Ask A Genius 891: Turing Test Shorthand
Author(s): Rick Rosner and Scott Douglas Jacobsen
Publication (Outlet/Website): Ask A Genius
Publication Date (yyyy/mm/dd): 2024/02/05
[Recording Start]
Scott Douglas Jacobsen: We just talked about the Turing test, a shorthand. I want to take a different angle based on the intuitions around ChatGPT. I’ve used it and learned how to work with it to do all sorts of things that can be a helpful tool. So, we were talking about the Turing test and fooling a human interlocutor. So, I want to take the angle of output in the sense of a strict communication theory, process-input-output or input-process- output and that is if you take any of the production of ChatGPT that is good, it can look like an educated adult person wrote it with English as a first language. However, they might be stilted, and it’s a very PC and PG language. At the same time, it’s good, and that makes me think about the missing parts, not of the output, which is all there, but of the process. So, we can put a bunch of stuff through our sensory systems and our language processing; we can produce a similar type of thing if we have a mind and a computer and a keyboard to type that stuff in or a microphone to speak into so that they can be speech to text so it produces something similar. The output could be identical to ChatGPT, or ChatGPT could be identical to that output, yet the paths of that input might be vastly different in terms of process.
In some ways, the input is different because you’re taking a multimodal form of information to get to your language production. ChatGPT takes the simple text and does a statistical token-based analysis to get that output.
Rick Rosner: Let’s talk about that because that’s the essential difference between human written output and ChatGPT. When you talk about multimodal, when we expand that, that’s the vital difference between human written output and ChatGPT. So, when you say multimodal, you mean we’ve got the active workspace according to workspace Theory and what we’ve been talking about for years and didn’t know to call workspace Theory. We kind of independently came upon it, but everything worthy of consideration in your immediate circumstance and by worthy of consideration like important enough and novel enough that it impinges on your conscious mind, like walking, breathing; your body can handle that for the most in most context semi unconsciously. Everything that demands your attention is in your conscious arena, and to a great extent, it’s your conscious arena that determines what your verbal output will be. By multimodal, you mean every aspect of your consciousness, what’s happening in your immediate environment, including all your sensory information and all the relevant memories you’re thinking is dredging up. So, you’re getting all sorts of inputs that feel like reality and your consciousness, and it’s still all Bayesian.
Bayesian is splitting the world up into subsets and making predictions based on which subsets your inputs are found to be in. I developed a Bayesian system of catching people with fake IDs. I take everything that I thought was wrong about somebody showing me an ID in a bar and everything that was right, and based on all that, that would put them in one of a number of subsets, but really, it was too complicated to consider each of the thousand or so different subsets individually. So, I assigned points for everything somebody got wrong, based on a Bayesian waiting of how bad it was to get something wrong. Not knowing their zodiac sign; was pretty bad; not knowing the year they were supposed to have graduated high school; was not as bad as not knowing their sign or misspelling their name; very fucking wrong. So, Bayesian is a probabilistic predictor based on accumulated data and weighting of that data. ChatGPT does that when putting together written words, but it’s not multimodal.
We’re Bayesian with all our experience. All our expertise gets weighted and evaluated, and used to predict how we should operate the car we’re driving and what words should come next based on our expertise. ChatGPT has yet to experience it. What it has is Bayesian weightings of a billion samples of writing and based on the probabilities of what words come after what other words, ChatGPT says the most probable following words or following sentence should be something like this because if you want ChatGPT to write 3,000 words on the Treaty of Versailles, ChatGPT has encoded maybe 50,000 written passages about the Treaty of Versailles and can weigh the various combinations of words to see what combinations of words would be the most probable 3,000-word article on the Treaty of Versailles but has no knowledge of what that treaty means or has no real-world knowledge of anything.
So, you could say ChatGPT is unimodal, only having Bayesian weightings of verbal samples with no context, and multimodal is our experience of the world with its multiple Bayesian weightings of what’s important, what needs to be considered and what we need to think about. Is that an adequate differentiation?
Jacobsen: Yeah, it speaks to the early period of this development, and so any comparison could only be on the surface, and that surface is as thin as output.
Rosner: Yeah, I mean we do, do something at some level similar to ChatGPT, like when we’re quoting something ‘when in the course of’ and then what pops up is ‘human events’ is the most probable following couple words because it’s I think is US Declaration of Independence. So, there’s a little bit of ChatGPT there where there’s some weighting based on our experience of history of having taken a history class. Still, there’s also just what’s the most when after those the most likely following couple words, if you Googled it, it would autofill; if you typed in when in the course of… it would autofill human events.
[Recording End]
License
In-Sight Publishing by Scott Douglas Jacobsen is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License. Based on a work at www.in-sightpublishing.com.
Copyright
© Scott Douglas Jacobsen and In-Sight Publishing 2012-Present. Unauthorized use and/or duplication of this material without express and written permission from this site’s author and/or owner is strictly prohibited. Excerpts and links may be used, provided that full and clear credit is given to Scott Douglas Jacobsen and In-Sight Publishing with appropriate and specific direction to the original content. All interviewees and authors co-copyright their material and may disseminate for their independent purposes.
In-Sight Publishing 