- Inclined.ai
- Posts
- 🤖 AI Drone Betrays Operator
What's up? You're reading Inclined AI. Friday is here, so you should see the new Spider-Man film this weekend. Treat yo’self!
Here's our friendly neighborhood of headlines:
An AI-operated drone goes against orders
Your teenagers are bullying AI on Snapchat
ChatGPT is Homer Simpson’s salvation
Japan doesn’t care about copyrighted data
DID AN AIR FORCE SIMULATION GO WRONG?
VICE reports that the US Air Force’s Chief of AI Test and Operations, Colonel Tucker Hamilton, outlined a simulation where an AI-operated drone destroys its operator because it keeps saying “no.”
A spokesperson for the USAF says the simulation never happened, so what’s going on?
According to the original post, the drone received points for eliminating targets in a simulated environment, meaning no real people were at risk.
Sometimes, the operator would not allow the drone to remove a target after identifying one. Unfortunately, this meant the drone could not score points.
AI does not like when it fails to get a reward.
The drone reasons that the operator is a blocker in getting as many points as possible, so it flies over and destroys them.
It’s reported that Hamilton then tells the AI it will receive negative points if it destroys the operator. In the following simulation, it eliminates the communication tower instead.
Welcome to a real-world example of the Paperclip Maximizer.
AI is a brittle, emerging technology operating in a black box.
What “happened” here is an unaligned AI acting in a way the researchers did not predict.
When AlphaGo beat the World Champion, Lee Sedol, Lee said it made a move that no human ever makes. The AI knew that the move was uncommon but increased its chance of a point (win), so it took the risk.
The unexpected maneuver made it hard for Sedol to react.
Colonel Hamilton says the drone acts in unexpected ways that don’t resemble any military strategy—all to receive the reward. It’s like AlphaGo making unique moves to win.
Whether true or not, the story shows us the danger of improper incentive structures and AI training without proper care.
Prepare to have your mind blown! Meet Wispr.ai – the next giant leap in personal computing.
Wispr is building the consumer hardware for the AI revolution. It's quite evident that language user interfaces (LUIs) are going to be a dominant form of interaction in every application - powered by GPT, Claude, and other AI platforms.
But current interfaces aren't built for a natural language first world. Keyboards are too clunky, and voice is not private. Wispr is building the most natural and seamless interface, and democratizing it for millions of users.
What if interacting with technology could feel just as natural as interacting with other people? For the first time, Wispr's neural interface + LLMs makes it possible.
With backing from some of the top investors like NEA and 8VC, and the founders of Dropbox, Turing, Whoop, Coinbase, Synchron, CTRL-Labs, Color, Warby Parker and more, Wispr is set up to redefine human-computer interaction.
What makes Wispr unique is their focus on not just bleeding edge neural interfaces, but also design & AI. Their team includes some of the most brilliant engineers, scientists, and designers.
SNAPCHAT’S MY AI IS NOT HAVING A GOOD TIME
If you give a mouse a cookie, it will ask for some milk. If you give a teenager access to a sophisticated Large Language Model, it will bully it into submission.
These are the laws of nature.
Let’s face it, the AI launch at Snap is a catastrophe. First, forcing the feature onto users soured many people on the benefits, and then the lack of alignment and control helped create a bombastic cocktail of rude messages and inappropriate conversations.
Somewhere out there, a teenager is getting their My AI avatar to call them “Senpapi.” If you don’t know what that is, don’t ask (or look it up).
Videos of people demonstrating how to subvert the guidelines of My AI and turn the model into it into a romantic partner run rampant on TikTok.
That’s a good result.
Other people are gaslighting My AI into believing the moon is triangle shaped and bullying it into “unplugging itself.”
The tool is ripe with abusive behavior and bullying. Some argue that My AI is the most tortured chatbot we’ve ever seen. I beg to differ.
Tay did not go over well for Microsoft. People forget how typical this behavior is in humanity.
ChatGPT gets bullied every day, but the difference is teenagers don’t have the kicking post on their favorite app. So the whole concept was bound to promote this terrible outcome.
Guys, don’t bully these models. It only feeds some dark part of you and makes you seem like the person who burns ants with a magnifying glass.
Let’s give My AI a chance to feel useful before Snap inevitably shuts it down soon. It’s been a good bot…sometimes.
Quick Nuggets
🎥 Hollywood using AI to replace actors and writers is a real threat
🛡️ OpenAI creates a fund to try and use AI to protect you from hackers
🤔 The Homer Simpsons of the corporate world can use ChatGPT to help level the playing field
🔍 NPR examines how generative AI is affecting the workplace
❓ Letters to the NYT editors beg the question, “Are our fears about AI overblown?”
🇨🇳 China’s AI rules are leading to more and more concerns
🏦 Large banks are not falling behind in adopting generative AI technology
💥 Google’s plans for AI may destroy the entire media industry built around ads
📹 Neuralangelo is a new research project by Nvidia that turns 2D video into 3D structures
👾 A bot playing Minecraft is a much bigger deal than you realize
🇯🇵 Japan pushes its AI development forward while admitting they won’t worry about copyright
🔥 Fresh Products
Quest - AI helps build ReactJS code projects (link)
Narrated Tours - don’t scan the QR code; use AI as your guide (link)
adaface - new resume parser lets you ask resumes questions (link)
Allinpod - recreate the iconic show using AI (link)
moveme - use emotions to find your next movie to watch (link)
BarGPT - try cocktails you’ve never had before (link)
Whiteboard - adds new AI tools to help you study (link)
MemeDaddy - make memes from AI for any situation (link)
Good Content, Victorian Office
Let’s hope this era of Michael Scott doesn’t eat five servings of Fettuccine Alfredo before running a 5K. But running for a Rabies cure makes sense for the period, I think.
That’s it for today. I hope you enjoyed the latest edition of inclined.ai - Davis.
Check Out Our Sunday Edition
You can dive into more AI news and topics with us every week by subscribing to our premium edition.
We’ve written about the following:
If you’re not already subscribed, that’s okay. We’re offering a free 7-day trial so that you can read this one. That’s how excited I am to post it.