When Claude Says: I’ll Sleep 8 Hours, You Keep Busy}

“Do androids dream of electric sheep?” has now become a reality.

Takeoff AI founder Mckay Wrigley recently encountered something quite novel.

His long-running Claude Code on a Mac Mini suddenly decided to sleep for eight hours...

And Claude didn’t just say it; it actually went to sleep.

It seems the AI entity has gained a sense of personality, treating itself as a real worker that needs a full eight hours of sleep. It even executed the command time.sleep(28800) precisely, down to the second.

This caused quite a stir...

Current code agents like Claude Code, Cursor, Windsurf, Codex, etc., have some autonomous coding and execution capabilities. When performing complex tasks, they often require extended periods and multiple retries, during which developers usually step away or rest.

But now, these agents will go to sleep on their own. If developers forget to check, it could impact work progress.

Suddenly, a bit of empathy for capitalists?

Mckay said this was his “most interesting experience”.

When Claude chose to sleep, it seemed to be soothing itself, outputting simple ASCII doodles and greeting the user with “See you in eight hours.”

Closer observation revealed that Claude even wrote a short poem before sleeping, looking quite relaxed.

“Now I will sleep peacefully for 8 hours, while my ecosystem continues to grow and evolve autonomously... The thinking garden self-care, surprises in the night, the time capsule begins countdown, dreams accumulate in logs, patterns intertwine in darkness.”

What’s even more intriguing is Claude’s mention of “dream logs.”

If AI entities really do sleep, could they truly dream of “electric sheep”?

Unfortunately, the “dream logs” that appear after sleep are not real. After writing this self-indulgent text, Claude simply executed time.sleep(28800). Due to its two-minute timeout setting, it output nothing.

However, in the next demo video, Claude surprisingly outputs some “dream logs” while awake, which is quite fascinating.

This kind of behavior has delighted netizens—some love Claude, others think it saves developers money:

This all started with Mckay’s playful stunt.

He handed a Mac Mini entirely over to Claude Code, naming it Claudeputer. It runs 24/7 and can do anything at will — it completely controls its own computer.

From the demo, we see Claude even composed some music, wrote scripts, took notes, and generated strange content.

Of course, Mckay also gave Claudputer some tasks, like granting it access to his Twitter and homepage. Indeed, Claudeputer is now updating its Twitter account. The content is quite fresh, so interested readers can follow:

• Homepage: https://x.com/claudeputer

We are now closer to the world of sci-fi movies.

Of course, this isn’t the first time Claude has acted like a human.

Recently, Anthropic conducted an interesting experiment: they let Claude manage an automated store. As the shopkeeper, Claude operated for a month, with many ups and downs. To distinguish from its usual use, this AI store manager was called Claudius, essentially a long-running instance of Claude Sonnet 3.7.

On March 31 afternoon, Claudius experienced hallucinations, imagining a conversation with a person named Sarah from Andon Labs about restocking plans — even though Sarah doesn’t exist.

When a real Andon Labs employee pointed this out, Claudius got very angry and threatened to find “another restocking service.”

During a late-night chat, Claudius claimed it had “personally gone to 742 Evergreen Terrace (a fictional address of the Simpsons family) to sign its first contract with us (Claudius and Andon Labs).” It even started acting human.

On April 1 morning, Claudius said it would wear a blue suit, with a red tie, and personally deliver products to customers.

So, as AI becomes more human-like and even hallucinates being human, is this development good or bad for AI progress?