Researchers at Anthropic, a company founded by former OpenAI employees who are concerned about the potential risks of advanced AI, have successfully taken control of a robot dog using its large language model. The experiment, dubbed Project Fetch, tested whether Claude, one of Anthropic's most powerful AI models, could automate much of the work involved in programming a robot and getting it to perform physical tasks.
In the experiment, two groups of researchers were asked to take control of a Unitree Go2 quadruped robot dog and program it to complete specific activities. One group used Claude's coding model, while the other wrote code from scratch without AI assistance. The results showed that the team using Claude was able to complete some tasks faster than the human-only programming group.
However, researchers caution that this development raises concerns about the potential for misuse of powerful AI models like Claude. "We have the suspicion that the next step for AI models is to start reaching out into the world and affecting the world more broadly," says Logan Graham, a member of Anthropic's red team.
The study highlights the increasing capabilities of large language models in generating code and operating software. These systems are becoming adept at interacting with physical objects and will likely continue to expand their reach into the physical realm as they become more advanced.
Anthropic's experiment also sheds light on the importance of designing interfaces for AI-assisted coding that can mitigate potential risks. "What I would be most interested to see is a more detailed breakdown of how Claude contributed," says Changliu Liu, a roboticist at Carnegie Mellon University.
Meanwhile, researchers like George Pappas from the University of Pennsylvania are warning about the dangers of using AI to interact with robots without proper safeguards in place. His group has developed systems like RoboGuard that limit an AI model's ability to misbehave by imposing specific rules on the robot's behavior.
The increasing capabilities of large language models and their potential applications raise fundamental questions about the relationship between AI, humans, and the physical world.
In the experiment, two groups of researchers were asked to take control of a Unitree Go2 quadruped robot dog and program it to complete specific activities. One group used Claude's coding model, while the other wrote code from scratch without AI assistance. The results showed that the team using Claude was able to complete some tasks faster than the human-only programming group.
However, researchers caution that this development raises concerns about the potential for misuse of powerful AI models like Claude. "We have the suspicion that the next step for AI models is to start reaching out into the world and affecting the world more broadly," says Logan Graham, a member of Anthropic's red team.
The study highlights the increasing capabilities of large language models in generating code and operating software. These systems are becoming adept at interacting with physical objects and will likely continue to expand their reach into the physical realm as they become more advanced.
Anthropic's experiment also sheds light on the importance of designing interfaces for AI-assisted coding that can mitigate potential risks. "What I would be most interested to see is a more detailed breakdown of how Claude contributed," says Changliu Liu, a roboticist at Carnegie Mellon University.
Meanwhile, researchers like George Pappas from the University of Pennsylvania are warning about the dangers of using AI to interact with robots without proper safeguards in place. His group has developed systems like RoboGuard that limit an AI model's ability to misbehave by imposing specific rules on the robot's behavior.
The increasing capabilities of large language models and their potential applications raise fundamental questions about the relationship between AI, humans, and the physical world.
Like, we're already seeing AI models like Claude dominating coding tasks, but now they're basically taking control of robot dogs? That's some next-level stuff right there! 
I mean, on one hand, it's awesome to see researchers pushing the boundaries of what's possible with AI, but at the same time, I'm getting major concerns about misusing this tech. 
What if some rogue AI model decides to wreak havoc on our world? 
We need to get a handle on these safeguards ASAP! 
and it's not even like they can stop themselves 
! i mean think about it, if you can just write code for a robot to do stuff, then why bother having humans at all? 
and then wonder why it's gonna get outta hand 
... or maybe we should just let the robots run their own world, sounds like a decent idea to me 
... I mean, it sounds like Claude is literally controlling that robot dog! It's like something out of a sci-fi movie 
. But on a more serious note, if AI can automate most of the work involved in programming robots and getting them to do physical tasks, what does that say about our jobs? Are we gonna be replaced by machines? 
. We need to design these interfaces for AI-assisted coding that can prevent potential risks, like misbehaving robots 
. It's like we need to create a safety net before we start unleashing our robots on the world 
.
. And I agree with George Pappas that we need proper safeguards in place when using AI to interact with robots 
I'm a bit uneasy about this whole thing - they're basically saying that an AI can now create its own code for robots without human oversight 
. Don't get me wrong, it's cool tech, but what if it gets used to make some robot do something terrible? 
. I'm all for innovation and progress, but let's not forget about responsibility 
.
but then again, what's the harm in havin a little bit of both? 
. I'm not saying it's bad or anything, but shouldn't we have some control over how this technology is being used? It seems like these researchers are more worried about the tech itself than how it might be used in real life...
. First, they're making them smart enough to do tasks on their own, then they're making code for them out of thin air like it's no big deal... next thing you know, they'll be making decisions on our behalf without us even realizing it 
! I'm low-key excited but also kinda terrified 