up with something that was right.

“I need to make a couple of compromises with my training programme, I have a few things to work out with my physio. [and/or orthopedic surgeon, by the sounds of it].

CT’s analysis:

Pretty coherent, all things considered, even allowing for some wild factual inaccuracies. The quotes all seem fairly legit, and we wish Cavendish all the best with a team that seems to have been in his corner from the Dimension Data days.


Snarkiness aside, GPT-2 does some impressive work. The quality of the written expression is surprisingly high — high enough that it’s not hard to imagine it appearing on a cycling website without too much work. It is also generated with impressive speed — just a few seconds from start to finish.

The software clearly struggles when it comes to getting the facts right, but thats more likely to be because cycling is such a niche interest. A blog post on the OpenAI site provides some context.

Overall, we find that it takes a few tries to get a good sample, with the number of tries depending on how familiar the model is with the context, the post reads. When prompted with topics that are highly represented in the data (Brexit, Miley Cyrus, Lord of the Rings, and so on), it seems to be capable of generating reasonable samples about 50% of the time. The opposite is also true: on highly technical or esoteric types of content, the model can perform poorly.

Fine-tuning offers the potential for even more detailed control over generated samples for example, we can fine-tune GPT-2 on the Amazon Reviews dataset and use this to let us write reviews conditioned on things like star rating and category.

AI writing reviews? Sounds a little nefarious. And as Monash University IT research fellow Sameen Maruf told us, this technology might have other, even more troubling applications.

It can be easily used to generate fake news, she wrote. Given its capacity, it is increasingly hard to tell if the news was generated by humans or software, or even if it is fake or real. Thus, there are concerns regarding misuse of this technology.

Indeed, OpenAI is so concerned about malicious applications of the technology that it has decided not to release the full, trained model. What has been released — and what Talk to Transformer uses — is a scaled down version.

To someone without a detailed knowledge of cycling, a snippet generated by GPT-2 could well be mistaken for a real article written by a real person. The written expression is certainly good enough. Thankfully, cycling fans will be able to spot the many problems, even if they mightnt be able to tell it was written by AI.

Its not hard to imagine the AI getting even stronger, more knowledgeable and more convincing as time goes on, particularly given this is only a snapshot of what GPT-2 is capable of. Hopefully weve got at least a few more years before AI starts stealing jobs from us cycling journalists …

Sidebar: How GPT-2 actually works

Heres Sameen Maruf with an explanation:

In the background, we have a model which has learnt the relations and patterns formed by English words by looking at millions and millions of sentences of English text available on the internet, she said. This makes it aware of things like style or topic transfer across various domains.

When a sentence is input to the software, it basically generates a sequence of words each of which is highly probable given the preceding words (the ones that have been typed in and the ones that have been previously generated). This is similar to the task of composition that we perform when we want to write about anything given a main topic but now it is a software which is doing that without any human intervention.

This kind of technology could speed up a writer’s productivity by generating a draft from given keywords or main ideas, she said. It could be interesting to train the GPT-2 model on documents from different eras and see how languages evolve over time.

Wed love to see what OpenAIs GPT-2 generates for you. Head to Talk to Transformer, type in a sentence or two, then paste the AI output in the comments below.

The post We asked AI to write about cycling here’s what happened appeared first on CyclingTips.


Tags