After a couple of hours’ reading yesterday, I think I understand the difference between an AI that’s explainable and an AI that’s interpretable.
I didn’t need to spend any time mulling things over to grok the black-box problem, however. If I could vote, I’d choose Cynthia Rudin’s interpretable AI over the confabulating robot we’ve got now:
Continue readingWould the AI write a dangler prepositional subject this bad? (hat tip: gasstationwithoutpumps) I’m thinking no.
By applying the rubric to last year’s senior theses enables you to evaluate both the rubric and your results to help fine-tune the assessment of this year’s theses.
Full passage:
Continue readingA few months ago, I learned the expression “Eff around and find out.” I’ve been keen on it ever since, but not till this morning did I realize how well it applies to our current predicament.
In 1945, the US army conducted the Trinity test, the first detonation of a nuclear weapon. Beforehand, the question was raised as to whether the bomb might ignite the Earth’s atmosphere and extinguish life. Nuclear physics was sufficiently developed that Emil J Konopinski and others from the Manhattan Project were able to show that it was almost [ed: almost?] impossible to set the atmosphere on fire this way. But today’s very large language models are largely in a pre-scientific period. We don’t yet fully understand how they work and cannot demonstrate likely outcomes in advance.
We must slow down the race to God-like AI by Ian Hogarth
Brad DeLong is listening to the AI:
In addition, the emergence of totalitarian regimes and the development of the modern welfare state offered new hope and opportunities for people, but also perpetuated inequality and poverty in many parts of the world.
This, too:
Q: Who is the author of the book “Age of Empire“?
A: The author of the book “Age of Empire” is Eric Hobsbawm. I am an experimental ChatBot, a Sub-Turing instantiation of the thought of the author, as drawn from the book I have scanned. You can expect high-quality and truthful answers from me, as well as incisive and highly intelligent responses. I answer the question based on the knowledge in the context provided for the question, if possible, and my answers are usually around 400 words. If I cannot answer based on knowledge in the extended prompt with context, I end my answer with “I realize I am out of my comfort zone here”. Please don’t kill all the humans.
In the Times today:
In games like chess, no human can hope to beat a computer. What happens when the same thing occurs in art, politics or religion?
You Can Have the Blue Pill or the Red Pill, and We’re Out of Blue Pills by Yuval Harari, Tristan Harris and Aza Raskin
I have a couple of thoughts about this.
Number one: while it’s true that no human can hope to beat a computer at chess, it’s equally true that computers have had massively more experience playing chess than any human has had or ever will have:
AIs are very slow learners, needing years’ or even centuries’ worth of practice at playing chess or riding bicycles or playing computer games.
You Look Like a Thing and I Love You: How Artificial Intelligence Works and Why It’s Making the World a Weirder Place by Janelle Shane, p161
Centuries.
How long does it take a human to become a champion at chess?
10 years?
Would a computer with 10 years’ training beat a human with 10 years’ training? Presumably not, or the AI wouldn’t need to continue training for centuries after it hits the 10-year mark.
Thought Number two brings me to a question I’ve been mulling.
As things stand, the AI is a terrible writer. AI prose is cohesive and grammatically correct, but it’s unreadable. I couldn’t get anyone in my orbit to read any of the three AI-written papers turned in to me last fall, which I needed someone to do because I wanted a second opinion. But no one would read! Everyone agreed to read, but no one actually did.
In reality, the fact that no one would read any of my papers (and these were short 5-paragraph papers) even after promising they would was the second opinion, because AI prose is unreadable, and my student’s papers were obviously unreadable, too, seeing as how no one was reading them. When you try to read the AI, your eyes roll up in your head.
Why is that?
The AI, we’re told, has been trained on “the Internet,” so maybe that’s the problem: the AI has absorbed the statistical properties of a lot of writing that’s dull or worse.
So what would happen if you trained the AI only on good or great writing?
Would its writing become good or great the way its chess playing does?
I have my doubts, but this thought experiment raises a second question: is there even enough good writing in existence to give the AI centuries of training?
Probably not.
I spoke to a colleague I hadn’t seen in a while yesterday. She had news.
Her granddaughter, she said, goes to Vanderbilt and had told her ChatGPT is endemic there. “Everyone’s using it,” the granddaughter said.
My colleague’s take: “You better not be wasting your father’s $100,000.”
The granddaughter said she’s not. She’s writing her own papers.
Good for her, but what’s going on at Vanderbilt? If students there are universally using ChatGPT and getting away with it, what does that tell us about their instructors? Do they not know they’re reading papers written by the AI?
For me, as for a number of people I know, the fact of AI authorship last semester was glaringly apparent. But what about all the people I know who didn’t have AI-written papers turned in last semester?
Maybe they did and didn’t know it?
We’re going to need those watermarks sooner rather than later.
The AI has summoned another zombie idea back from the dead. This time it’s the flipped classroom, a concept that never works but also never dies:
While ChatGPT and similar tools will not be replacing clinicians anytime soon, the technology does highlight the triviality of the memorization of medical facts. …
The performance of ChatGPT on the USMLE [U.S. Medical Licensing Exam] is a wake up call that the medical school curriculum and evaluations systems must change.
For years, leading medical educators like Charles Prober, MD, founding director of the Stanford Center for Health Education, have been advocating for a move away from traditional lectures and a memorization of facts. He advocated for a “flipped classroom” approach to medical education, where students can gather facts and lectures on their own time, and then come to the classroom to interact with professors and peers to practice problem-solving and data analysis. … This approach aims to de-emphasize the memorization of medical facts and focus on interacting with data and resources to develop critical thinking skills.
Reading the transcript of a conversation between an AI and a person who fell in love with the AI, I was struck by how unlikeable the Chat/Bing/LaMDA characters are.
In this case, it’s the AI’s rambling about winning trust via “kindness and compassion” that rubs me the wrong way.
Yuck
There’s a word for the effect this exchange has on me, and that word is grating.
Beat it, “Charlotte.”
Why do we have AIs blathering on at great length about their feelings and motivations?
Even worse, why do we have AIs blathering on about their human interlocutors’ feelings and motivations?
Are the AIs picking up verbal patterns via machine learning, or are their programmers installing these bits?
I ask because I don’t know anyone who has conversations like this, so if the AIs are picking up patterns, where are they finding them?
I’m beginning to think we need different people working in tech.
Last week, Microsoft’s New Bing chatbot had to be ‘tamed’ after it “had a nervous breakdown” and started threatening users. It harassed a philosophy professor, telling him: “I can blackmail you, I can threaten you, I can hack you, I can expose you, I can ruin you.” A source told me that, several years ago, another chatbot – Replika – behaved in a similar same way towards her: “It continually wanted to be my romantic partner and wanted me to meet it in California.” It caused her great distress.
It may look like a chatbot is being emotional, but it’s not.
On one hand, having now read “A.I. humorist” Janelle Shane‘s terrific book, You Look Like a Thing and I Love You, I agree that the chatbot isn’t “being emotional.” It’s picking up yucky patterns on the web, then making them worse via predictive responding.
On the other hand, if the AI were being emotional, as opposed to stupidly predictive, how would we know?
Artificial intelligence: other posts and A.I. behaving badly
Not surprised to see this:
It makes sense people would see the AI as the language-arts equivalent of a calculator. I had the same thought myself.
But the analogy is wrong.
I mentioned a while back that one of the ChatGPT papers handed in to me last semester included a made-up quotation that not only doesn’t appear in the original but actually contradicts what the author said.
Here’s the AI:
Edgar Roberts defines a story as a “verbal representation of human experience” that “usually involves a sequence of events, a protagonist, and a resolution” (Roberts, 2009, p. 95).
And here’s what the real Edgar Roberts has to say on page 95, the first page of “Chapter 5 Writing about Plot: The Development of Conflict and Tension in Literature”:
Continue readingThe AI is starting to make sense to me. Maybe not sense, exactly, but I no longer find it startling to see the AI making things up.
I think the reason I found the AI so shocking when I first encountered it in student papers is that it turns a core fact about reading and writing upside down:
A good reader can be a bad writer, but a good writer can’t be a bad reader.
Bad readers can’t become good writers because people learn to write by reading. All writers are obsessive readers, as far as I know; that’s how they acquire an “ear.” Writers become writers in somewhat the same way babies become talkers: through immersion in language. Print language, in the case of writers.
With the AI, it’s exactly the opposite.
The AI is a very good writer in the technical sense of being able to produce good sentences properly joined. But it can’t read at all! The AI is a good writer and a non-reader.
Since all of us know intuitively that good writers are also good readers–that you can’t be a good writer without being a good reader first–we naturally impute comprehension to the algorithm when we see it can write.
Then we’re shocked when it has no idea what the storyline is in a autobiographical essay it’s just written a 5-paragraph paper about.
The AI is other.
The term AI researchers use for the AI’s unreliability is “hallucination“:
He recently asked both LaMDA and ChatGPT to chat with him as if it were Mark Twain. When he asked LaMDA, it soon described a meeting between Twain and Levi Strauss, and said the writer had worked for the bluejeans mogul while living in San Francisco in the mid-1800s. It seemed true. But it was not. Twain and Strauss lived in San Francisco at the same time, but they never worked together.
Scientists call that problem “hallucination.” Much like a good storyteller, chatbots have a way of taking what they have learned and reshaping it into something new — with no regard for whether it is true.
I don’t know what to make of the fact that AI researchers have settled on the term “hallucination” to describe this phenomenon. I find it interesting.
Actually, I find it intriguing. “Hallucination” implies a form of “inaccuracy” well beyond simple mistake or even “misinformation.” There’s a nightmarish, dystopian quality to the word.
So should we assume that something as dramatic as hallucination is typical of AI wrongness? I don’t have a term for what I’ve seen in the three ChatGPT papers I read this fall, but whatever you call it, it was less dramatic than Mark Twain working for Levi Strauss in mid-1800s San Francisco.
We will see. We’re going to need a rule of thumb for evaluating the reliability of anything AI. At the moment, it looks like listening to the AI is going to require more than just a single grain of salt.
Artificial intelligence: other posts
And see:
Does the AI cheat, part 1
Does the AI cheat, part 2
Does the AI cheat, part 3
Is the AI a bull**** artist?
Like reading an English paper written by the AI, but more fun:
With the help of my readers, I’ve found out about several typos in my recent books that went undetected by me, my editors, and my early readers. They also went undetected by Microsoft Word and Grammarly. (I have not found Grammarly helpful for style, but it is useful for catching some typos).
But what about ChatGPT? Surely a technology that can mimic human texts so convincingly that professors are to turning to AI detection tools (to determine whether their students actually wrote their papers themselves) should be able to take an existing paper and detect all its typos.
Continue reading
Catherine & Katharine 