Time-stamp: <2023-04-23 19:50>
I was called a »doomer« when it comes to Large Language Models (LLM). I have a master's degree in Computational Linguistics and spent the last 20 years of my professional career in Enterprise Search. I will have to explain to customers why replacing an old-school inverted-index-based search solution with an LLM will not solve their knowledge discovery challenges in the near and mid-term future. An LLM will make up a cantina menu instead of delivering today's actual menu from the corporate intranet if it's not supported and/or reigned in by an old-school search engine, or by -- gasp -- human input.
Recently, I was trying to remember what energy source powers the starships in Kurt Vonnegut's 1959 novel Sirens of Titan. I thought this might be a question ChatGPT could have an answer to. And it had. With utter conviction, it proclaimed:
In the novel "The Sirens of Titan" by Kurt Vonnegut, the space ships are powered by a substance called "chrono-synclastic infundibulum."
Followed by a twice-as-long disclaimer on this substance being entirely fictional.
It is also entirely wrong. I told ChatGPT I'm not satisfied with its answer, but it failed to come up with a different one, it just worded the same one differently.
Most likely, the model never ingested the novel. It is copyrighted and hidden from royalty-free access even though written 64 years ago. What the model did ingest is what's available for free on the web, including articles discussing the novel. Chrono-synclastic infundibulum is a core idea in Vonnegut's story. The ships, however, are powered by Universal Will to Become, as Google's autocomplete helped me remember without skimming the book again.
SALAMI¹ are freeloading the work of millions of humans that is available openly on the internet. Undoubtedly to make a profit in the near future.
I heard from friends and colleagues about them now using ChatGPT to aid with programming tasks, thereby replacing Programming by Stackexchange. But are they? Basically, they are replacing Programming by Stackexchange with Programming by Stackexchange via ChatGPT. ChatGPT profits from the work of millions of programmers who actually read library documentations and programming language guides, creating easily digestible tidbits of information ideally suited for ingestion by a parasitic LLM.
Stackexchange thrives on the merit-incentivized goodwill of fellow programmers to answer questions about concrete challenges. If users largely switch to ChatGPT for those answers, this well will run dry, and it will destroy the parasite with it. When humans stop producing the input an LLM needs to create a functioning model, the SALAMI rots. And the more the web will be filled with regurgitations of the stochastic parrot, the more everything we rely on today for truthful information -- will rot with it.
Walter Krämer's decades-old bestseller »How to Lie Using Statistics« could be the tag line for LLMs.
We can probably agree that trying to model natural languages top-down using Chomsky-style phrase structure grammar didn't produce any results usable for automation. There isn't a half-way complete PSG even for a language most linguists speak: English.
Using local grammars, the approach of modeling languages bottom-up, is much more promising and has been applied for subsets of language successfully, but failed at large because the tedious work cannot be outsourced to unskilled, exploited labor in developing countries, and, as my professor used to put it, »linguists are lazy.«
So now we're using statistics, training the models on content without honoring the sources, trying to fix inherent flaws or predetermined political correctness issues by using the intelligence of exploited labor. It's interesting to me that copyright does not seem to be a problem here, when corporations violate it instead of individual consumers.
A system that produces grammatically correct language and has no clue about its meaning, is that what we wanted? What is the problem that it solves again? We cannot trust it to answer questions correctly. It's inherent in the system that it »hallucinates« statements. On its own it's useless as an information source.
Conversational interfaces? Maybe. If one is willing to wait orders of magnitude longer for answers while a computationally extremely expensive process interprets the question and produces a result using data from a trustworthy source. At least the consumed CPU cycles produce something, unlike crypto. But what problem does this solve, really? A good suggest is still way quicker for information retrieval than a chatbot, while using way less ressources.
Producing text in the style of whatever: why?
An LLM can copy Stephen King's writing style and possibly write a story about a demonic SUV killing people. It might contain current trivia to make it relatable, as in a King novel. But would it, as »Christine« did, convey a message of teenage angst, of bullying and abuse, first love, coming of age, and tragedy? The things that made this novel worth reading to 14-year-old me? And if it wouldn't: Why bother?
An LLM cannot invent chrono-synclastic infundibulum. Maybe it can derive from the original, but, again, to what end? The story that needed it does not need to be written again. Vonnegut wrote it, back in 1959. And it is still holding up.
¹ SALAMI: Systematic Approaches to Learning Algorithms and Machine Inferences, a term coined by an Italian politician and popularized by Computational Linguist Emily Bender to replace the misleading »AI«
----------
✍ Wolfgang Mederle CC BY-SA 4.0
✉ <madearl+gemini@mailbox.org>
language: en
date: [2023-04-23 Sun]
tags: AI, IT, SALAMI