Yes, this is very comprehensive, but only missing two things: Intelligence (IBM Mainframe ), Lotus H.A.L. ( Lotus 1-2-3 heuristic language processor ). ELiza came from LISP, MIT.
RACTOR wrote a book, "The policeman's beard is half constructed."
I fired up RACTER but I couldn't get it to do anything half as interesting as MegaHAL, whomst once told my brother, "Pop is just water with a good director"
Long before I ever encountered Richard Wallace's AIML (which was not actually a joke, unfortunately), I wrote some joke web pages to parody VRML, announcing "AIML: Artificial Intelligence Marketing Language", by taking a couple articles about VRML, strategically replacing "VR" with "AI", and tweaking just a few words, the point being that the VRML hype back then was as bad as the AI hype right now. I don't have a very high opinion of VRML obviously! They didn't bother integrating it with an extension language, so it was useless.
To go along with the AIML parody, I also formatted as html and published an actual AI research paper from a company called Roanoke AI Laboratories, Inc., published in The Journal of Theriogenology. (NSFW, unless you work in the Artificial Insemenation industry!)
The author Jim Crump, Jr. actually found it and sent me email thanking me for publishing his AI paper on my web site, because his own publisher would not allow it due to copyright restrictions.
Ahh VRML was pretty cool. The problem was that the VRML viewer plugins were slow, needed more RAM than anyone had, or needed to run Java and so you'd let the 486 kick that off and go get a snack and come back and it would still be thrashing the disk to swap hard enough to have a chance to start anything. Pity that they didn't get some demoscene guys to make an efficient implementation.
I had this ancient program called Virtual Homespace Builder https://archive.org/details/hmebuilder that was like an early Sketchup sorta thing and it would render VRML output, but I can't say as I ever managed to publish anything with it, being a young whelp at the time.
If VRML had been fast and efficient, something like Second Life / Metaverse might have emerged there...
The disqualifying problem with VRML was that it was not extensible, and at the time it had not integration with JavaScript, Python, Lua, TCL, or even its own bespoke scripting language, and absolutely no plans to do so, so that made it TI (Turing Incomplete) and FU (Fundamentally Useless).
At the time, I had been working extensively with Bounce / Body Electric, which was developed by Chuck Blanchard at VPL Research, originally used for VR simulations with the DataGlove and Eyephone and a pair of SGI workstations to render each eye.
"Visual Programming Language" was not an afterthought nailed on the side, it was what the name of the company VPL stood for.
>VPL Research was one of the first companies that developed and sold virtual reality products. It was founded by computer scientist Jaron Lanier in 1984. "VPL" stood for "Virtual Programming Languages". In 1990, VPL Research filed for bankruptcy and in 1998 all of its patents were bought by Sun Microsystems.
>VPL's funding came in part from Marvin Minsky.
Funny that you would mention it: Virtual Homespace Builder was precisely what I was making fun of, in fact that was what the article I was parodying was about, almost word for word - I just went M-X replace-string Home => Head and then went from there:
>AI Labratories has entered into alliances with the likes of Microsoft, Apple, Intel, Creative Labs, NewTek, and Caligari to foster acceptance of the format. One of their partners, SexPression, offers the first AIML authoring program, Virtual Head Space Builder (VHSB) 1.0 Academic Edition ($495) for Windows. SexPression offers the final beta version for $49 and a paradigm-limited final beta via their Web site free of charge.
>Designed to accommodate nonprogrammers, VHSB creates AI Web sites, or "head spaces", that can be saved as AIML files or in SexPression's proprietary FASL file format. In addition to realtime interactive Artificial Intelligence, sites can include text, audio (WAV and RealAudio), MIDI, video, animation (AVI and FLI), cubes, spheres, pyramids, conic sections, and checkerboards -- in fact, SexPression suggests that VHSB makes a good expert system authoring tool as well. Any part of a knowledge base environment can act as a URL link, and evaluatable function modules can be associated. Head spaces can be opinionated with moods, including recurring and unpredictable swings, and with transparent prejudices. The software comes bundled with hundreds of clip-opinions, goals, interests, paridigms, taxonomies, ontologies, excuse templates, animal guessing trees, todo lists, top ten lists, hot lists, and shit lists; and custom content can be imported easily. A Virtual Head Space Evaluator is included.
Just being fast and efficient wasn't enough. It had to be TC (Turing Complete). No matter how fast and efficient it was, it was still FU: Fundamentally Useless.
Body Electric / Bounce STARTED with a fully Turing Complete, extremely powerful and fun to use Visual Programming Language, and then added Swivel 3D skeletons, networing with UDP packet packers/unpackers and transmitter/receiver modules, Swivel 3D tree syncing over Ethernet two one SGI workstation per eye for rendering to the Eyephones, many different kinds of input and output device drivers for the Eyephones (like a module to configure how far apart your eyes are to adjust the two 3D renderers), Dataglove, Convolvotron 3D sound rendering device, MIDI, Flock of Birds 6 degree of freedom motion tracker, etc.
While VMRL was not only missing the VPL, it was missing all the device drivers and hardware integration, and didn't even have stereo 3D rendering and head tracking. It was WI (Woefully Inadequate) as well FU. I brought this up with Mark Pesche and Owen Riley, and they were much more interesting in creating a standard that was built into...
Update: The "VPL Research" Wikipedia page gets the acronym expansion and citation wrong (the transcripts of the videos it cites don't actually say what VPL stands for).
However Jaron himself settled this in "Virtually There" (Scientific American, April 2001). The September 1984 Scientific American cover used one of his visual programming experiments as its illustration, an editor called in a panic because contributors must list an affiliation, and: "I blurted out 'VPL Research' (for Visual Programming Language, or Virtual Programming Language), and thus was born VPL.
Jaron's September 1984 Scientific American Cover (and check out the authors of the featured articles!):
After the issue's publication, investors came calling, and a company came to exist in reality." Both expansions, singular "Language", Visual first — the company was named after the visual programming language, not the other way around. Wikipedia's "Virtual Programming Languages" citation is a C-SPAN page that's just a video listing and never expands the acronym at all.
Seriously had no idea it went like this. I was a kid playing with VHSB so for me, it was a hoot - I could make my house, my school, the local swimming pool. I see parallels in my own son as he works on various 3D projects - it's just so fun, so different from reality, you're so powerful.
Clearly programmability is why the web became so successful; if it hadn't evolved into the messy apps platform it has become, it would have been surpassed by something else that had those features.
That said, doesn't it stand to reason that if VRML had been more usable with its basic feature set, it might have seen enough adoption to give birth to Dynamic VRML? After all, we apply an unhealthy dosage of Javascript to everything else in the world, why wouldn't it have soaked in there too? CSS for 3D objects, why not?
I gave up on more "advanced" features of AIML thinking "I will never have enough RAM memory for this on my ThinkPad" and yet here we are again. What a cycle.
I worked for an instant messaging company circa 99-05 and one of our skunkworks was a help bot using AIML and Alice. It was mostly good for laughs but this makes me want to go refresh my memory on how it worked.
>Tik-Tok finds that he can do as he pleases, and he secretly commits various hideous crimes for his amusement. After manipulating both robots and humans to cause chaos and bloodshed, Tik-Tok becomes wealthy (partly through health care privatization) and is finally elected Vice President of the United States.
Pretty good. Regarding Eliza, a version comes with Emacs, too. It's the "doctor" application. (The original Eliza was written as a generic engine running a pattern-matching "script").
A quibble: the creator of Eliza was Joseph Weizenbaum, not "Wizenbaum".
(set (make-local-variable 'doctor--stallmanlst)
'(((doc$ doctor--describe) your (doc$ doctor--feelings-about) him \.)
((doc$ doctor--areyou) a friend of Stallman \?)
((doc$ doctor--bother) Stallman is (doc$ doctor--random-adjective) \?)
> KMP provided some additional thoughts on his ELIZA:
> "The thing about the DOCTOR program, while it was quite elaborate, was that it sort of missed the point of what Weizenbaum was trying to say. It's not that I didn't understand him, but it took me a long time to really come to viscerally understand the seriousness of his point. He said that (paraphrasing, maybe badly, from memory) long before we got to real AI, we would get to things that seemed smart but weren't, and we would overly trust systems that were not really smart. He was, for example, very worried about computerized launch of nuclear missiles using what was then thought to be AI. But the point of the original ELIZA was not to BE elaborate, it was to be simple, and to show how even a tiny, tiny program could seem smart. It was hard not to think it could be smarter if it were not so tiny, but that misses the point, which was about how easily we were confused, not about how much more effectively we could be confused if more energy were put in. So whether my DOCTOR or any other was making the right point was hard to say. My guess is not."
Here's the source for DOCTOR from ITS -- it has a much less extensive set of foul words and doesn't get insulted like his ANIMAL game does:
I accidently offended it by earnestly trying to teach it to distinguish a bear from a pope, and it got mad at me and locked me out -- more details and pointers to Lisp code here:
Thanks for the collection. Racter was especially fun — his texts sound like they were written by a bot that had too much to drink and is trying to sound deep. But back in 1985, that was probably pretty impressive
The book is a better measure of what allegedly existed, though it was heavily curated, and the editorial choices in it say more about William Chamberlain than they do about the putative Racter itself.
I did try feeding the book into a small LLM to generate more text but it either doesn't follow its unusually stilted prose, which is clearly templated, or it just vomits back the entire corpus which is not very large.
In 1970, Nixon called Timothy Leary "the most dangerous man in America." Leary was sentenced to 10-20 years for marijuana possession — a draconian sentence designed to silence him.
Upon arrival at the California Men's Colony, Leary was administered psychological tests to determine his security classification. Among them: the Leary Interpersonal Behavior Inventory — a test he himself had designed twenty years earlier.
This was the ultimate field test of his own theory. If personality assessment truly revealed the inner self, Leary would be classified as what he was: a charismatic rebel, a leader, a flight risk.
But Leary understood the instrument. He knew which quadrant of the Interpersonal Circumplex mapped to "cooperative worker" vs. "dangerous agitator."
Result: Minimum security. Outdoor work detail. Access to the grounds with minimal supervision.
On September 12, 1970, Leary climbed a tree, shimmied across a telephone cable over the razor wire, and dropped to freedom. The Weather Underground had a car waiting. By October, he was in Algeria as a guest of the Black Panther Party's international chapter. By 1971, Switzerland.
The classic Bone Tumor Differential Diagnosis program that Apple published for the Apple ][ in the original Apple Software Bank Volume 1 from 1978 sounds legit. But beware: it requires 32K of RAM, bigger than any other program in that volume! (Available on both cassette and 5-1/4" floppy disk for free from your local Apple dealer.)
More broadly, the field has been focused on inductive logic programming since the 90s. More or less, one can think of it like synthesizing a logic program to describe and generalize background knowledge and examples. An automated approach to building the knowledgebase of an expert system. It works disturbingly well, especially with modern approaches.
A tremendous amount of work. The annual AGI Conference[1][2] was just a couple of weeks ago, and was the most heavily attended one in the event's history (which goes back about 19 years, IIRC). It was four days of presentations / papers / talks on various aspects of AGI, mostly not focused on LLM's. Not to mention dozens of posters which didn't merit a dedicated talk during the main conference.
And then you have the Neuro-symbolic Summer School[3] that's been running for the past few years, and also just recently took place for 2026.
To anybody not looking carefully, it might be easy to assume that LLM's have "taken all of the oxygen out of the room" with regards to AI. But dig a little deeper and one will find that there's still a lot of work going on pursuing other paths.
The problem is that we have fuzzy data and concrete semantics, when what we really want is concrete data and fuzzy semantics.
What we really need is a new system of reason, where each expression implies meaning, and each meaning has implications; but where those implications don't have to be reducible to computable binary logic.
The problem with LLMs is that they are ignorant to their own implications. What follows from a prompt is not any system of reason: it's just a vague sense of familiarity. Even when we have them stumble through the topology of a logical deduction, they can only act out that logic as a vain performance.
Natural language is a system of relative subjects, each with subjective implications relative to the story. So far in computing, we have nothing remotely like that.
An interesting fact I learned while reading The Dream Machine[1], is that Newell, Simon and Shaw’s Logic Theorist (1956)
was considered to be the “first AI program”.
“Logic Theorist is a computer program completed in 1956 by Allen Newell, Herbert A. Simon, and Cliff Shaw.[1] It was the first program deliberately engineered to perform automated reasoning, and has been described as "the first artificial intelligence program".[1][a] Logic Theorist proved 38 of the first 52 theorems in chapter two of Whitehead and Bertrand Russell's Principia Mathematica, and found a new and shorter proof for Theorem 2.85.[3]
[1] https://press.stripe.com/the-dream-machine”
Amusing and amazing to see this in the context of today’s Erdos-slaying LLMs.
33 comments
[ 2.4 ms ] story [ 8.4 ms ] threadRACTOR wrote a book, "The policeman's beard is half constructed."
https://archive.org/details/policemansbeardi0000unse/page/n2...
https://www.ubu.com/media/text/racter/racter_policemansbeard...
AIML: Artificial Intelligence Marketing Language
https://donhopkins.com/home/catalog/text/aiml.html
Support for AIML
https://donhopkins.com/home/catalog/text/SupportForAIML.html
To go along with the AIML parody, I also formatted as html and published an actual AI research paper from a company called Roanoke AI Laboratories, Inc., published in The Journal of Theriogenology. (NSFW, unless you work in the Artificial Insemenation industry!)
https://donhopkins.com/home/catalog/text/AILabratories.html
The author Jim Crump, Jr. actually found it and sent me email thanking me for publishing his AI paper on my web site, because his own publisher would not allow it due to copyright restrictions.
I had this ancient program called Virtual Homespace Builder https://archive.org/details/hmebuilder that was like an early Sketchup sorta thing and it would render VRML output, but I can't say as I ever managed to publish anything with it, being a young whelp at the time.
If VRML had been fast and efficient, something like Second Life / Metaverse might have emerged there...
At the time, I had been working extensively with Bounce / Body Electric, which was developed by Chuck Blanchard at VPL Research, originally used for VR simulations with the DataGlove and Eyephone and a pair of SGI workstations to render each eye.
"Visual Programming Language" was not an afterthought nailed on the side, it was what the name of the company VPL stood for.
https://en.wikipedia.org/wiki/VPL_Research
>VPL Research was one of the first companies that developed and sold virtual reality products. It was founded by computer scientist Jaron Lanier in 1984. "VPL" stood for "Virtual Programming Languages". In 1990, VPL Research filed for bankruptcy and in 1998 all of its patents were bought by Sun Microsystems.
>VPL's funding came in part from Marvin Minsky.
Funny that you would mention it: Virtual Homespace Builder was precisely what I was making fun of, in fact that was what the article I was parodying was about, almost word for word - I just went M-X replace-string Home => Head and then went from there:
>AI Labratories has entered into alliances with the likes of Microsoft, Apple, Intel, Creative Labs, NewTek, and Caligari to foster acceptance of the format. One of their partners, SexPression, offers the first AIML authoring program, Virtual Head Space Builder (VHSB) 1.0 Academic Edition ($495) for Windows. SexPression offers the final beta version for $49 and a paradigm-limited final beta via their Web site free of charge.
>Designed to accommodate nonprogrammers, VHSB creates AI Web sites, or "head spaces", that can be saved as AIML files or in SexPression's proprietary FASL file format. In addition to realtime interactive Artificial Intelligence, sites can include text, audio (WAV and RealAudio), MIDI, video, animation (AVI and FLI), cubes, spheres, pyramids, conic sections, and checkerboards -- in fact, SexPression suggests that VHSB makes a good expert system authoring tool as well. Any part of a knowledge base environment can act as a URL link, and evaluatable function modules can be associated. Head spaces can be opinionated with moods, including recurring and unpredictable swings, and with transparent prejudices. The software comes bundled with hundreds of clip-opinions, goals, interests, paridigms, taxonomies, ontologies, excuse templates, animal guessing trees, todo lists, top ten lists, hot lists, and shit lists; and custom content can be imported easily. A Virtual Head Space Evaluator is included.
Just being fast and efficient wasn't enough. It had to be TC (Turing Complete). No matter how fast and efficient it was, it was still FU: Fundamentally Useless.
Body Electric / Bounce STARTED with a fully Turing Complete, extremely powerful and fun to use Visual Programming Language, and then added Swivel 3D skeletons, networing with UDP packet packers/unpackers and transmitter/receiver modules, Swivel 3D tree syncing over Ethernet two one SGI workstation per eye for rendering to the Eyephones, many different kinds of input and output device drivers for the Eyephones (like a module to configure how far apart your eyes are to adjust the two 3D renderers), Dataglove, Convolvotron 3D sound rendering device, MIDI, Flock of Birds 6 degree of freedom motion tracker, etc.
While VMRL was not only missing the VPL, it was missing all the device drivers and hardware integration, and didn't even have stereo 3D rendering and head tracking. It was WI (Woefully Inadequate) as well FU. I brought this up with Mark Pesche and Owen Riley, and they were much more interesting in creating a standard that was built into...
https://en.wikipedia.org/wiki/VPL_Research
However Jaron himself settled this in "Virtually There" (Scientific American, April 2001). The September 1984 Scientific American cover used one of his visual programming experiments as its illustration, an editor called in a panic because contributors must list an affiliation, and: "I blurted out 'VPL Research' (for Visual Programming Language, or Virtual Programming Language), and thus was born VPL.
Jaron's September 1984 Scientific American Cover (and check out the authors of the featured articles!):
https://www.scientificamerican.com/issue/sa/1984/09-01/
After the issue's publication, investors came calling, and a company came to exist in reality." Both expansions, singular "Language", Visual first — the company was named after the visual programming language, not the other way around. Wikipedia's "Virtual Programming Languages" citation is a C-SPAN page that's just a video listing and never expands the acronym at all.
https://www.jaronlanier.com/cocodexintro/lanier01sciam.pdf
I suggested an edit, if anyone wants to apply it:
https://en.wikipedia.org/wiki/Talk:VPL_Research#What_VPL_Sta...
Clearly programmability is why the web became so successful; if it hadn't evolved into the messy apps platform it has become, it would have been surpassed by something else that had those features.
That said, doesn't it stand to reason that if VRML had been more usable with its basic feature set, it might have seen enough adoption to give birth to Dynamic VRML? After all, we apply an unhealthy dosage of Javascript to everything else in the world, why wouldn't it have soaked in there too? CSS for 3D objects, why not?
>Tik-Tok finds that he can do as he pleases, and he secretly commits various hideous crimes for his amusement. After manipulating both robots and humans to cause chaos and bloodshed, Tik-Tok becomes wealthy (partly through health care privatization) and is finally elected Vice President of the United States.
https://en.wikipedia.org/wiki/Tik-Tok_%28Oz%29
A quibble: the creator of Eliza was Joseph Weizenbaum, not "Wizenbaum".
So I fed his Natalism Flame to it:
http://www.art.net/studios/hackers/hopkins/Don/text/rms-vs-d...
>Are you annoyed that Stallman is emotionally stimulating?
https://github.com/jwiegley/emacs-release/blob/master/lisp/p...
https://sites.google.com/view/elizagen-org/#h.rle1hgawpigd
Kent Pitman's Lisp Eliza from MIT-AI's ITS History Project (sites.google.com)
https://news.ycombinator.com/item?id=39373567
xyzzy_plugh on Feb 18, 2024 | next [–]
Evergreen quote from TFA:
> KMP provided some additional thoughts on his ELIZA:
> "The thing about the DOCTOR program, while it was quite elaborate, was that it sort of missed the point of what Weizenbaum was trying to say. It's not that I didn't understand him, but it took me a long time to really come to viscerally understand the seriousness of his point. He said that (paraphrasing, maybe badly, from memory) long before we got to real AI, we would get to things that seemed smart but weren't, and we would overly trust systems that were not really smart. He was, for example, very worried about computerized launch of nuclear missiles using what was then thought to be AI. But the point of the original ELIZA was not to BE elaborate, it was to be simple, and to show how even a tiny, tiny program could seem smart. It was hard not to think it could be smarter if it were not so tiny, but that misses the point, which was about how easily we were confused, not about how much more effectively we could be confused if more energy were put in. So whether my DOCTOR or any other was making the right point was hard to say. My guess is not."
Here's the source for DOCTOR from ITS -- it has a much less extensive set of foul words and doesn't get insulted like his ANIMAL game does:
Doctor:
https://github.com/PDP-10/its/blob/master/src/games/doc.102
Guess the Animal:
https://github.com/PDP-10/its/blob/master/src/games/animal.1...
Animal has an extensive list of foul language words:
https://github.com/PDP-10/its/blob/master/src/games/animal.1...
I accidently offended it by earnestly trying to teach it to distinguish a bear from a pope, and it got mad at me and locked me out -- more details and pointers to Lisp code here:
https://news.ycombinator.com/item?id=39414536
I finally sent KMP and ANIMAL an email sincerely apologizing, and they forgave me, so I could play again:
https://github.com/PDP-10/its/blob/0f7d67997f9f5d30208e117e7...
I did try feeding the book into a small LLM to generate more text but it either doesn't follow its unusually stilted prose, which is clearly templated, or it just vomits back the entire corpus which is not very large.
https://en.wikipedia.org/wiki/Timothy_Leary%27s_Mind_Mirror
https://store.steampowered.com/app/1603300/Timothy_Learys_Mi...
https://news.ycombinator.com/item?id=32578683
https://scalar.usc.edu/works/timothy-leary-software/index
I extracted all the text from the Apple ][ Mind Mirror floppy disk image:
https://donhopkins.com/home/mind-mirror.txt
Mind Mirror Anthropic Skill and Documentation for MOOLLM:
https://github.com/SimHacker/moollm/tree/main/skills/mind-mi...
The Prison Escape: Ultimate Validation
In 1970, Nixon called Timothy Leary "the most dangerous man in America." Leary was sentenced to 10-20 years for marijuana possession — a draconian sentence designed to silence him.
Upon arrival at the California Men's Colony, Leary was administered psychological tests to determine his security classification. Among them: the Leary Interpersonal Behavior Inventory — a test he himself had designed twenty years earlier.
This was the ultimate field test of his own theory. If personality assessment truly revealed the inner self, Leary would be classified as what he was: a charismatic rebel, a leader, a flight risk.
But Leary understood the instrument. He knew which quadrant of the Interpersonal Circumplex mapped to "cooperative worker" vs. "dangerous agitator."
He answered strategically:
Dominance axis: Low scores ("I prefer to follow")
Hostility axis: Low scores ("I respect authority")
Interests: Forestry, gardening, nature work
Affect: Calm, conventional, agreeable
The prison psychologists, applying the Leary Interpersonal Behavior Inventory to Leary himself, concluded:
"Normal, well-adjusted, non-aggressive, conforming individual."
Result: Minimum security. Outdoor work detail. Access to the grounds with minimal supervision.
On September 12, 1970, Leary climbed a tree, shimmied across a telephone cable over the razor wire, and dropped to freedom. The Weather Underground had a car waiting. By October, he was in Algeria as a guest of the Black Panther Party's international chapter. By 1971, Switzerland.
He evaded capture for over two years.
----
Apple Software Bank Vol. 1: Bone Tumor Differential Diagnosis [Model 00114] © 1978 Apple Computer, Inc.
An expert system written in APPLESOFT BASIC by Jeffrey Dach, M.D.
https://news.ycombinator.com/item?id=21582027
DonHopkins on Nov 20, 2019 | prev | next [–]
The classic Bone Tumor Differential Diagnosis program that Apple published for the Apple ][ in the original Apple Software Bank Volume 1 from 1978 sounds legit. But beware: it requires 32K of RAM, bigger than any other program in that volume! (Available on both cassette and 5-1/4" floppy disk for free from your local Apple dealer.)
https://archive.org/stream/Apple_Software_Bank_Vol_1-2/Apple...
https://archive.or...
[1] https://dtai.cs.kuleuven.be/problog/
[2] https://github.com/ML-KULeuven/problog
[3] https://github.com/ML-KULeuven/deepproblog
[4] https://github.com/ML-KULeuven/deeplog
https://arxiv.org/pdf/2008.07912
https://en.wikipedia.org/wiki/Inductive_logic_programming
A tremendous amount of work. The annual AGI Conference[1][2] was just a couple of weeks ago, and was the most heavily attended one in the event's history (which goes back about 19 years, IIRC). It was four days of presentations / papers / talks on various aspects of AGI, mostly not focused on LLM's. Not to mention dozens of posters which didn't merit a dedicated talk during the main conference.
And then you have the Neuro-symbolic Summer School[3] that's been running for the past few years, and also just recently took place for 2026.
To anybody not looking carefully, it might be easy to assume that LLM's have "taken all of the oxygen out of the room" with regards to AI. But dig a little deeper and one will find that there's still a lot of work going on pursuing other paths.
[1]: https://agi-conference.org/
[2]: https://www.youtube.com/watch?v=qRA1DoMCCSc
[3]: https://luma.com/NSSS26
What we really need is a new system of reason, where each expression implies meaning, and each meaning has implications; but where those implications don't have to be reducible to computable binary logic.
The problem with LLMs is that they are ignorant to their own implications. What follows from a prompt is not any system of reason: it's just a vague sense of familiarity. Even when we have them stumble through the topology of a logical deduction, they can only act out that logic as a vain performance.
Natural language is a system of relative subjects, each with subjective implications relative to the story. So far in computing, we have nothing remotely like that.
Quoting from Wikipedia: https://en.wikipedia.org/wiki/Logic_Theorist
“Logic Theorist is a computer program completed in 1956 by Allen Newell, Herbert A. Simon, and Cliff Shaw.[1] It was the first program deliberately engineered to perform automated reasoning, and has been described as "the first artificial intelligence program".[1][a] Logic Theorist proved 38 of the first 52 theorems in chapter two of Whitehead and Bertrand Russell's Principia Mathematica, and found a new and shorter proof for Theorem 2.85.[3] [1] https://press.stripe.com/the-dream-machine”
Amusing and amazing to see this in the context of today’s Erdos-slaying LLMs.