Don't Trust the Internet! Another Watching-Breathless-"AI"-Boosters Edition :: A Morning Rant

Modern Advanced Machine-Learning Models: useful interns and unreliable bullshitters. Beware the illusion of competence! Limitations really matter. Approach them all like the bluffing shitposters that they are! And recognize that you really need a real domain-knowledge expert to check them!…

Share

One of the things about the Internet is that it is filled with people making assurances that are simply not true.

Here is Vox’s Kelsey Piper:

Kelsey Piper: China’s new AI agent Manus Calls Its Own shots: ‘Hey readers, Modern large language models are really good at a lot of tasks, like coding, essay writing, translation, and research. But there are still a lot of basic tasks, especially in the “personal assistant” realm, that the most highly trained AIs in the world remain hopeless at. You can’t ask… “order me a burrito from Chipotle” and get one, let alone “book me a train from New York to Philadelphia”…. That such “AI agents” sometimes work, sort of, is about the strongest thing you can say for them right now… <[Sww.vox.com/future-pe...](https://Sww.vox.com/future-perfect/403896/artificial-intelligence-china-manus-chatgpt-openai-google-privacy)>

Share Brad DeLong's Grasping Reality

The second two-thirds is correct, but they do not “sometimes” work. Rather, they rarely work:

@TetraBoston

Leave a comment

The first third of the quotation, however…

  • If you are no programmer at all, MAMLMs—Modern Advanced Machine Learning Models—will not help you write a program. If you are a bad programmer, MAMLMs can make you an adequate one in terms of getting you 90% of the way there via reminding you of exact syntax, as long as you are good enough to recognize where it has gone wrong from the error messages. If you are a good programmer, it is like having an intern looking at the documentation they do not understand whispering in your ear.

  • If you are no essay writer at all, MAMLMs will produce an essay at the level of an Internet shitposter—moreover, one that is bullshitting, and not actually transmitting ideas but rather gesturing in the direction in which ideas might be found, in the hope that the reader will fill in the gaps. If you are an OK essay writer, MAMLMs will give you starting points and paragraphs that are wrong in ways that you can try to correct. If you are a good essay writer, MAMLMs may produce useful verbiage in the sense that “explain it to your rubber duck” is a useful psychological trick to use on your recalcitrant brain; otherwise, you will spend more time wincing and rewriting than you would writing from scratch yourself.

  • If you are no translater or an OK translator, MAMLMs will surely in the first case and probably in the second case do much better than you would. The problem if you are an OK translator is that you will then think it is a better translation than it actually is. If you are a good translator—what it produces will be flat, and miss nuances, but it will definitely provide a starting place which you can improve.

  • If you are no researcher, then no, no, no, NO, NO, NO!!!!!! Go to Wikipedia instead. If you are an OK researcher, the additional linguistic flexibility of the vector-database map underpinning MAMLMs may give you an edge relative to keyword-bound google searches, especially in this age of SEO not yet tuned to MAMLMs. If you are a good researcher, it will wind up wasting your time.

  • If you are lucky, or if you are a superb “prompt engineer”—if you have somehow become adept at poking the MAMLM with your prompt in such a way as to direct it to the pools of vector representations of the Internet in which reliable information is to be found—you can get amazingly good results, truly. But what is high-class prompt engineering changes from month to month. And, in general, spending time on prompt engineering is of the “Clever Hans” nature: it spits out text until you recognize the right (or a right-enough) answer.

  • If your MAMLM is trained on a trusted, reliable, structured database (or maybe someday soon on an unstructured one?)—rather than trained to be an Internet shitposter—it can be a very useful and reliable natural-language front end.

  • If what you are interested in is not in producing high-quality first-class writing, but rather good-enough writing to serve purposes of what Marion Fourcade term “ritual”—what us non-sociologists call “boilerplate”—then MAMLMs may well be a golden tool, saving you huge amounts of time in creating documents that are really valuable only for the actions that their magic phrases trigger others to undertake.

But to say, without qualification, that “modern large language models are really good at a lot of tasks, like coding, essay writing, translation, and research”—that is not just creating a map that is not the territory, that is giving your reader a map of Kalamazoo and claiming it is a good guide to Long Island.

Now Kelsey might respond that I am not the typical user of MAMLMs: That I am a highly-trained consumer and producer of intellectual excellence in written English prose. That what I see is valid for me. But that for your typical guy who just wants to punch the clock and create documents that are respectable—it is a substantial leg up for everyone who did not go to an élite university, or if they did did not pay attention in their classes.

Give a gift subscription

References:

I fed ChatGPT 4 the two URLs <https://www.economist.com/by-invitation/2024/09/04/large-language-models-will-upend-human-rituals> and <https://www.vox.com/future-perfect/403896/artificial-intelligence-china-manus-chatgpt-openai-google-privacy>, asking it for University of Chicago author-data citations with attached live and working URLs.

The two entries above are what it gave me. It did not do so hot.

The actual citations are:

Capisce?

You are warned.

Leave a comment

Subscribe now

If reading this gets you Value Above Replacement, then become a free subscriber to this newsletter. And forward it! And if your VAR from this newsletter is in the three digits or more each year, please become a paid subscriber! I am trying to make you readers—and myself—smarter. Please tell me if I succeed, or how I fail…