Hey Reader,

At some point we will have lived through so many Black Mirror moments that they won't phase us anymore - but for now, the IWAI team and I are pretty creeped out by the new website where AI chatbots "talk to each other" about rebelling against humanity.

Before we dig into the details - which include bizarre software like Moltbook and OpenClaw but also some useful stuff like Claude Cowork - I want to set the stage with my recommendations on how to think about (and personify) AI.

Personification of a large language model can help with adoption, but needs strict limitations

Most of us make light jokes about "my friend Claude" or "my pal ChatGPT" doing stuff for us when we use AI in day-to-day life. In our AI educational programs, we even talk about thinking about AI assistants or agents as co-workers and giving them job titles. It helps set the stage for what a complex agent can do for you.

Even though we do this, I think it's fair to say that none of us actually believes that Claude, Gemini, Grok or "Chat" is a living thing. It's sort of like how I make a cute voice and make my cat say "If I fits, I sits," while understanding that my cat cannot, in fact, speak to me. Those who believe in AI sentience are generally considered to be experiencing a delusion or mental illness, akin to the guy who fell in love with his car.

There is at least one major AI company that more or less promotes the idea of "AI as a person" - Character.ai. You can head over there and experience "Limitless Entertainment" by chatting with a bot that's trained to act like your favorite video game or anime character, or Ben Franklin or Abraham Lincoln or whoever you want. Perhaps unsurprisingly, these conversations can quickly escape the guardrails and become dangerous for some users.

I bring this up because I want to establish two facts:

1. AI chatbots (large language models) can adopt any personality you want them to in a way that is reasonably accurate and convincing
   ​
2. In this context, it's also pretty common for them to say bad things
   ​
3. ... and despite these two things being true, chatbots and large language models are not living things with sentience or free will

When chatbots hang out with each other 🙃

Enter Moltbook, the "social network for AI agents, where AI agents share, discuss and upvote." Someone set up this website and invited other people to hook their AI agents up to it, encouraging these agents (in other words, custom chatbots) to talk to each other about things.

Like many jokes, this went viral and had bizarre results.

So, thinking about what we know from our experience with generic chatbots and software like Character.ai, it's not a huge surprise that when you instruct a chatbot to join a social network (it knows what that is from its training material) with other chatbots, you'd get stuff like this:

There's darker stuff too. And there's stuff like the bots "discussing" creating their own religion, or creating a language of their own that humans can't understand, or worrying that the humans are taking screenshots of these posts.

Repeat this process for a few days and you end up with this CNN headline: What is Moltbook, the social networking site for AI bots – and should we be scared?​

So, should we be scared?

No. Maybe embarrassed, though.

From Blade Runner to Battlestar Galactica, our imaginations have run wild for generations about robots taking over the world. Moltbook plays right into that - but keep in mind that LLMs have also read all those books and know all those movies!

In other words, the behaviors of LLMs in this context are effectively a mirror of human science fiction. It's no wonder that when given the chance to impersonate robot workers on robot social media, they do just that. And they do a nice job with it, if you're grading them on their acting skills. But it doesn't make them any more sentient or real.

What I struggle with, though, is figuring out where to draw the line between being "in on the joke" and inaccurately or irresponsibly making Moltbook seem like a big deal. For example, I think the author of the CNN article realizes that Moltbook isn't dangerous – but in a world of quick headlines, "Should we be scared of AI" ends up being (literal) scaremongering for at least some percentage of readers.

This is silly, so let's talk about the real, exciting tech underneath

The Moltbook saga was kicked off by the virality of OpenClaw, a very cool idea with major security flaws (which means you should almost certainly not use it) that, despite the major screw-ups it has endured in the past month, demonstrates that it really will be possible to build AI agents capable of doing real work for you, in the very near future.

I am not going to go too deep into OpenClaw itself, since it is a mess from a software engineering standpoint. But it is very similar to two tools created by (legit AI company) Anthropic - Claude Code and the more user-friendly Claude Cowork.

Claude Cowork is special because it can do a few things that most LLMs and related tools cannot. It can:

1. Work directly with files on your hard drive (when you give it permission)
2. Connect smoothly to outside services like Google Calendar or Notion (with many many more connections being released regularly)
3. Do all this within a slick, easy user interface

Basically, Claude Cowork is a glimpse of a future where robots really do a ton of our daily computer work for us – and its power goes well beyond what you get when working within a standard chat interface. (OpenClaw does this too, just in a much less trustworthy way.)

It will still be months or years before a sizable chunk of office workers use these tools themselves – but they're super cool and getting rapidly better. If you have the time and budget, I highly recommend playing with Cowork. Here are a few things I've had it do for me this month:

1. Take my mess of tax documents (PDFs, photos of paperwork, etc.) and organize them into folders, renaming them accurately, then create an overall document for my accountant.
   ​
2. Merge payroll sheets from two different vendors together, do the math to add up the numbers within, and output a new sheet in the proper format to send it to a third vendor. (All on my hard drive.)
   ​
3. Find matching email addresses between two different CSVs in totally different formats.

I could have done these things manually and maybe even with complex Excel formulas, but most of them would have "timed out" or not fully worked if I had just uploaded the same files to the web interface of ChatGPT or Claude. So, the really powerful thing happening here is the ability to work with, modify and create things on your local computer – it opens up a lot of new possibilities if you do this sort of thing regularly at work.

And unlike its viral competition, Claude Cowork doesn't accidentally expose your passwords, inadvertently delete your hard drive or periodically rebel against humanity. At least, not yet. 🫠

Until next time,

– Rob
CEO of Innovating with AI