#generative artificial intelligence

I'm on a 20+ city book tour for my new novel PICKS AND SHOVELS. Catch me in CHICAGO with PETER SAGAL on Apr 2, and in BLOOMINGTON at MORGENSTERN BOOKS on Apr 4. More tour dates here.

A law professor friend tells me that LLMs have completely transformed the way she relates to grad students and post-docs – for the worse. And no, it's not that they're cheating on their homework or using LLMs to write briefs full of hallucinated cases.

The thing that LLMs have changed in my friend's law school is letters of reference. Historically, students would only ask a prof for a letter of reference if they knew the prof really rated them. Writing a good reference is a ton of work, and that's rather the point: the mere fact that a law prof was willing to write one for you represents a signal about how highly they value you. It's a form of proof of work.

But then came the chatbots and with them, the knowledge that a reference letter could be generated by feeding three bullet points to a chatbot and having it generate five paragraphs of florid nonsense based on those three short sentences. Suddenly, profs were expected to write letters for many, many students – not just the top performers.

Of course, this was also happening at other universities, meaning that when my friend's school opened up for postdocs, they were inundated with letters of reference from profs elsewhere. Naturally, they handled this flood by feeding each letter back into an LLM and asking it to boil it down to three bullet points. No one thinks that these are identical to the three bullet points that were used to generate the letters, but it's close enough, right?

Obviously, this is terrible. At this point, letters of reference might as well consist solely of three bullet-points on letterhead. After all, the entire communicative intent in a chatbot-generated letter is just those three bullets. Everything else is padding, and all it does is dilute the communicative intent of the work. No matter how grammatically correct or even stylistically interesting the AI generated sentences are, they have less communicative freight than the three original bullet points. After all, the AI doesn't know anything about the grad student, so anything it adds to those three bullet points are, by definition, irrelevant to the question of whether they're well suited for a postdoc.

Which brings me to art. As a working artist in his third decade of professional life, I've concluded that the point of art is to take a big, numinous, irreducible feeling that fills the artist's mind, and attempt to infuse that feeling into some artistic vessel – a book, a painting, a song, a dance, a sculpture, etc – in the hopes that this work will cause a loose facsimile of that numinous, irreducible feeling to manifest in someone else's mind.

Art, in other words, is an act of communication – and there you have the problem with AI art. As a writer, when I write a novel, I make tens – if not hundreds – of thousands of tiny decisions that are in service to this business of causing my big, irreducible, numinous feeling to materialize in your mind. Most of those decisions aren't even conscious, but they are definitely decisions, and I don't make them solely on the basis of probabilistic autocomplete. One of my novels may be good and it may be bad, but one thing is definitely is is rich in communicative intent. Every one of those microdecisions is an expression of artistic intent.

The complexity of AI systems lies in their architecture, which is a labyrinth of algorithms, data structures, and neural networks. These components are meticulously designed to mimic cognitive functions. Yet, they are bound by the constraints of their programming and the quality of their input data. The notion that AI can autonomously solve any problem is a fallacy. It is akin to expecting a Swiss Army knife to perform the specialized tasks of a surgeon’s scalpel. Each tool, or in this case, each AI model, has a specific purpose and context in which it excels.

AI’s decision-making process is a cascade of probabilistic inferences, derived from training data. This process is not inherently intuitive or adaptable beyond its training scope. The protocols that guide AI are deterministic, meaning they follow a predefined path unless explicitly programmed otherwise. This rigidity is both a strength and a limitation. It ensures consistency but lacks the flexibility of human reasoning.

Moreover, AI’s reliance on data is a double-edged sword. While vast datasets can enhance its learning, they also introduce biases and errors. The GIGO principle—Garbage In, Garbage Out—remains a pertinent concern. AI systems are only as reliable as the data they are fed. This dependency underscores the importance of data integrity and the potential pitfalls of over-reliance on AI without human oversight.

In practical applications, AI is a tool that augments human capabilities rather than replaces them. It excels in tasks that require pattern recognition and data analysis at scales beyond human capacity. However, it falters in areas requiring empathy, ethical judgment, and contextual understanding. The complexity of human experience cannot be distilled into binary code or algorithmic logic.

This makes me a gardener.

I'm making gardening accessable.

The food I stole is rightfully mine to sell and profit off of. The people who put the food in the stores should have thought about that before putting food in stores where I can just steal it. Food is for everyone and I am just making gardening accessable.

I am the best gardener ever. All other gardeners are gatekeeping food.

And the problem is that statistically appropriate =/= factually correct.

There is no actual intelligence there so if your going to use it you need to bridge those gaps yourself. You need to be able to recognize when the AI's answer is wrong even if it sounds appropriate.