Welcome back to The Algorithm! 

I actually have a chair of shame at home. By that I mean a chair in my bedroom onto which I pile used clothes that aren’t quite dirty enough to scrub. For some inexplicable reason folding and putting away those clothes looks like an amazing task when I’m going to bed at night, so I dump them on the chair for “later.” I’d pay good money to automate that job before the chair is roofed by a mountain of garments. 

Due to AI, we’re slowly inching towards the goal of household robots that may do our chores. Constructing truly useful household robots that we are able to easily offload tasks to has been a science fiction fantasy for many years, and is the final word goal of many roboticists. But robots are clumsy, and struggle to do things we discover easy. The kinds of robots that may do very complex things, like surgery, often cost lots of of 1000’s of dollars, which makes them prohibitively expensive.

I just published a story on a brand new robotics system from Stanford called Mobile ALOHA, which researchers used to get an inexpensive, off-the-shelf wheeled robot to do some incredibly complex things by itself, equivalent to cooking shrimp, wiping stains off surfaces and moving chairs. They even managed to get it to cook a three-course meal—though that was with human supervision. Read more about it here.

Robotics is at an inflection point, says Chelsea Finn, an assistant professor at Stanford University, who was an advisor for the project. Up to now, researchers have been constrained by the quantity of knowledge they will train robots on. Now there’s so much more data available, and work like Mobile ALOHA shows that with neural networks and more data, robots can learn complex tasks fairly quickly and simply, she says. 

While AI models, equivalent to the big language models that power chatbots, are trained on huge datasets which were hoovered up from the web, robots must be trained on data that has been physically collected. This makes it so much harder to construct vast datasets. A team of researchers at NYU and Meta recently got here up with a straightforward and clever option to work around this problem. They used an iPhone attached to a reacher-grabber keep on with record volunteers doing tasks at home. They were then in a position to train a system called Dobb-E (10 points to Ravenclaw for that name) to finish over 100 household tasks in around 20 minutes. (Read more from Rhiannon Williams here.)

Mobile ALOHA also debunks a belief held within the robotics community that it was primarily hardware shortcomings holding back robots’ ability to do such tasks, says Deepak Pathak, an assistant professor at Carnegie Mellon University, who was also not a part of the research team. 

“The missing piece is AI,” he says. 

AI has also shown promise in getting robots to reply to verbal commands, and helping them adapt to the usually messy environments in the true world. For instance, Google’s RT-2 system combines a vision-language-action model with a robot. This enables the robot to “see” and analyze the world, and reply to verbal instructions to make it move. And a brand new system called AutoRT from DeepMind uses an identical vision-language model to assist robots adapt to unseen environments, and a big language model to give you instructions for a fleet of robots. 

And now for the bad news: even probably the most cutting-edge robots still cannot do laundry. It’s a chore that’s significantly harder for robots than for humans. Crumpled clothes form weird shapes which makes it hard for robots to process and handle.

However it might just be a matter of time, says Tony Zhao, considered one of the researchers from Stanford. He’s optimistic that even this trickiest of tasks will someday be possible for robots to master using AI. They only need to gather the info first. Perhaps there’s hope for me and my chair in spite of everything! 

A Birthday Special

How MIT Technology Review got its start

We’re turning 125 this 12 months! Thanks for sticking with us all these years. Here’s the way it all began—and the way the fledgling magazine helped rally alumni to oppose a merger with Harvard.

Did you already know? When the publication was founded in 1899, The Technology Review, because it was first titled, didn’t concentrate on the applying of scientific knowledge to practical purposes. It was a magazine about MIT itself—or “Technology,” as its earliest alumni fondly called it. Read more from Simson Garfinkel here. 

Bits and Bytes

Meet the girl who transformed Sam Altman into the avatar of AI
An amazing profile of Anna Makanju, OpenAI’s vp of worldwide affairs. She is the girl who orchestrated Sam Altman’s global tour meeting world leaders, transforming him into the AI sector’s ambassador in the method. (The Washington Post) 

It’s “inconceivable” to create AI models without copyrighted material, OpenAI says
In a submission to a committee within the UK’s House of Lords, the AI company said it couldn’t train its large AI models, such GPT-4 and ChatGPT, without access to copyrighted work. The corporate also argued that excluding copyrighted content would result in inadequate systems. Critics, equivalent to professor emeritus at NYU Gary Marcus, called this “self-serving nonsense” and an try and avoid paying licensing fees. (The Guardian) 

US firms and Chinese experts engaged in secret diplomacy on AI safety
With the blessing of presidency officials, OpenAI, Anthropic and Cohere met with top Chinese AI experts last 12 months. The meetings were in regards to the risks referring to the technology, and inspiring investment in AI safety research. The “ultimate goal was to search out a scientific path forward to securely develop more sophisticated AI technology,” writes the FT. (The Financial Times) 

Duolingo has cut 10% of its contractors because it creates more content with AI
The language-learning app company has fired a few of its contractors and has began using more generative AI to create content. The corporate says it is not a direct alternative of employees to AI, but a results of its employees using more AI tools. It would be interesting to see how well this can serve Duolingo in the long run, knowing how flawed and biased generative AI will be. (Bloomberg)