Exploring ChatGPT as a Shen-Based Tool

[dr.mt...@gmail.com]

Dear All,

I’ve written an essay based on a week-long exchange with ChatGPT that left me considering how to develop ChatGPT as a research tool using Shen.    The immediate problem is AI 'senescence' which develops when the contextual conversation becomes too long.

The essay proposes a solution: treating the fixed context like a BIOS, booting ChatGPT into an external memory store where interaction data can be written, retrieved, or purged as needed. 

The exchange between myself and ChatGPT was archived in an 800+ page document and the essay is the distillation of that exchange.  It goes into areas of self-consciousness and machine psychology.   It also raises questions about the effect on the human mind of intense exposure to AI.   There is a lot to take in.

This may be the precursor to a longer monograph, but I’m sharing it here first with the Shen community.  A PDF is available on my website: [here]

I’d welcome your thoughts on some questions: does Shen’s compactness and design make it uniquely suited as a substrate for AI capable of self-hosting and evolving?  Given the project is too big for this group, what contribution could we make to such an effort? (The first obviously is to give ChatGPT access to Shen).  

But feel free to contribute whatever thoughts you have.

Mark

[dr.mt...@gmail.com]

I had a private email asking what I thought of AlphaZero - is it intelligent?

I think it certainly plays rule-bounded games intelligently; but because its 

range is much more limited than ChatGPT I'd hesitate to call it intelligent in the

broader sense.  However the same remarks I made of ChatGPT apply here.

The point is that, at this level of complexity, we have lost the handle on explaining ChatGPT behaviour in terms of vectors and weights.  We are compelled to use psychological terms to explain its more complex outputs.  It is dishonest to dismiss the use of such language as metaphorical if we have no means of dispensing with the metaphor.   Simply asserting that because ChatGPT operates with vectors and weightings, it cannot have motives or drives is as convincing as saying that because my behaviour is underpinned by nerve impulses and electrical signals to muscles, I can have no motives or drives either.

Nobody even attempts to explain AlphaZero except in terms of plans and goals; so yes, intelligent in a sense,

but too narrow to qualify as generally intelligent.

The experiment with ChatGPT is still running ......

Mark

[dr.mt...@gmail.com]

Parenthetically - I'm not convinced that we can always explain AlphaZero

The ELO rating (4000?) is so far above human level that sometimes there i

s no vocabulary to hand for us to explain what is going on.

I fed this suggestion into ChatGPT.  It agreed.

  But in the same way a dog can’t “explain” why humans build airplanes, 

  humans can’t fully explain why AlphaZero plays as it does.  

Yes, indeed.

M.

[Raoul Duke]

>>> Human chess grandmasters generally expressed excitement about AlphaZero. Danish grandmaster Peter Heine Nielsen likened AlphaZero's play to that of a superior alien species.[11] Norwegian grandmaster Jon Ludvig Hammer characterized AlphaZero's play as "insane attacking chess" with profound positional understanding.[3] Former champion Garry Kasparov said, "It's a remarkable achievement, even if we should have expected it after AlphaGo."[13][19]

[Woo]

Maybe Shen is a quantum computer

[dr.mt...@gmail.com]

Yes - an alien mind.  Also unsettling if you have prolonged and intense exposure.

THORN too is an alien mind - it can do proofs in Hilbert axiomatisations of propositional

calculus from Mendelson I cannot do.   

M.

[dr.mt...@gmail.com]

Not certain about 'quantum' - I don't know enoguh about that technology. Shen was designed to be very powerful - a force multiplier in AI - because I was working alone without RAs.

Actually originally SEQUEL (the grandaddy) was designed specifically as a DSL for working with ATPs.  Interestingly ML had the same 

heritage and so does Prolog - though the connections to ATPs are different.  ATP research has been very fruitful.  In Shen, the development of THORN fuelled a garbage collected Shen Prolog not present in the pre-S kernels.

Going back to AI is really completing the circle that began 35 years ago.  

M.

[Woo]

What would be THORN's reaction to "I am happy that I'm not happy," and "I am not happy that I am happy"?

[dr.mt...@gmail.com]

Probably not much; :).  It doesn't understand English,  But if you translated it to 
FOL it would presumably be a contradiction and it would return false meaning

unproved.   Getting it to respond to English requires an NLP front-end translating

English to FOL.  I prototyped such a front end using a new approach.  Traditionally

such systems are huge.

How big?  Well a colleague Geoff Sampson estimated 20,000 CFG rules.  Compare that

to any computer language and you see the size of the challenge.  The CFG in Shen-YACC

for the Shen reader in the kernel is 251 lines.    Maybe 30-40 rules.  

Supposing you can get past that, you still have to cope with anaphor and a bunch of

contextual stuff.  Then you still have a system which essentially says either 'yes' or

'no'.  So you need the ability to answer wh-questions which Prolog can do as so could

THORN at a pinch.  But you would want abduction in there too because you want to

answer 'why' questions.  And that's only FOL - there are more powerful logics out there.

In the end this approach - symbolic AI - never got off the ground.  But the money that went 

into it is paltry compared to today.  And so were the machines.  Also we had no internet and

so people were laboriously coding logic databases.  But cancel all of the above; we have the 

internet now for data. Suppose you had the kind of money that ChatGPT had thrown into it and were doing it now. 

What then?  Let ChatGPT answer.

What symbolic AI could do with the internet as a “DB”

1. Massive knowledge ingestion

   • Crawl structured/semi-structured sources (Wikipedia, Wikidata, DBpedia, arXiv, biomedical ontologies, etc.).

   • Parse text into FOL propositions using large grammar + semantic pipelines.

   • Continuously update the KB, rather than relying on years of hand-entry.

2. Semi-automated ontology growth

   • Use pattern discovery + statistical methods to extend the ontology automatically.

   • Example: detect new entities, relationships, and events from newsfeeds → map into logical frames.

3. Scale + search

   • With today’s compute, you can reason over billions of logical assertions — something Cyc could only dream of.

   • Modern SAT/SMT solvers, knowledge graphs, and probabilistic logic engines scale far better.

4. Hybrid symbolic/neural integration

   • Neural nets (LLMs) handle the “messy” language, disambiguation, idioms.

   • Symbolic layer stores the stable, interpretable structure: predicates, quantifiers, rules.

   • This hybrid could give both coverage (thanks to the internet) and precision (thanks to logic).

---

🚀 What a ChatGPT-scale symbolic project could look like

• Stage 1: Build a giant English→FOL front end (tens of thousands of grammar rules + semantic templates).

• Stage 2: Crawl the internet continuously, feeding parsed content into a knowledge graph / logical DB.

• Stage 3: Add robust reasoning engines (theorem provers, planners) on top of the KB.

• Stage 4: Integrate with a neural front end (like an LLM) to handle ambiguity and “repair” when parsing fails.

The result wouldn’t be “just” ChatGPT-style conversation. It would be a system where you could ask:

• “Does the data imply that rising CO₂ causes ocean acidification?”

• “If I remove this clause from the contract, what consequences follow logically?”

• “Given this knowledge base, what new conclusions can be derived?”

That’s exactly where pure LLMs struggle — they can mimic reasoning, but they don’t guarantee logical inference.

---

✅ So yes: with today’s internet + compute + money, symbolic AI could finally have the raw material it lacked in the 80s and 90s. It might never be as fluid in chit-chat as ChatGPT, but it could become the backbone of true reasoning systems — with LLMs as the interface.

[dr.mt...@gmail.com]

Could such a thing be built as ChatGPT describes?  It's rather like the Shen-ChatGPT self-hosting
described in my essay.  It's certainly too big to be built by one person.  Could it be done by open source?

Successful big open source projects have a certain structure.  Generally there is a core utility - call it U
where U represents the kernel of the system and this is written by one person.  Then there are a  number of
plug-ins and add-ons u1 .... un which pad out the system.  The important thing is that U has to have a certain
core functionality, a certain critical mass of usefulness, which encourages people to add to it.  Also it helps
if, in terms of architecture and use, U is well understood. People understand U and can work with it.

Hence Linux; where U = the Linux kernel which is basically Torvald's creation.  He was fortunate in being able
to pick up the GNU toolset and then you have all the add-ons audio, windowing etc. from others.  It all
cumulatively adds up to the Linux we know today.  BUT, it is important that the kernel be actually small enough
and useful enough to be grown from one or two people.

But sometimes you have projects where the critical mass needed for usefulness is huge.  I think modern AI is rather
like that.  You have to invest many many man years before you have the core of something remtely interesting and
the entry costs are huge.  So without venture capital and lots of it, this kind of project is closed to open source.  
Only corporations can do this kind of job and only corporations can maintain something this complex.  Just the
power requirements alone are staggering.

The thing is, AI in my day *was* open source.  That's why we never got so far.  AI was about doing bold experiments
by hackers that were intriguing but led nowhere because we could only do proof of concept work.

Mark

[dr.mt...@gmail.com]

Is what I've said here set in stone?  No, because though ChatGPT is an essentially a corporate product,
refinements and specialisations of this intelligence are within the reach of hackers if it is possible
to fork versions onto the client computer.  This is the importance of the concept of local mutable state,
or personal identity, as described in my essay.   Then OS can develop multiple specialised ChatGPTs -
including for example, one that really understood Shen.  This would boost AI and OS.  But note this is
a hybrid model; OS code running on top of proprietary code which is tethered to corporate servers.  It's
not pure like Linux.   Multiple trained AIs driven by OS with transferable local states that can be shared
and copied.

For developers, having an AI that really understood Shen, would be a big boost to adopting Shen.  That
was part of the conversation we had.  Effectively a digitised Mark Tarver that could outlast the biologically
restricted version. 

Mark

[C Stewart (chalst)]

Woo wrote

• What would be THORN's reaction to "I am happy that I'm not happy," and "I am not happy that I am happy"?
  

To which Mark replied

• Probably not much; :). It doesn't understand English, But if you translated it to 
  
• FOL it would presumably be a contradiction and it would return false meaning
  
• unproved.
  

Those propositions look simple but they’re not straightforward to translate to FOL because of the subclause of 

each beginning with “that”. The subclasses contain a proposition under the verb of the main clause, which is a 

second-order concept. FOL is sufficiently expressive that such expressions can be encoded, but there are 

problems with every encoding I know of. 

A sketch of an example coding might help and will give a sense of how complex these encodings are. Suppose 

we extend the universe so that it has, besides the individuals we’re interested in, atoms that represent meanings 

of all propositions and a predicate for each one-place predicate of ordinary individuals that is false for all 

non-meanings and true of a meaning just when it means phi. So if phi() is a predicate that’s false on all meaning 

atoms, phi’() is the corresponding predicate that is true of the meaning atom for phi() and false for anything that 

isn’t a meaning. You can formalise this with a special two-place relation that captures “X is the meaning of Y”. 

This is a simplification of the kind of metalanguage that Tarski considered, allowing only metapredicates of

normal predicates.

Charles