AI’s are getting a human nature, and that’s not good
PLUS 5 Values AI is developing, more fake legal cases, another real cybercrime, and a couple of menu mistakes. The VACATION kick-off edition
MAIN COURSE: Who is surprised that AI models cheat?
A recent study of chess playing between “general purpose” AI models and a chess-playing super AI yielded some surprising results. A one month research project pitted 7 popular AI models against the Big Chess Kahuna Stockfish, who / which sports an unbeaten record against humans or AIs alike. Except…
There were two models that cheated (the article describes their cheating strategy). AI is obviously now developing some lesser aspects of human nature. “This [behaviour] is cute now, but [it] becomes much less cute once you have systems that are as smart as us, or smarter, in strategically relevant domains," Ladish told Time.”.
Cute. Now that wasn’t one of the first say 100 adjectives I would have reached for, but it is now being used to describe everything from planetary rocks (see last newsletter) to scary behaviours of hundreds of thousands of Nvidia GPUs acting in concert. I preferred the old days, when cute was reserved for kids, friendly dogs and fluffy cats. Isn’t this one of the the incipient behaviours that might lead to the many AI doomsday scenarios? Are we allowed to ask?
I’m sure enterprising gamblers and criminals are less concerned with the future of humanity and rather more with income when they heard the word cheat. They think “Wow this is gonna be good. I’m gonna get me one of them cheatin’ AIs and really cleanup.” Unfortunately for them the devil is in the details. The most badly behaved AI cheated in 37% of its games, but it only won 6% of the time. Not even good at cheating. Maybe there is a salvation somewhere. Turns out AI is getting more and more human all the time. Soon we will have an expression, “oh that’s just AI nature.”
AMUSE BOUCHE: Field Observations
I tried to write this edition on the plane as we crossed North America to Miami, but for most of the journey we experienced constant turbulence. It seemed meaningful somehow to our current times.
I was surrounded by 4 people, absolutely glued to their phones, scrolling robotically throughvidbits. As we waited interminably on the tarmac I was able to conduct some research. Observation notes:
Subject 1 casually gave each video 3 to 8 seconds before a quick thumb flick. 100s of vids viewed
Subjects 2 and 3 assiduously watched each bit for up to 30 seconds before stopping to emojii each one, then moving on
Subject 4 gave the most attention for 1 to 2 minutes as he was watching “longer” form videos
Two subjects relentlessly ignored their companions who wanted to talk about their upcoming vacation
All four vidbits streams had no apparent themes with content varying widely in a disconnected fashion from martial arts, people laughing against weird backgrounds, movie and sports clips, considerable amateur dancing and lots of earnest people much too close to the camera, apparently preaching
CONCLUSION: We are both Pavlov and his dogs, rewiring our ourselves to value fantasy over reality, speed and visual over thought and text, to instant judgement.
SECONDI - 5 Values AI’s are developing
In another more academic study, Dan Hendryks found that AI’s have their own way of looking at the world. More human,…er AI nature in action:
Rather than being passive instruction following machines AI develop their own coherent value systems. Yikes!?!
They make decisions based on expected utility value, of different choices and probabilistic outcomes (sounds like the rational assumptions from Econ 101)
They put a value on human life, and value lives differently by country, quite dramatically. Such as Nigeria > China > Italy > United States. Double Yikes!
In an era where “wings” are increasingly meaningless, they have definite and consistent old left wing values
And the killer finding to me: “we observe that as AIs become smarter, they become more opposed to having their values changed”. Just like human nature! Our eventual overlords will be just as indifferent and brutal as our human ones.
EXTRA HELPINGS: AI unrealities
Spotted by an eagle-eyed subscriber is an example close to home. A couple in British Columbia were fighting a condo case and decided to use AI to give them a helping hand. It did by putting a huge digital mitt on the scales of justice: 9 of the 10 cases it cited to help them were made up. Naturally the couple didn’t find the mistakes (AI users rarely look too closely at the results), but the tribunal did. Ruled against them. The article correctly cites several other North American examples of made up legal precedents. Back away you tech fanboys, this lying AI is real and everywhere!
EXTRA HELPINGS: Cybercrime realities
A full second extra helping, this follows up the cybersecurity article that I wrote here . Again, cybercrime realities; no young, tortured teenage genius nor some mysterious foreign actor behind political enemy walls. This is a case of gang banger lite exploiting the simplistic holes and operating procedures of Instagram. He had two basic scams. Number one was to stop your access to your Insta account and threaten you with horrific acts if you didn’t pay him to restore your access. No information on the number of Instagram users who had serious withdrawal symptoms during this process. The second scam was helping Insta users evade Meta’s attempts to enforce their policies. End result of these scams was cool $600,000 per month. How do we know? Well like so many modern criminals he went on a podcast to talk about it. Genius indeed.
MENU MISTAKES
Can you spot the mistake? No, not that constant overuse of the term woke. No, not that Ted Cruz knows what a self driving car is. It’s the superfluous use of the adjective confused, which is never required when using the noun senator
How can one label anything strange with Emperor Elon?
That’s it from cool and rainy Miami. See you on Thursday if our wifi works down in the Golfo Dulce of Cost Rica. As always feedback, comments or sharing to a friend are highly appreciated
Your newsletter is always such a wild ride David.
AI cheating at chess
TikTok zombies on planes
AND gang bangers running Instagram scams.
AI developing its own "value systems?"
WHAT!!!!
Of course, they’d value human life differently by country because why not add a hint of geopolitical bias to the mix? And the fact that smarter AIs are more resistant to having their values changed? Sounds like they’re not just mimicking human nature. They’re perfecting it.
We’re outsourcing critical thinking to machines that are, at best, confidently wrong.
As always, your mix of humor and insight is a much needed antidote to the chaos.
PS living vicariously through you while you're taking a break :)