Thinking about the Pace of Machine Learning
I recently met with a colleague and explained how Machine Learning (ML) works and she applied for a job in the field the same day so I may be responsible for this in some way.
This is a collection of Twitter and Mastodon messages that I wrote about ML and the emerging Large Language Model (LLM) technology behind recent spectacles like ChatGPT and competitors like Jasper AI, Bard AI, Bing, CoPilot, Elsa Speak, YouChat, Poe by Quora, My AI on Snapchat, etc. I placed these messages in a timeline to provide a sense of how quickly things changed.
It started with asking the computer to differentiate between a traffic cone and a toddler. Next, we were feeding images and JSON (JavaScript Object Notation, an open standards data interchange format) descriptions into it and asking it to paint pictures with the prompt “Bacon Nancy Pelosi”. Now, people are giving this thing the passwords to their bank accounts and it is sending printed letters to airlines to request refunds for unused in-flight Wi-Fi access and the “chat bots” are integrated into Snapchat where children are being groomed to interact with them. Next, Google will be replacing Google Search with the technology. It is May 13, 2023.
April 13, 2022 I’ve been reading about audio deep learning and have been surprised to see the use of the Python language, and also that after the usual prep process of classification and analysis, the next steps homogenize the audio then transform it into visual snapshots before it is loaded.
I think about olfactory deep learning where we bottle an isolated scent (sandalwood) and then convert it mathematically into a 2-d sine wave and feed that picture to the machine with our prepared analysis. “Machine, this is what sandalwood looks like.”
Years from now your cellphone has a “scent camera” that captures father in law flatulence, converts it into a picture of the stench, and reaches out to the cloud in a desperate attempt at isolating and identifying the Guinness and Maker’s Mark notes.
It might also measure components in your halitosis and offer up suggestions; “peppermint not spearmint, also remember to floss.”
With a large enough sample from a single person perhaps your phone could make more accurate meal recommendations. “Curry breath detected, chicken tikka masala recommended.”
We could also define scents like “athlete’s foot” “Democrat” and “boomer” and the way consumers smell could be used to create entirely new markets where we cater to, mask, or embrace them. “Girl, you smell BTS today.”
I also imagine a “history of scents in the cloud” like a collection of vacation, celebration, and other memorable scents. Stored for individuals in their scent wallets. Scanned by their cloud providers for illicit scents: ammonium nitrate, nitroglycerin, Drakkar Noir. ;–)
I guess what I’m trying to say is that modern machine learning appears to have only one sense: sight. So everything we feed into it has to be transformed into a picture. And that’s kind of silly. The inefficiency is galling and hilarious to me.
I have a feeling this is not how Shazam works but I’ll need to do some research to know for sure.
How Shazam works isn’t that dramatic, it is that the client creates a fingerprint of the downsampled input and then the server uses machine learning and compares the fingerprint against a library of other fingerprints to locate the culprit.
It actually works more like a real sensory organ using the nerves to transmit the information to the brain than that other strangeness I was reading about.
July 11, 2022 I spent the morning teaching a computer to design a car made of raw bacon. You’re welcome, world. I also taught it to make raw bacon boats because I am an artificial intelligence genius. :–)
September 03, 2022 An article entitled, AI Artwork Wins First Prize.
This reminds me of the rise of marketplaces for digital goods leveling the playing field (and diluting the market) for artists in music, books, movies (via Amazon’s upload anything marketplace), etc. But now the computers are doing it all.
I remember complaining in the nineties that pretty soon you would only hear music made with keyboards and no one would take the time to learn the guitar or drums etc.
Another wild time to be alive I guess.
This is another way of saying “in the nineties I could not appreciate the work and artistry that went into mastering the keyboard and making electronic music because I was ignorant”, by the way.
The way AI art is being created and presented here should have been unpacked more. The “artist” is sending words (and possibly accompanying art) into an art cow that masticated, digests, and excretes “the art”. The outcome is sometimes altered by the artist as well.
But I think that many artists who wanted to spend hours on a single work and produce something extremely valuable are watching another person send the words “raw bacon grandfather clock” into a laptop and comparing the output in a traditional art contest (and the market)
I also think of all of my own art that I sent to be recycled or destroyed after years of dragging it around the country hoping a market would appear for it. I created it in isolation, exposed it to a market that didn’t want it, and when I discovered how the professional art market really works I didn’t want to spend my life painting what I thought was throwaway corporate garbage in order to afford to make what I wanted. I also didn’t like other people or my self very much.
Now, my objective and outlook are collaborative. I’m not pretending the computer is “a bicycle for the mind” but more like an entity that I am learning with. It would be cool to ask it for “beautiful artwork” and get something that was both to be able to trust it and credit it for its own creation. But I wouldn’t try to enter it in a contest against humans. Or, to pass it off as my own creation and be like, “it’s like a collage, trust me”.
Also, my hope is that AI can take over menial tasks so that folks can free up time to do what they love. That doesn’t mean erasing whole sectors of markets so that the people who own the capital can own more.
February 02, 2023 This young filter on TikTok is making a lot of people sad. I guess it uses AI though so it’s not a reliable way to make you think about what you’ve done with your life and how a computer can inspire introspection.
It is a reliable way to make you hate the verve pipe tho ;–)
February 27, 2023 So these folks who were using AI to model molecules in hope of finding helpful drugs and they switched variables for toxicity and in under six hours it spat out over forty thousand variations of VX nerve agent.
And when it was made public they were invited to speak at the Convergence conference by the Swiss Federal Institute for Nuclear, Biological and Chemical Protection, Spiez Laboratory.
March 22, 2023 What's chilling is learning that the AI chat bot knows what Finding Nemo is. Maybe we should not be allowing “the company that sells you things while you are searching the Internet” to pioneer this stuff.
March 23, 2023 They are imagining these AI services doing a lot of work from simple formatting of a script to determining who gets hired to work on the films and which movies get the green light and which do not.
When you put enough people in a room does it become one of these “AI services”? How many people are needed in order to simulate something so homogenous. I remember in the nineties reading someone describe the process of selling music online as 'feeding music to the organism that returns money' in a very general way. The artist didn't want to learn more than they had to in order to make the music.
April 08, 2023 I wonder if any medium has had as convincing a method of deceiving people as the AI deepfake videos that appear to be on the way. And if so, what people did to cope with the possibility that so much of the media they consume could potentially be lies.
May 09, 2023 It’s kind of funny to imagine a world where the ML models are so advanced that captcha no longer work and we learn that because humans can not can solve them anymore because the machine has exceeded our ability.