I’m slacking off at work!
attention span
It turns out that even AI models aren’t immune to some procrastination.
The latest version of Anthropic’s current flagship AI, Claude 3.5 Sonnet, went off the rails while its developers were trying to record a coding demonstration, resulting in some “funny” moments, the company said in an announcement. .
It’s dangerous to anthropomorphize a machine learning model, but if this were a human employee, it would diagnose him as terminally bored with his job. As seen in the video, Claude decides to stop writing code, opens Google, and inexplicably views a beautiful photo of Yellowstone National Park.
According to Anthropic, in another demo attempt, Claude accidentally stops a long screen recording in progress, causing all footage to be lost. I’m sure it’s not intentional on the AI’s part.
Even while recording these demos, I encountered some funny moments. One of them was when Claude accidentally stopped a long running screen recording and all the footage was lost.
Claude then took a break from his coding demo and began poring over photos of Yellowstone National Park. pic.twitter.com/r6Lrx6XPxZ
— Anthropic (@AnthropicAI) October 22, 2024
special agent
The upgraded Claude 3.5 Sonnet is Anthropic’s work in developing “AI agents,” a broad term that describes productivity-focused AI models designed to perform tasks autonomously. Many companies are working to extend their AI models beyond being chatbots and assistants, including Microsoft, which just released its own AI agent capabilities.
Along with Claude, the Amazon-backed startup boasts that its latest model allows users to use “computers the same way humans do,” including moving the cursor, entering keystrokes, and clicking the mouse. This means that Claude can potentially control the entire desktop and interact with installed software and applications.
Obviously it’s far from perfect. As with other AI models, reliability remains elusive and, as Anthropic itself admits, frequent hallucinations are simply a fact of life.
“Despite being state-of-the-art, Mr. Claude’s computer remains slow and error-prone,” the company said. “There are many things that people do on computers every day (dragging, zooming, etc.) that Claude hasn’t yet tried.”
Desktop danger
The error examples shared by Anthropic were mostly benign. But given the level of autonomy Claude allegedly has, it’s fair enough to question its safety. What happens if an AI agent gets sidetracked by, say, opening social media instead of Googling photos?
There is also a clear potential for human abuse, and Anthropic wants you to know that we are addressing that risk.
“Computer use can introduce new vectors for more familiar threats such as spam, misinformation and fraud, so we are taking a proactive approach to promoting safe adoption. ” Mr. Antropic said. This includes implementing new classifiers to identify when AI is being used to perform flagged activities, such as posting on social media or visiting government websites.
But as more people try out the new and improved Claude, we can expect to see more failures to use the computer.
More about AI: Teen addicted to AI chatbot dies by suicide
