What Is The Monster Thinking?
Decoding the vectors that AI models use to represent abstract concepts; a chick from French Lick, Indiana; and a guide to fact-checkers.
The Week
Well, we did it. There’s no going back. We are now gleefully building monsters. I don’t mean monsters in a pejorative sense. These might be very nice, helpful monsters. (I, for one, welcome…) By monsters, I only mean powerful creatures that are beyond our control or understanding. Merely that.
But there is a glimmer of hope. One company may have developed a technique for learning what’s going on inside the monster’s brain. I’ll tell you about that in this week’s post.
It’ll be a short post, because I’m busy packing up for a week on the road. I am adding some links to the Blogroll and inflicting three more limericks, though.
Oh, the picture is from my morning walk. The cow was lonely until the sheep arrived, and now they are inseparable. Here she is keeping watch on her flock.
Quote of the Week
“We create monsters and then we can’t control them.
— Joel Coen
Inside the Monster’s Brain
Generative Artificial Intelligence systems don’t model the world, or language, the way humans do. They can (sometimes) carry on coherent conversations, answer questions, apologize for their errors, and in doing all this they can sound frighteningly intelligent. But the truly frightening thing, to me, is that we don’t know what’s going on in there. The technology of machine learning allows AI systems to construct complex models, but models of what? We’ve built systems that we are able to train by rewarding them for getting the right answers, but we have lost the ability to understand how they get those answers. We forget to tell them to show their work.
As reported by Timothy B. Lee in the blog Understanding AI, Anthropic, which describes itself as Anthropic is an AI safety and research company, has made a breakthrough in our ability to understand how Large Language Models actually work.
LLMs represent linguistic elements (an unattractive phrase I’m using to avoid saying words, which would be incorrect) as lists of thousands of numbers. These lists are vectors in a “word space” (still incorrect) of thousands of dimensions.
These vectors represent meaningful parts of words rather than words per se, but I give up, I’ll just say words from here on.
Words with similar representations are close in this vector space, and thus close in meaning. But there’s the rub: whose meaning? The LLM’s meaning, of course. The words are close in the multi-dimensional model the LLM has developed, the model of the linguistic data on which it has been trained, the model that is its efficient engine for generating acceptable output. And we don’t know what that model is, although we can be pretty sure it isn’t our model. It isn’t ourmeaning.
What Anthropic has shown is that it is possible to use the mathematical nature of the representation to extract meaningful (to a human) features, which can be combined in meaningful ways just by doing arithmetic on the features.
From Understanding AI:
“[I]f you subtract the vector for Paris from the vector for France, you get a new vector that maps any country to its capital. Calculate Paris minus France plus Italy and you get a vector for Rome. Compute Paris minus France plus Japan and you get a vector for Tokyo.”
Why is this important?
“The Anthropic team identified features associated with deception, the creation of bombs and bioweapons, racism, and other undesirable behaviors. Anthropic’s research could lead to new tools to detect model misbehavior — or perhaps even to prevent it altogether.” [Timothy B. Lee, Understanding AI.]
Verse
If you’ve been following this blog for over a year, you may recognize some of these. I just feel a need to put some verse in each post. I’ll try to include a new one from time to time. For a while it’s just going to be limericks, but eventually I’ll get to the sonnets and villanelles and whatnot.
Three Limericks
A chick from French Lick, Indiana
Once got sick from a green-ish banana
That she bought in South Bend
And that is the end
Of this nugget of Hoosier arcana.
The cowboy star Hopalong Cassidy
Made this claim of uncertain veracity:
My Topper is bigger
Than Roy Rogers’ Trigger,
And Roy said that Hoppy’s flops flaccidly.
As we broil in the burbs or the Barrio
It’s too late to just wait till tomorrow yo.
While the pols was debatin’
We’re left to our fate in
This climate-change worst-case scenario.
Of Interest
When I was editor of Dr. Dobb’s Journal, we had a section in the back of the magazine titled “Of Interest.” Here are some things that I think you might find Of Interest.
Blogroll
I’ve added another important fact checker, The Trusted News Initiative, which targets disinformation in real time.
FACT CHECKERS
Judd Legum’s Popular Information blog/newsletter is not merely a fact-checker, but an exercise in investigative journalism. He recently took on the disinformation regarding a hot topic: Donald Trump’s 34 felony convictions.
I try to avoid politics on this blog, but I do get annoyed when people just blatantly lie for political gain. I hope always to be on the side of truth.
Here are some of the fact-checking sites I use:
The Trusted News Initiative is a partnership, founded by the BBC, that includes organisations from around the globe including; AP, AFP, BBC, CBC/Radio-Canada, European Broadcasting Union (EBU), Financial Times, Information Futures Lab, Google/YouTube, The Hindu, The Nation Media Group, Meta, Microsoft, Thomson Reuters, Reuters Institute for the Study of Journalism, Twitter, The Washington Post, Kompas — Indonesia, Dawn — Pakistan, Indian Express, NDTV — India, ABC — Australia, SBS — Australia, NHK — Japan.
TNI members work together to build audience trust and to find solutions to tackle challenges of disinformation. By including media organisations and social media platforms, it is the only forum in the world of its kind designed to take on disinformation in real time.
Snopes is the granddaddy of fact-check sites. It got its start in 1994, investigating urban legends, hoaxes, and folklore. Now it’s the oldest and largest fact-checking site online, widely regarded by journalists, folklorists, and readers as an invaluable research companion.
Media Bias / FactCheck is your go-to site to fact-check the fact-checkers. From CNN to Breitbart and even browser extensions like Newsguard, MBFC rates them all.
FactCheck is a non-partisan fact-checking website that focuses primarily on US politics. It is a non-profit project and regularly debunks claims by politicians and viral fake news.
Lead Stories focuses on trending stories, images, videos, and posts with false information. It includes fact checks, analysis, and corrections. The site also provides fact-checking services to social media companies TikTok and Facebook.
SOURCES ON ARTIFICIAL INTELLIGENCE
One Useful Thing
Wharton professor Ethan Mollick shares research-based views on the implications of AI.
The AI Edge
A daily newsletter to help you keep up with the latest news and trends.
Big Technology
A newsletter about big tech and society by independent journalist Alex Kantrowitz.
Creators’ AI
AI insights, tools, guides for creators and entrepreneurs.
AI Supremacy
News at the intersection of Artificial Intelligence, technology, and business. Includes Op-Eds, research summaries, guest contributions, and info on AI startups, by Michael Spencer.
Artificial Intelligence Made Simple
AI made simple by Devansh.
AI: A Guide for Thinking Humans
A blog about interesting developments in artificial intelligence by Melanie Mitchell, Professor, Santa Fe Institute.
marcwatkins
How generative AI is impacting education, by Marc Watkins, Academic Innovation Fellow, Director of the Mississippi AI Institute, Lecturer of Writing and Rhetoric at the University of Mississippi.
The AI Optimist
Exploring the possibilities of AI, against the drawbacks. By Declan Dunn.
Machine Society
Mike Elgan’s technology newsletter, formerly known as “Mike’s List.”
Ahead of AI
Machine Learning & AI research by Sebastian Raschka.
Mark Watson’s AI Books and Blog
Read his books for free online.
GOOD GUYS
Doctors Without Borders
Every day, Doctors Without Borders teams deliver emergency medical aid to people in crisis, with humanitarian projects in more than 70 countries.
World Central Kitchen
WCK is first to the frontlines, providing meals in response to humanitarian, climate, and community crises.
FRIENDS AND ASSOCIATES
Kent Beck’s advice for geeks
Tales from the Jar Side
Bookshop.org
Pragmatic Bookshelf
My Day Job
I edit books on technology for The Pragmatic Bookshelf. Although I’m working on the road for the next week, mostly I work at home in Oregon. But PragProg authors travel the world, sharing their knowledge at conferences, meetups, and private training sessions. Here’s what they are up to now.
ICYMI
Thanks for reading. You can read all the back issues of Swaine’s World at my blog home.
Coming Attractions
In the coming weeks, more Swaine’s Flames flashbacks, Dirt Road Diaries, bulletins from the AI revolution, tech history, books, and random verse.