When Rodney Brooks talks about robotics and artificial intelligence, you need to listen. He is currently the Panasonic Professor Emeritus of Robotics at MIT and co-founder of three major corporations, including Rethink Robotics, iRobot, and his current project Robust.ai. Brooks also directed the MIT Computer Science and Artificial Intelligence Laboratory (CSAIL) for ten years starting in 1997.
In fact, he likes to make predictions in regards to the way forward for AI and maintains a scorecard on his blog how well he’s doing.
He knows what he's talking about, and he thinks it might be time to place an end to the hype around generative AI. Brooks thinks the technology is impressive, but perhaps not quite as powerful as many think. “I'm not saying LLMs aren't necessary, but we have to be careful how we evaluate them,” he told TechCrunch.
He says the issue with generative AI is that while it’s definitely able to performing a certain set of tasks, it cannot do every part a human can, and that humans are likely to overestimate its abilities. “When a human sees an AI system do a task, they immediately generalize to things which can be similar and estimate the competence of the AI system; not only performance on that, but competence in that area,” Brooks said. “And they are often very overoptimistic, and that's because they’re using a model of an individual's performance on a task.”
He added that the issue is that generative AI is neither human nor human-like, and so it’s improper to attribute human capabilities to it. He says people see it as so powerful that they even wish to use it for applications that don't make sense.
Brooks cites his latest company, Robust.ai, a warehouse robotics system, for instance of this. Someone recently suggested to him that it will be cool and efficient to inform his warehouse robots where to go by constructing an LLM for his system. However, in his estimation, this is just not a useful use case for generative AI and would actually slow things down. Instead, it is far easier to hook the robots up to a knowledge stream coming from the warehouse management software.
“If you could have 10,000 orders that just got here in and you’ll want to ship them inside two hours, you’ll want to optimize that. Language isn't going to assist with that; it's just going to slow things down,” he said. “We have massive data processing and big AI optimization techniques and planning. And that's how we will process the orders quickly.”
Another lesson Brooks has learned about robots and AI is that you would be able to't try too hard. You should solve a solvable problem that robots can easily integrate into.
“We have to automate where every part is already clean. My company, for instance, does quite well in warehouses, and warehouses are literally quite a restricted work environment. The lighting doesn't change in these big buildings. There's no stuff lying around on the ground because people pushing their shopping carts would bump into it. There are not any plastic bags flying around. And by and huge, it's not within the interest of the people working there to harm the robot,” he said.
Brooks explains that it's also about robots and humans working together, which is why his company designed these robots for practical purposes related to warehouse operations, moderately than constructing a robot that appears like a human – on this case, it looks like a shopping cart with a handle.
“So the shape factor we're using is just not humanoids walking around – although I've built and shipped more humanoids than anyone else. These seem like shopping carts,” he said. “They have handlebars, so if there's an issue with the robot, an individual can grab the handlebars and do whatever they need with it,” he said.
After all these years, Brooks has learned that it's all about making technology accessible and relevant to the aim. “I all the time attempt to make the technology easy for people to know so we will deploy it at scale. I all the time have a look at the business case; return on investment can be very necessary.”
Nevertheless, Brooks says we must accept that there’ll all the time be hard-to-solve corner cases in AI that would take a long time to resolve. “Without careful specification of how an AI system is deployed, there’ll all the time be a protracted series of corner cases that can take a long time to find and resolve. Paradoxically, all of those solutions are themselves AI-complete.”
Brooks adds that this misconception is especially as a result of Moore's Lawthat there’ll all the time be exponential growth in technology – the concept if ChatGPT 4 is so good, what’s going to ChatGPT 5, 6 and seven be like? He sees a flaw on this logic, namely that despite Moore's Law, technology doesn’t all the time grow exponentially.
He cites the iPod for instance. In some versions, the storage capability actually doubled from 10 to 160 GB. If this trend had continued, he calculates that by 2017 we’d have an iPod with 160 TB of storage, but in fact that was not the case. The models sold in 2017 actually had 256 GB or 160 GB because, as he identified, no person needed more.
Brooks acknowledges that LLMs could eventually help with household robots, where they might take over certain tasks, especially given an aging population and never enough people to look after them. But even that, he says, could bring its own unique challenges.
“People say, 'Oh, the large language models will enable robots to do things they couldn't do otherwise.' But that's not the issue. The problem of with the ability to do things has to do with control theory and all kinds of other hard mathematical optimizations,” he said.
Brooks explains that this might ultimately result in robots with useful voice interfaces for people in care situations. “In the warehouse, it doesn't make sense to inform a single robot to exit and get one thing for an order, but for home look after the elderly, it might be useful for people to have the ability to inform the robots things,” he said.
