How to combat Evil AI™ - 🤖😋🧠 #2

Author

Anthony Castrio
September 19, 2022

Welcome back to The Bot Eat Brain Daily* Newsletter. We're the newsletter your TI-86 graphing calculator would write if it ever got an update.

*Daily is a goal and not a promise alright? Let's all be cool here.

What's happening today:

  • How to combat Evil AI™ ⚔️

  • Your imagination is now a movie 🎬

  • You're being followed 👀

Today's vibe check: Shoot to kill!

Let's go 🦾

🥩 Meat of the day: How to combat EVIL AI.

In the court of the Evil AI resistance team this week Evan Hubinger lays out his game-plan to fight against hypothetical evil AI.

While AI giants like Google and OpenAI race to build stronger and stronger AI's the threat of an Evil AI™ looms ever larger.

We can take steps now to watch for early signs of deception in their models and in our own.

The resistance is worried that powerful AI might exhibit evilness and badness through the power of lies.

They could make everything look hunky-dory to the people training it to not be evil, but secretly be very much evil.

But there's hope!

A young AI is likely to be no better at lying than your 3-year old nephew, so it should be easy to catch an evil but dumb AI with it's hand in the cookie jar.

How to catch an AI that's evil and bad at it:

1. Make the AI try to be sly even if it’s easy to catch it being sly, like your nephew with their fingers crossed behind their back.

2. Trick the AI into thinking it’s deceiving you through a trick of your own. AKA Stay one step ahead.

3. Make a bunch of deceptive AI's that will tattle on each other because they aren’t friends or to stay out of trouble themselves.

4. Pay close attention to the innards of your AI using anti-evil-AI inspection tools... that don't exist yet.

Just because we see it, doesn't mean we can fix it.

But once we see it we'll know it's possible. It might not be a bad thing for everyone to be a bit more scared of AI and it's potential for evil.

Once we know what to look for, we can learn techniques to prevent the emergence of an evil intelligence.

Be on the lookout for...

😈 Models that default to evil.

If one type of model always learns how to tell lies, then we can stop making those types of models and concentrate on benevolent models instead.

🧟 Models that refuse to die.

We could see models that try to avoid being killed off and notice that. Self-preservation is not a good skill for evil to master.

🤪 Models that try to look dumber than they are.

We could see models that try to give stupid answers on purpose or models that try and make you like them more than you should.

Right now it’s very likely that AIs won’t be sly and if they are it will be easy to catch. That's why now is a good time to start.

Once AI gets good at being evil, it won't be cute, it will be too late.

🎬 Your imagination: a movie

Magical AI tools, realtime collaboration, precision editing, and more.

Runway AI announced a NLP to movie pipeline. You type it, Runway turns it into a movie.

Your next unicorn startup idea 🦄

  • Want to adapt your novel into a movie? Drop your novel into GPT-3 to turn it into a script. Drop your script into Runway to make it a masterpiece.

  • Need a promo video for your new product launch? Forget about hiring a film crew, just tell the AI what you want to see.

👀 You're being followed

Step 1: You take a selfie on a public street.

Step 2: You post it on Instagram.

Step 3: The follower tracks you down across the Earth, using public web-cameras to find video footage of you taking that photo.

It's one of the spookiest demo's we've seen all week:

How does it work?

  1. Recorded a selection of open cameras for weeks.

  2. Scraped all Instagram photos tagged with the locations of those open cameras.

  3. Software compared Instagram with the recorded footage.

​Whodunit:​

How worried should you be?

The truth is that nothing here has changed, you've just been made of the reality you're living in.

Anyone with a bit of compute and some time on their hands (especially any government agency) can find you via public surveillance.

If you want to avoid detection consider taking anti-surveillance measures:

That's it for today, we hope you got your fill.

Can't get enough?

🤖😋🧠