Following is the latest announcement of Permion products. This is a major step upward from our old VivoMind Company, which terminated all activity in 2010. Arun Majumdar is the CEO of Permion, and I am a Permion Fellow. That means I am in an advisory position, but not actively working on day-to-day developments.
Note the two terms in the phrase "neural pattern recognition with symbolic reasoning." Tthe first uses LLMs for Generative AI and the second is an extended upgrade of the symbolic reasoning in our VivoMind system of 2010.
The purely symbolic applications of VivoMind in 2010 were very impressive, even without any Generative AI components. For examples of some applications (all of which were developed for paying customers) see slides 44 to 64 of https://jfsowa.com/talks/cogmem.pdf .
Note that those applications, which are more than 15 years old, are still very impressive today. The latest extensions by Permion can perform all of them in addition to the LLM-style of processing. They can use the symbolic methods to detect and eliminate LLM errors and hallucinations.
John
___________________
Permion Listed in AWS "ICMP" for the US Federal Government
Permion's XVMTM and DSMETM AI Product Suite Now Available to US Government Customers in ICMP.
February 19, 2025 08:55 AM Eastern Standard Time
WASHINGTON--(BUSINESS WIRE)--Permion Inc., a software company focused on delivering trusted AI for mission critical applications, today announced that it listed XVMTM and Digital Subject Mater Expert (DSMETM) in the AWS Marketplace for the U.S. Intelligence Community (ICMP). ICMP is a curated digital catalog from Amazon Web Services (AWS) that makes it easy to discover, purchase and deploy software packages and applications from vendors that specialize in supporting government customers.
"As the need for trustworthy and scalable AI rapidly grows, we are excited to leverage the AWS Marketplace to deliver solutions to the government," said Permion COO John Sankovich. "Customers require innovative AI capabilities that can adapt to evolving challenges and needs while maintaining the highest standards of security and reliability. Permion’s solutions deliver exactly that – powerful, flexible AI tools that transform how organizations process information and make critical decisions."
Permion's XVM™ is an artificial intelligence (AI) virtual processor that combines neural pattern recognition with symbolic reasoning, empowering developers to create more intelligent and trustworthy AI systems. Through its open API and SDK, the XVM™ toolkit harnesses the power of large graph and language models, delivering flexibility and performance while maintaining robust security standards. XVM™ integrates with modern development environments to significantly reduce development time and operational costs while helping ensure AI systems operate at peak efficiency.
Permion's DSME™ is an application built on the robust foundation of XVM™ and optimizes how operators, analysts and businesses process and understand large sets of complex information. By analyzing vast amounts of data across multiple formats and languages, DSME™ allows users to make faster, more informed decisions while maintaining accuracy and reliability.
About Permion
Permion is a pioneer in developing advanced AI and data solutions that transform industries and unlock value for customers. Our Permion XVM™ platform offers advanced AI and machine learning capabilities within our AI virtual processor for complex, real-world problem-solving and business outcomes. Based in Washington D.C., Permion is committed to delivering trusted, secure, and effective AI solutions for mission-critical applications. Learn more at www.permion.ai and info(a)permion.ai.
Greg and Alex,
I agree that graph structures support algorithms that are more flexible than the typical operations with predicate calculus. I also agree that it's important to support methods that deal with approximations or "fuzzy" kinds of truth values. There has been a huge amount of theoretical and practical R & D 0n these issues in the past 50 years.
GS: The subject of pattern logic is developed over several illustrated webpages and culminates in the topic of "decisions" which represents a novel theory of propositional truth grounded in the structure of interpreted patterns. This expanded theory of truth encompasses both the typical truth values, as well as intermediate degrees of certainty and contradiction.
I don't know the details of your system, but from your notes, I believe that you have a more advanced system than the so-called ":semantic web stack" that is based on the 2005 "layer cake". But as I have said many times, the "decidability gang" destroyed the vision and specifications in the winning proposal by Tim Berners-Lee in 2000. The so-called "Semantic Web Stack" of 2005 was a pale shadow of what Tim B-L had proposed.
In 2003, another branch of the Federal Gov't saw that the Semantic Web was headed in the wrong direction, and they funded a much more advanced and more ambitious project called IKRIS. See https://jfsowa.com/ikl/ . It was funded for two years (2004 to 2006) and it included some of the most advanced AI projects and researchers from industry and Academia. Arun Majumdar and I were just two of the many researchers involved.
Unfortunately, there were some cutbacks in gov't funding in 2005, and neither the IKRIS project nor the SW project were continued. For a survey of developments from the late 1970s to 2011, see Semantics for Interoperable Systems, documents collected and related by John Sowa, https://jfsowa.com/ikl/ .
You don't have to believe me. I wrote the overview, but I include links to the original R & D articles by everybody I mention in the reviews. I also have more reviews and publications, but this is enough for now.
John
----------------------------------------
From: "gregsharp73" <gregsharp73(a)gmail.com>
The requested pattern logic primer can be found here: https://patternslanguage.com/pattern-logic
The subject of pattern logic is developed over several illustrated webpages and culminates in the topic of "decisions" which represents a novel theory of propositional truth grounded in the structure of interpreted patterns. This expanded theory of truth encompasses both the typical truth values, as well as intermediate degrees of certainty and contradiction.
For a discussion of why a cyclic pattern language is an improvement over linear formal languages and offers an approach to making the meaning in natural language accessible to computation see the following article: https://patternslanguage.com/articles/f/a-cyclic-pattern-language
Thanks,
Greg
On Sunday, February 16, 2025 at 9:39:21 AM UTC-7 Alex Shkotin wrote:
Just small Sunday's evening addition: If we put brackets we get this
(∄(Philosopher) ⊇ ∂(Person))
i.e. we have two unary modifiers (∄ ∂) of unary predicates (concepts) into something that can be modified by the binary operator ⊇ to create a proposition.
Very interesting.
вс, 16 февр. 2025 г. в 14:42, alex.shkotin <alex.s...(a)gmail.com>:
by the way " ∄Philosopher ⊇ ∂Person" [1] is not a FOL, but HOL like this
∃p:unary_predicate, ∃x:Person p≠Philosopher ∧ p(x)
[1] https://patternslanguage.com/articles/f/unifying-logic-traditional-premises…
пятница, 14 февраля 2025 г. в 17:29:39 UTC+3, Gregory Sharp:
ꓱPhilosopher ⊇ ∂Person is a label given to the central occasion of a 9 occasion pattern that follows the general form of a logical statement. This particular statement is analogous to an Aristotelian I-premise which can be rendered in English as "some person is a philosopher". The general form of a logical statement requires quantification of both of its terms and a copula. The copula here is called "predication". There are three other copulas. The concept philosopher is existentially quantified. The concept person is partially quantified. There are two other quantifiers used in the analogous Aristotelian system. They are universal and non-existential quantification. There are six additional quantifiers in pattern logic. The four copulas and ten quantifiers set the boundaries for pattern logic "proper" in the ADEPT LION "first consideration" which encapsulates the "grammar" of the broader pattern language. The second consideration is the vocabulary and the third consideration is the syntax.
From the standpoint of pattern logic, "wife" is a concept, or a non-logical term.
Greg
On Fri, Feb 14, 2025, 4:28 AM Alex Shkotin <alex.s...(a)gmail.com> wrote:
Just FYI:
"17.02.2025, jointly with S.I. Adian seminar Alexei Miasnikov (Stevens Institute of Technology): First-order classification, non-standard models, and interpretations
In this talk I will focus on three things:
1. First-order classification: in particular, how one can describe ALL groups which are first-order equivalent to a given one.
2. Non-standard models of groups: in particular, I will describe non-standard models of the finitely generated groups with decidable or recursively enumerable (or arithmetic) word problems and explain how they naturally appear as non-standard Z-points of the general algebraic schemes.
3. Theory of interpretations: it seems a new rich theory is emerging right now. I will show several interesting results based on interpretations."
https://www.mathnet.ru/eng/conf876
Enjoy,
Alex
Gary,
The knowledge encoded in genes is not specified in language or in what we might call an ontology. But it's encoded in methods of interacting with the world and the kinds of things and events we encounter. The knowledge we encode in language (verbal or gestural) is invented by us to communicate, and that involves many level of encoding in genes for language, for communication, and for interacting with other humans and beasts. An infant is born with genes that can enable such learning, but there are many, many kinds of things that must be learned befor that ability can be activated.
GBC: One might ask if there is human a priori knowledge via say evolution for ideas such as space and time if Kant would argue for any immediately relevant a priori knowledge of what we now talk about via models in quantum physics or the notion of space-time in relativity."
Ideas are not encoded in genes. A new-born infant has the ability to register sensations and the ability to move around. By trial and error, it discovers which actions have favorable results, which ones are painful, and which ones get milk, warmth, and good feelings.
There is a huge amount of pre-verbal learning in the first few months. The first interactions are with a warm soft milk source. After a while, baby learns that the milk source comes and goes, and other similar things don't provide milk. That is the first step in distinguishing mama from daddy.
GBC: But I would hazard the argument that we can distinguish QM probabilistic aspects, say quantum uncertainty, as more recently developed cognitive tools developed as part of collective, social and scientific experiences building models.
Baby has no idea about QM vs Newton. By age 3, the child is a very long way from being a baby, but he or she is nowhere near the level that could distinguish a Newtonian world from a QM world.
I would guess that a child that could appreciate the story of Alice in Wonderland would be able to distinguish QM effects from Newtonian effects. But that ability would have to be learned from some kind of unusual experience, not intuited from everyday experience.
And the words for describing it would be "book learning" or something taught by a special kind of teacher.
Basic point: There is a huge jump from learning a language to learning the use the technical terms for grammar and semantics. Another huge step: learning to deal with things that behave differently from one time to another and learning that some behaviors involve an abstract notion of probability.
GBC: After more than a century from its birth, Quantum Mechanics (QM) remains mysterious.
Yes, but people have learned how to use it for better AND worse.
John
----------------------------------------
From: "Gary Berg-Cross" <gbergcross(a)gmail.com>
John,
A few quick thoughts in response to your " Every species from bacteria on up inherits a huge amount of knowledge about space, time, edible items, dangerous items, and species-specific methods for dealing with them. "
Which was in response to my :
"Every species from bacteria on up inherits a huge amount of knowledge about space, time, edible items, dangerous items, and species-specific methods for dealing with them.
One might ask if there is human a priori knowledge via say evolution for ideas such as space and time if Kant would argue for any immediately relevant a priori knowledge of what we now talk about via models in quantum physics or the notion of space-time in relativity."
Certainly evolution prepares organisms to survive in the world which includes navigating space and having some sense of time. Systems biology among other things gives us a bit more of a dynamic idea of some of things going on here. Even simple autonomous agents are able to distinguish and select external entities by virtue of a simple chemistry that hosts and affords cognitions like symbols and signs. (but also hypotheses and models not inherited) One conceptualization uses the idea of a processing system. Evolution has developed what some call innate “programmed in” software operating on inherited hardware as the result of natural selection. But natural selection also allows learning via systems such as self-organizing maps that optimize the results of try and error behaviors (such as your bee example).
This system's view is one that combines an innate view and an empirical/experiential one where “becoming” is as much or more meaningful than just “innate/natural being”. But innate infrastructure provides affordances. Self-organizing maps are perhaps just one simple example of reflective loops that provide/afford higher sign-creating activity and eventually symbolic thinking that allows building useful models of perceived reality.
And certainly we have inherited reasoning methods that can deal with probabilities. Probability seems to provide a rich framework for vision and motor control, as well as higher order functions for learning, language processing, and reasoning (Tenenbaum, Joshua B., et al. "How to grow a mind: Statistics, structure, and abstraction." science 331.6022 (2011): 1279-1285.) So we can say that an agent learns what is likely to happen if they do X. But I would hazard the argument that we can distinguish QM probabilistic aspects, say quantum uncertainty as more recently developed cognitive tools developed as part of collective, social and scientific experiences building models. As D’Ariano says “Quantum Mechanics (QM) is a very special probabilistic theory...After more than a century from its birth, Quantum Mechanics (QM) remains mysterious. (D’Ariano, Giacomo Mauro. "Probabilistic theories: what is special about quantum mechanics." Philosophy of quantum information and entanglement 85 (2010): 11.)
Gary Berg-Cross
Potomac, MD
240-426-0770
On Thu, Feb 13, 2025 at 5:43 PM John F Sowa <sowa(a)bestweb.net> wrote:
Gary,
There is evidence of a priori knowledge in the genome. It's knowledge that is learned by the species and encoded in genes rather than neurons: The total genome of any species has an enormous amount of information. And very little of that information has been decoded.
Gary B-C: One might ask if there is human a priori knowledge via say evolution for ideas such as space and time if Kant would argue for any immediately relevant a priori knowledge of what we now talk about via models in quantum physics or the notion of space-time in relativity.
Every species from bacteria on up inherits a huge amount of knowledge about space, time, edible items, dangerous items, and species-specific methods for dealing with them.
For example, just consider honey bees. The methods for finding honey-bearing flowers and bringing back both honey and pollen are extremely complex, and there is no way that a newly born bee could have learned that method by experience or by teaching from other bees. There are also complex methods by which a returning bee communicates the distance, direction, and amount of honey by a dance back in the hive.
There is also evidence of learning how to communicate better. Bees that forage outside the hive acquire larger amounts of brain tissue that is devoted to memory of spatial distance and direction than their fellow bees that work inside the hive.
Furthermore, the regions of a honeybee's brain for spatial info are analogous to regions of the hippocampus in mammals. Guess what? Humans and squirrels who need to remember large amounts of spatial info also acquire enlarged regions in the hippocampus.
For humans, this fact was discovered by observing London taxi drivers who had to memorize a huge amount of information about streets and locations. They had significantly larger regions in their hippocampus. But now that taxi drivers use computer displays for that information, there is no enlargement of the hippocampus.
For squirrels that bury nuts in autumn and must remember the locations for several months, their hippocampus grows larger in autumn and decreases in size in the spring.
Re quantum mechanics and relativity: A critical aspect of QM computation is the need for reasoning about probabilities. That would imply that humans would need to have inherited reasoning methods that can deal with probabilities. But such reasoning methods would be useful for many non-QM reasoning as well. Reasoning about space time is also essential, and it's hard to distinguish QM aspects from the many other factors involved.
John
Gary,
There is evidence of a priori knowledge in the genome. It's knowledge that is learned by the species and encoded in genes rather than neurons: The total genome of any species has an enormous amount of information. And very little of that information has been decoded.
Gary B-C: One might ask if there is human a priori knowledge via say evolution for ideas such as space and time if Kant would argue for any immediately relevant a priori knowledge of what we now talk about via models in quantum physics or the notion of space-time in relativity.
Every species from bacteria on up inherits a huge amount of knowledge about space, time, edible items, dangerous items, and species-specific methods for dealing with them.
For example, just consider honey bees. The methods for finding honey-bearing flowers and bringing back both honey and pollen are extremely complex, and there is no way that a newly born bee could have learned that method by experience or by teaching from other bees. There are also complex methods by which a returning bee communicates the distance, direction, and amount of honey by a dance back in the hive.
There is also evidence of learning how to communicate better. Bees that forage outside the hive acquire larger amounts of brain tissue that is devoted to memory of spatial distance and direction than their fellow bees that work inside the hive.
Furthermore, the regions of a honeybee's brain for spatial info are analogous to regions of the hippocampus in mammals. Guess what? Humans and squirrels who need to remember large amounts of spatial info also acquire enlarged regions in the hippocampus.
For humans, this fact was discovered by observing London taxi drivers who had to memorize a huge amount of information about streets and locations. They had significantly larger regions in their hippocampus. But now that taxi drivers use computer displays for that information, there is no enlargement of the hippocampus.
For squirrels that bury nuts in autumn and must remember the locations for several months, their hippocampus grows larger in autumn and decreases in size in the spring.
Re quantum mechanics and relativity: A critical aspect of QM computation is the need for reasoning about probabilities. That would imply that humans would need to have inherited reasoning methods that can deal with probabilities. But such reasoning methods would be useful for many non-QM reasoning as well. Reasoning about space time is also essential, and it's hard to distinguish QM aspects from the many other factors involved.
John
Greg,
The term "modern logic" is hopelessly confusing. Every decade there is a new modern logic. Anybody and her brother can concoct a new notation and call it "modern logic". Unfortunately, there are huge numbers of such notations and no systematic method for relating them.
For the ISO Common Logic standard, we started with an ABSTRACT specification that has NO readable or writable notation of any kind. The standard itself specified three different readable and writable DIALECTS. But anybody who specifies any notation that can be mapped to and from the abstract specification can call it a dialect of Common Logic.
In fact, it's very easy to specify an English-like dialect of first-order predicate calculus. All you need are 7 reserved words: and, or, not, if, then, some, every. This version is immediately readable by anybody who can read English. With LLM technology, It can also be translated to and from anybody's favorite notation, whatever they may call it.
I agree with the importance of the six theories you listed below. Each of them can be specified with a few more key words added to the basic seven.
Unfortunately, we made a mistake in specifying the ISO Common Logic standard: We did not include an English-like dialect in the official document. I strongly recommend that anybody who proposes any kind of logic include a translation to and from an English version that uses only those 7 key words for the base logic plus whatever key words are needed for the additions described below (and/or any additions for any branch of science and engineering).
And by the way, there have been many versions of logic that do enable users to type whatever English assertions or questions that they prefer. Then the system can translate them to their preferred English-like version and ask the question: "Is this what you mean?"
Those systems have been very user friendly. Unfortunately, they required a considerable amount of work to design them and to keep then user friendly. With the current LLM technology, it is much easier to implement the translation from arbitrary English to whatever official version of controlled English may be specified.
And by the way, LLM technology can also map any natural language to and from a controlled dialect. There can be official dialects of French, German, Russian, Chinese, Arabic, Malaysian, Vietnamese, etc. And they would all be exactly compatible with their translations to whatever formal notation is the foundation.
John
________________________________________
From: "gregsharp73" <gregsharp73(a)gmail.com>
To further unpack my cryptic response to Paul’s understandable incredulity regarding the integration of Traditional and Modern Logics and to follow up on Alex’s request for an example of the formal notation used in Pattern Logic, I put together an article: https://patternslanguage.com/articles/f/unifying-logic-traditional-premises…
I believe this article nicely closes the loop on where my interest lay in raising this question about Nicola’s last slide. I do see a means of unifying these various theories using the same toolbox of fixed patterns that I demonstrate in unifying Traditional and Modern Logic in the article. To sketch out an approach:
1. Theory of Parts (Mereology) à universal and partial quantifiers of pattern logic with mediation copulas
2. Theory of Unity and Plurality à generic quantifiers (individual and total) and copulas of pattern logic
3. Theory of Essence and Identity à general quantifiers and mediation copulas as explained in the article
4. Theory of Dependence à quantified monadic predicates of Modern Logic as explained in the article
5. Theory of Composition and Constitution à universal and partial quantifiers of pattern logic with union and overlap copulas
6. Theory of Properties and Qualities à quantified monadic predicates of Modern Logic as explained in the article
These discoveries have been made in obscure and isolated labor over many years now and I would welcome collaboration as it has arrived at this point of so many different branching paths. Please contact me if you have an interest in walking any of them with me.
Thanks,
Greg
(Apologies if you receive multiple copies of this call)
--------------------------------------------------
CALL FOR ABSTRACTS
Workshop on Human-Centred Machine Learning: Bridging Design, Development, and Social Impact
March 6-7, 2025, Lugano, Switzerland (in-person)
--------------------------------------------------
The B4EAI team (Best 4 Ethical AI) invites abstracts for the Workshop on Human-Centred Machine Learning: Bridging Design, Development, and Social Impact (March 6-7, 2025, Lugano – Switzerland), a two-day event focused on addressing critical questions about the ethically and socially responsible design and deployment of AI systems.
This workshop seeks to advance human-centred design principles and participatory approaches in AI development and contribute to the responsible adoption of these technologies.
Our workshop features a distinguished lineup of keynote speakers, including (in alphabetical order):
- Federico Cabitza (University of Milano-Bicocca, IT): "Measuring Social Impact in AI Systems"
- Cristina Conati (University of British Columbia, CAN): "Community Perspectives in AI Research"
- Marcello Ienca (Technical University of Munich, DE): "Policy Perspectives on AI Governance"
- Caterina Moruzzi (Edinburgh University, UK): "Future Directions in Human-Centered AI"
- Elisa Rubegni (Lancaster University, UK): "Human-Centred Design Principles in AI Development"
- Niels van Berkel (Aalborg University, DK): "Successful Case Studies in Human-Centred AI"
The programme includes dedicated panel sessions on the topics of "Bridging Technical Innovation and Social Impact", "Ethical Considerations in AI Development", "Implementation Challenges and Solutions" and "Shaping the Future of Human-Centred AI".
Finally, a session will be dedicated to spot presentations (10 minutes) by participants that would like to present their own research.
**Submission Guidelines**
For the spot presentations session, we welcome short abstracts addressing, but not limited to, the following topics:
- How can we design AI systems that meaningfully incorporate diverse human perspectives and needs throughout their development cycle?
- What methodological frameworks can bridge the gap between technical AI capabilities and real-world social challenges?
- How do we measure and evaluate the social impact of AI systems beyond traditional technical metrics?
- What are effective approaches for participatory design in AI development, particularly when working with marginalised or vulnerable communities?
- How can we ensure AI systems remain accountable to their intended social benefits throughout their lifecycle?
We invite the submission of abstracts for the meeting from early career scholars (students, postdoctoral researchers, and junior faculty).
Abstracts should be up to 250 words long (excluded references) and they should include name, affiliation, and title in one .pdf file.
Abstracts should be submitted by February 21 to: b4eai.info(a)gmail.com
**Important Dates**
Submission Deadline: February 21, 2025 AoE
Notification of Acceptance: February 24, 2025
Workshop Dates: March 6-7, 2025, Lugano – Switzerland
Location: Dalle Molle Institute for Artificial Intelligence, Campus Universitario USI/SUPSI (Via la Santa 1, 6962 Viganello)
For further details, please contact alessandro.facchini(a)supsi.ch
** We are committed to fostering diversity and equality. Submissions from underrepresented groups are particularly welcome. **
Michael D and Michael DB,
The answer to that question is very clear and simple:
MDB: Where is there a purely symbolic AI system that can support NLP anywhere in the same ballpark as LLMs? If such a thing exists would love to know about it. Keep in mind the system has to do a lot more than turn natural language into SPARQL. It needs to keep the context of a discussion thread, it needs to be able to respond appropriately when you give it feedback like: "give me more/less detail" "change the tone to [more playful] [more professional] [more creative]" etc. It needs to be able to do common sense reasoning and handle very hard NLP problems like anaphora.
Answer: We did that with the purely symbolic VivoMind technology (from 2000 to 2010). The language understanding part was far better than the language generating part. In accuracy and precision, the 2010 version far exceeds any and every current LLM system that does not use a symbolic component to detect and evaluate the inevitable errors and hallucinations. Our largest clients (some US gov't agencies) attest to that fact.
For an overview and documentation of the old VivoMind system, see https://jfsowa.com/talks/cogmem.pdf . The date of this presentation is 2016, but it is an update of slides that were originally presented in 2011. See the applications that begin on slide 45. Every one of them was implemented in 2010 or earlier. And in precision and accuracy, they are far ahead of anything that can be implemented by LLMs today without using symbolic technology for reasoning and evaluation.
Our new Permion Inc. system combines an upgraded version of the 2010 symbolic technology with LLMs for an enhanced natural language OUTPUT that has a more readable and more elegant syntax. Permion also uses LLMs for some of the searching, but it only requires a modest amount of storage. For customers with deep pockets, however, the Permion technology can scale up to use as much storage as the customer provides.
For a talk on YouTube, see Evaluating and Reasoning with and about GPT by John Sowa and Arun Majumdar, https://www.youtube.com/watch?v=6K6F_zsQ264 .
That talk only had 544 views. But the following YouTube, from the 2024 Ontolog Summit, had over 10 thousand views: Without Ontology LLMs are Clueless by John Sowa, https://www.youtube.com/watch?v=t7wZbbISdyA .
We strongly suspect that a large percentage of the viewers are in China. One reason is that I have been invited to give talks (expenses paid) in places like Hong Kong, Macau, Singapore...
Another reason: Arun has also received an offer from China to buy our Permion company. However, some of our key customers would not be happy with that idea.
John
----------------------------------------
From: "Michael Denny" <regattago(a)gmail.com>
Michael,
I was a little surprised to see you say
“What LLMs do is to solve the problem of common sense reasoning better than any other approach to date. LLMs know that things fall down, that water is wet, etc. “
When I asked Claude 3.5 the following:
“Is it accurate to say that LLMs solve the problem of common sense reasoning and that LLMs know that things fall down, that water is wet, etc.?”
The response was:
“No, that's not accurate. The relationship between LLMs and common sense reasoning is much more complex and limited.
LLMs can often produce text that appears to reflect common sense understanding, but this is primarily because they've learned statistical patterns in human-written text about how objects and phenomena are typically described. They don't have genuine understanding or reliable common sense reasoning capabilities.
For example, an LLM might correctly state that objects fall downward or that water makes things wet, but it doesn't truly "understand" gravity or wetness the way humans do. Some key limitations:
- Inconsistency: LLMs can make basic physics or common sense errors when questions are asked in novel ways or contexts
- No real-world grounding: Their "knowledge" comes from text patterns, not from physical experience with the world
- No causal understanding: They don't truly grasp cause and effect relationships
- Pattern matching vs. reasoning: They often give sensible-looking answers by matching patterns rather than through actual reasoning
This is why LLMs can sometimes make surprisingly basic errors about physical reality or common sense relationships that would be obvious to humans. They might correctly state that "water flows downhill" in one context but then make an elementary mistake about fluid dynamics in another context.
What are your thoughts on the nature of common sense reasoning in AI systems? I'm curious about your perspective on this.”
I think the key point is that LLMs do not understand naïve physics or any deterministic relationships as humans do, nor have the inferential capabilities symbolic logics do. But as you point out, they do make a great interlocutor.
Mike
From: ontolog-forum(a)googlegroups.com [mailto:ontolog-forum@googlegroups.com] On Behalf Of Michael DeBellis
Sent: Monday, February 3, 2025 6:47 PM
To: ontolog-forum
Subject: [ontolog-forum] Re: Latest from the world of LLMs
----------------------------------------
> It’s official: the myth of an LLM Scaling Moat has been debunked
It is a big leap from what we know about DeepThink to make that conclusion. For one thing, I don't trust the mainstream press (CNBC) for any info about a complex scientific or technology question. For another, there is so much we don't know with confidence about DeepThink. How well does it really do on benchmarks? How much money or computing resources did the Chinese government or other groups associated with them pump into DeepThink? For another, there are things that Open AI can do that DeepThink can't do such as their recent Deep Research announcement: https://www.youtube.com/watch?v=YkCDVn3_wiw ; Also, see the Playlist "12 Days of Open AI" on Youtube: https://www.youtube.com/playlist?list=PLOXw6I10VTv9lin5AzsHAHCTrC7BdVdEM
But even supposing that every claim made about DeepThink is true, it doesn't mean "the myth of an LLM Scaling Moat has been debunked" It just means the moat isn't as large as we thought. And if that moat isn't as large as we thought, I don't see how that justifies the conclusion: " All roads lead back to good old symbolic reasoning and inference—now more accessible and understandable when combined with a multi-modal natural language interface that enhances interaction possibilities." If anything, it makes LLMs even MORE usable and accessible because it doesn't require the huge resources that we used to think to create one (from what I know, I'm deeply skeptical but I actually hope that the claims of DeepThink are true because I think it is a good thing that we don't restrict foundational LLMs to just a few companies)
Where is there a purely symbolic AI system that can support NLP anywhere in the same ballpark as LLMs? If such a thing exists would love to know about it. Keep in mind the system has to do a lot more than turn natural language into SPARQL. It needs to keep the context of a discussion thread, it needs to be able to respond appropriately when you give it feedback like: "give me more/less detail" "change the tone to [more playful] [more professional] [more creative]" etc. It needs to be able to do common sense reasoning and handle very hard NLP problems like anaphora.
What LLMs do is to solve the problem of common sense reasoning better than any other approach to date. LLMs know that things fall down, that water is wet, etc. It doesn't make sense to also expect them to be able to solve that problem AND have completely reliable domain knowledge. There are several approaches to using curated, authoritative knowledge sources with an LLM to reduce or eliminate hallucinations. One of them is Retrieval Augmented Generation (RAG). That's an example of how semantic tech and LLMs complement each other: https://journals.sagepub.com/eprint/RAQRDSYGHKBSF4MRHZET/full also see: https://www.michaeldebellis.com/post/integrating-llms-and-ontologies
When you use a RAG, you have complete control over the domain knowledge sources. Building a RAG is the best way I've ever seen to put an NLP front end to your knowledge graph.
Another way to get domain knowledge into an LLM are to train LLMs with special training sets for the domain. One excellent example of this is Med-BERT, a version of the BERT LLM specially trained with healthcare data. See: L. Rasmy, Y. Xiang, Z. Xie, C. Tao and D. Zhi, "Med-BERT: pretrained contextualized embeddings on large-scale structured electronic health records for disease prediction," Nature npj digital medicine, vol. 4, no. 86, 2021.
Still another (much newer) way is to use reinforcement fine-tuning to train an LLM with domain specific data. This doesn't take as much work as the domain specific LLM but is much more work than creating a RAG. See: M. Chen, J. Allard, J. Wang and J. Reese, "Reinforcement Fine-Tuning—12 Days of OpenAI: Day 2," 6 December 2024. https://www.youtube.com/watch?v=yCIYS9fx56U&t=29s
