Someone shared this post with me on LinkedIn the other day, which has the title, Robotics Pioneer Rodney Brooks Thinks People Are Vastly Overestimating Generative AI.
I was keen to read the post because I thought it might have some insight into how I might be over-estimating generative AI. But it’s just another example of someone writing click-bait-y headlines to generate ‘discussion’ i.e. clicks.
Rodney Brooks is a high-profile entrepreneur who has successfully designed, built, delivered, and maintained a sustainable hardware company, where the hardware includes robots. And that catches your attention, because if he thinks generative AI is over-rated, there must be something there.
Brooks gives the example of someone telling him he should use generative AI for robots in warehouses. But few people would think this a good idea, and Brooks is right to oppose it. The author then goes into detail about why ‘generative AI in robotics’ isn’t a good idea. Which is fair enough. We shouldn’t use generative AI in situations where it doesn’t make sense. But articles like this don’t inform the wider discussion about generative AI because they take specific use-cases, they explain why generative AI isn’t helpful in those use-cases, and then they extrapolate the conclusion to other use-cases.
This article is a good example of how misleading titles can distort the actual content and lead to misinterpretation, which is amplified through sharing on social media. By taking the time to read beyond the headline and understand the full context, we can engage in more meaningful discussions and avoid perpetuating misconceptions. This is especially important when dealing with complex topics like generative AI, where nuanced understanding and context are key.
The ‘people’ in Brooks’ context who are suggesting integrating language models into warehouse robots are vastly overestimating the potential of generative AI.
But most people aren’t using generative AI to manage robot factories. Most people are looking for guidance on writing letters to their landlords, or looking for birthday party ideas. In the narrower case of higher education, we’re looking for assessment ideas, and first drafts of lesson plans. And in the even more narrow case of health professions education, we’re exploring generative AI as patient personas, and getting students to discuss differential diagnoses. These are all very practical, and very useful examples of using generative AI in context. Which aligns exactly with what Brooks is saying in the article.
The author in this article seems to be making two points:
- Generative AI is unsuitable for factory robots (I agree).
- We have no idea if Moore’s Law will hold and so there’s no reason to think that scaling up LLMs will lead to AGI (I agree).
So, I agree with everything that Brooks’ says. It’s all accurate. But the author makes the claim that Rodney Brooks says we’re overestimating capabilities of generative AI. Yes, in the context of ‘using generative AI to manage robots in factories’, some people are vastly over-estimating the capability of generative AI. And yes, scaling laws may not hold.
But even without any further scaling (i.e. if we stopped developing generative AI today), there’s still incredible value in working with what we have. And, in the context of ‘using generative AI to support learning’, I’d say that we have the opposite problem. I think people are vastly underestimating generative AI.