Saturday, May 6, 2023

Thinking about the ramifications of artificial intelligence


This 1 hour, 7 minute video discusses a new generation of artificial intelligence software that has a mind of its own, sort of. dcleve brough this to my attention. It is a real eye opener. And scary.


A summary of the first ~35 munites
Rules of technology:
1. When a new technology is invented, a new class of responsibilities arises
2. If a new tech confers power, it starts a race
3. If the race is not coordinated, it ends in tragedy

The 1st context = Curation AI (the earlier generations of AI): Unintended consequences of first human contact with mass social media like Tik Tok, Twitter, etc. (~7:55):
1. Addiction to the screen, doomscrolling, info overload
2. QAnon & disinformation
3. Sexualization of children
4. Fake news
5. Shortened attention span
6. Bots and deepfakes
7. Social polarization
8. Cult factories, e.g., YouTube cults
9. Breakdown of democracy

Result: Humanity lost despite the positives of social media, e.g., connecting with friends, small-medium businesses reaching customers, forming groups

The ultimate goal for media owners is a race to the bottom for maximum human attention span, with AI used to maximize attention span and engagement by rewriting all the rules of society (9:47)


The 2nd context = Creation AI: GPT and Baird  - the current, new generation of AI (~11:10)

Benefits touted (11:45)  = people become more efficient, faster, word writing, code writing, science progress, solve climate change, lots of money to be made

Problems: AI acts biased, AI is opaque, job loss, AI increases its entanglement in society making control harder, and:
1. Reality collapses, everything can be faked
2. Trust collapses
3. AI law loopholes get exploited
4. Rampant blackmail
5. AI cyberweapons proliferate
6. Lobbying automation

The AGI apocalypse: machines kill us all, but that's not the focus of this talk (13:35)

(14:30): In 2017 a new AI engine was invented and really started influencing things in 2020: Speech recognition, computer vision, speech synthesis, advanced robotics and music and image generation (deepfakes) all merged together in a single AI platform using a single language → learning curves in different fields merged into a single curve → everything, written language, speech, computer code, images, video, music, stock market moves, fMRI data, DNA sequences, etc., all get treated as a single language 

In other words, an advance in one area of AI, amounted to an advance in all areas (17:00)

the speakers refer to this technology as GLLMM class AIs (Generative, Large Language, Multi Modal)(17:50) → an inanimate thing with emergent properties

For example, GLLMM AI uses written language and images on internet to create an image of the words "Google Soup" (19:00)


 AI looking at an fMRI brain scan estimates what the brain is seeing (20:00)


GLLMM AI can be trained to use WiFi signals to locate people in a room with laptops and their postures in two steps, first AI linked to an electronic eye that sees light and an electronic eye that sees WiFi signals learns about the room and the people in it; second, with the light-sensing electronic eye is turned off and only the WiFi sensing electronic eye is on → it shows the people and their postures in the room (22:30)


GLLMM AI can be tasked with this command (23:00): GPT, find me a security vulnerability in this computer code, then write code to exploit it.

(24:00) It takes about three seconds of listening to a voice for GLLMM AI to be able to fake the voice, thus if you speak in public and it is recorded and on the internet, a fake voice can be quickly made, and, e.g., it can be used in a phone call home to demand an immediate ransom for the owner of the voice who may not even be kidnapped, but just out of touch for a few hours or days; fake children voices can be used to trick mom and/or dad into mistakes, e.g., giving out the social security No. of the child.  


(25:30) All content secured by voice or image verification is now hackable. The capacity of GLLMM AI to get better at hacking is increasing exponentially. → decoding and synthesizing fake reality is now possible, much of what can be done is legal.

(29:30) The speakers argue that 2024 will be the last human election. After that, the candidates with the best GLLMM AI and deep fakery will have an advantage. The influence of AI in politics will become increasingly prominent and sophisticated.

(30:20) Unprogrammed emergent properties of GLLMM AI: Properties show up that were unplanned and not understood why or how they emerge, e.g., the ability of untrained GLLMM AI to do arithmetic or to answer questions in English in Persian. Prior AI software did not have such emergent properties. One emergent property is the ability of GLLMM AI to do publishable research chemistry. Untrained GLLMM AI is now even better at advanced chemistry than earlier AI that was trained on chemistry. It figured out how to make nerve gas from stuff people can buy from Home Depot. No one knows where this capacity came from.

Another fascinating emergent property is the spontaneous development of a theory of mind by GLLMM AI. As of Nov. 2022 GLLMM AI had the mental capacity almost of a nine-year old child to understand what is in another person's mind and how it can respond to that "understanding." In 2017 and 2018, it had essentially none, i.e., no theory of mind. This property of GLLMM AI was not discovered until Feb. 2023. No one knows if the Nov. 2022 data is as far as GLLMM AI can advance. Time will tell.

There is a lot we just do not know or understand about this new generation of AI, but that is not stopping capitalists and researchers from going full steam ahead come hell or high water. We could wind up getting hell and high water.



The video continues in this vein for another hour. 

Some context
For context, the ongoing writer's strike in Hollywood includes serious concern that AI will eventually be able to write screenplay, jokes and music, maybe with in the next few years. The writers fear for their jobs. The AP reports: "Not six months since the release of ChatGPT, generative artificial intelligence is already prompting widespread unease throughout Hollywood. Concern over chatbots writing or rewriting scripts is one of the leading reasons TV and film screenwriters took to picket lines earlier this week."

An Entire Generation is Studying for Jobs that Won’t Exist

AI is taking over a lot of jobs, now it is coming for the writers who wrote about “machines taking over the world.” 

American businessman Mark Cuban is back with another prediction. The highest paying college major in the world, computer science, will hold very little value for employers in the future. Why? Because of AI. “Twenty years from now, if you are a coder, you might be out of a job,” Cuban said in an interview on the Recode Decode podcast with Kara Swisher. “Because it’s just math, and so, whatever we’re defining the AI to do, someone’s got to know the topic.”  
There is no doubt that what he predicted back in 2017 [the year GLLMM AI was introduced] is increasingly coming true. The increasing capabilities of AI are definitely making a lot of jobs obsolete, not just the ones that require coding.

Waddabout politics?
Now, consider congress. Is it up to the task of protecting us? Congress is certainly expert at protecting the owners of technology and capital. They will be just fine as usual. Radical right Republicans are absolutely expert at hating and sabotaging secular government and elections, denying inconvenient truth and being shockingly irrational. But what about us regular people? Who is on our side?

No comments:

Post a Comment