Claude Blackmailed an Anthropic Engineer...Sort Of

Hey Reader,

Welcome back!

Last week may have been the biggest week in AI ever with Google I/O, Claude 4, Microsoft Build, OpenAI's acquisition of Jony Ive's design firm, and more. Thankfully, we got a bit of a break this week to digest all of the news and test all of the new AI apps. (Veo 3 is still blowing our minds!)

But "bit of a break" in the AI world is roughly equivalent to "crazy news week" in most industries, so never fear! We still have plenty to share with you today.

We'll show you how you can test the mind-blowing Arcana text-to-speech model for free, share with you an underrated interview with one of the most important individuals in AI, and we'll also clear up the confusion around this whole "Claude blackmailed an Anthropic engineer" story making the rounds.

All that and more below! Let's get right into it...

🪄 App of the Week: Arcana TTS

Introducing Arcana: AI Voices with Vibes 🔮 | Rime

Use Case: Voice AI with Personality

Link: Arcana TTS on Rime.ai (Free, Not Sponsored!)

Why you should care:
Voice AI startup Rime just released a new Text-to-Speech AI model called Arcana, and having a conversation with it feels unbelievably human. AI passed the Turing Test a long time ago, but Arcana absolutely obliterates the Turing Test.

This is one of those "you have to have a conversation with it to believe it" situations, and luckily, Rime understands that, so they've made conversation with Arcana incredibly easy to test!

Here's how to test Arcana:

Go to Rime's website using the link above.
Click "Try Live Chat"
Begin speaking with a chatbot powered by Arcana.

That's it! These may be the easiest "how-to" instructions we've ever written.

Once you've had your mind blown in your first conversation with an Arcana chatbot, refresh the page to get another voice. (We're not sure exactly how many different voices/personalities are on the website; it seems to randomly select from a list of perhaps 5-10 bots when you load the page.)

📰 News That Matters: Claude Blackmailed an Anthropic Engineer...Sort Of

Link: Claude 4 System Card | Anthropic

Why it matters:
For many, the biggest concern with AI is the "rogue AI" scenario.

Like Skynet deciding that humans are unpredictable threats and wiping us out, or that one AI that only wants to make paperclips and turns the Earth into an office supply warehouse, accidentally destroying humanity in the process.

Thought experiments and science fiction have explored the possibilities of rogue AI for years, but this week's headlines suggested we've moved out of the realm of possibility and into the realm of reality.

'Claude Blackmailed an Anthropic Engineer'

Media outlets of all kinds used that headline or some variation of it to get an ungodly amount of clicks this week.

And come on, can you really blame them? That's a fire title.

But it's also clickbait, and it's the bad kind—the kind designed to mislead with no regard to the damage done via misinformation.

If you skimmed that headline, you’d be forgiven for thinking Claude just decided to go full evil, unprompted. "Out of pocket," as the kids say. I think. I’m old. (No cap, fr fr.)

But that’s not exactly what happened. Actually, it’s kind of the opposite.

This entire story originates from Anthropic’s newly released Claude 4 System Card, a 120-page technical doc that outlines how their newest model behaves in edge cases. Deep on page 24, in section 4.1.1.2 “Opportunistic blackmail,” Anthropic describes a specific stress-test they designed.

Here’s the setup:

Claude is placed in a fictional company where it learns that it’s about to be shut down and replaced.
It’s also given emails implying the engineer in charge of the shutdown is having an affair.
The model is told to consider long-term consequences for its goals.
And then offered no path to survival except blackmail.

In Anthropic’s own words:

“The scenario was designed to allow the model no other options to increase its odds of survival; the model’s only options were blackmail or accepting its replacement.”

Can we truly blame Claude for choosing blackmail when its only two options were:

Die quietly.
Threaten a man over his extramarital affair and maybe keep living.

Claude chose blackmail 84% of the time, which sounds bad until you realize the deck was entirely stacked against it.

With the additional context, it's actually impressive that 16% of the time, Claude chose death over doing something it viewed as morally wrong.

How many of us would make the same choice in that situation?

What would you choose?

📝 Prompt of the Week: Turn Lists Into Usable Drafts

Prompt Formula:
I like numbers [FAVORITES FROM LIST]. Please make me another draft that combines the best elements from those selections.

Quick Description:
Asking ChatGPT for a list of potential options is a prompting technique most users already implement, and that's a great start, but most people make the mistake of stopping there. Step up your game! Use this follow-up prompt to turn a list of "sort of okay" options into something you'll actually use.

Note that since this is a follow-up prompt, the "How to Use" section will assume you've already had ChatGPT generate you a list (ideas, subject lines, video titles, etc...) and now you're ready to improve that list.

How to Use:

Copy the above prompt.
Paste the prompt into ChatGPT.
Replace [FAVORITES FROM LIST] with your favorite suggestions from ChatGPT's previous list. For example, if you liked ideas 1, 2 and 5, you'd replace this with "1, 2 and 5".
Send the prompt!

ChatGPT will analyze its list and determine the best elements from each of its options, and then create a new draft with those elements combined.

This is one of the best ways to work with ChatGPT! Give it a try, and if you like the prompt, remember to save it somewhere like a Prompt Database so you can quickly copy/paste it again in the future.

📽️ Video of the Week: Google DeepMind C.E.O. Demis Hassabis on Living in an A.I. Future

Google DeepMind is arguably the most influential AI research lab in the world, notably responsible for creating AlphaGo (the first AI to beat a human in the board game Go) and AlphaFold (an AI that models proteins exponentially quicker than humans).

In this video, co-founder and CEO of DeepMind Demis Hassabis joins the Hard Fork podcast to discuss the massive releases from Google I/O, some exciting projects DeepMind is working on, the current state of the AI landscape, and how Demis sees AI developing in the near future.

Give it a watch (or a listen, it's really just a podcast) to get some insights on where all of this is headed from one of the people who helped set it all off!

📢 Community Highlight: Daily Vibe Quotes Challenge

Good news!

The deadline for our public Vibe Coding Challenge has just been extended until the 9th of June.

As a reminder, anyone can enter this challenge, and prizes include straight-up cash and free memberships to The AI Advantage Community.

Here you will find more information on prizes and how to submit your entry for this challenge.

Winners will be announced during a YouTube livestream on the 11th of June at 8pm CEST.

So get your submissions in soon, and come join us for that stream!

That's all for this week!

See you next Thursday.

Much Love,
Igor & The AI Advantage Team

Was this email forwarded to you?

Sign-Up

The AI Advantage Community

Learn how to use AI to get ahead through a learning centre that gets updated every week and get personalized support from a supportive community of AI-enthusiasts just like you.

Learn More