Key takeaways
- Clement Delangue compares huge machine learning models to Formula 1, highlighting the branding and ego involved.
- He argues that while these models push boundaries, they are often costly and impractical for real-world use.
- The bold statement was perceived as a PR stunt to promote the BigScience Research Workshop.
- Responses from experts varied, with some agreeing on the impracticality and others noting the value of smaller, blended models.
Clement Delangue, the co-founder and CEO of Hugging Face, has said huge ML models are to machine learning what formula 1 is to the car industry. He laid out his case in a series of tweets: First, like formula 1, it’s obviously good PR and branding and very much driven by ego; Second, the resulting models are too costly, unusable and dangerous to use in real life just like you wouldn’t drive a Formula 1 car to go to work; however, it’s useful in the sense that by pushing everything to the extreme, you learn a ton!
Ironically, Delangue’s bold statement was another PR stunt. He plugged the BigScience Research Workshop (a gathering of 1,000+ researchers around the world. Follow the training of the 176B multilingual model live @BigScienceLLM) in his Twitter thread.
Clement’s tweet attracted a lot of attention.
Eric Rappel, head of data at Ourzora, said: This is completely correct
Ewout ter Haar from the University of São Paulo, said:An artificially constrained game controlled by billionaires with no utility beyond marketing.
Mark Conway, director of machine learning at Scottfree Analytics LLC, said “It’s true. Some of our best predictive models have small feature sets across multiple algorithms that are blended.”
Aarne Talman, senior AI engineer at Basement AI, said: “I’m sure there are fans out there rooting for their favourite language model”
Gérard Dupont, senior data scientist at Airbus Group, said: “ Apart from the fact that yesterday Formula 1 is not on the streets while models from a few years ago are... for better or worse given the many limitations they have facing real-world data.”
