James,
Thanks for that link. It provides important insights into the kinds of jobs that can benefit from current versions of AI. People classified as "developer" use AI tools far more than people classified as "programmer" or "data scientist".
The percentages for the last two are 7% and 8%. They need absolute precision, and an occasional bug takes far more time to correct than the time they gain by using AI tools,
As I keep repeating, precise evaluation is absolutely essential for a wide range of critical applications. For such tasks, any gain in speed is lost in the delays and even disasters caused by AI systems that have a "high percentage" of being correct.
As one programmer said, AI systems let me write a program in 5 minutes that would usually take an hour. But then I have to spend a week in debugging it.
People called developers are probably looking for suggestions that they can pass along to a programmer. And those suggestions may be good. But the programmer has the responsibility for mapping a broad suggestion to a precise solution that is designed for the exact situation.
John
----------------------------------------
From: "James Davenport' via ontolog-forum" <ontolog-forum(a)googlegroups.com>
https://www.infoworld.com/article/3489925/github-survey-finds-nearly-all-de… which is based on https://github.blog/news-insights/research/survey-ai-wave-grows/
The following article shows an important use of GPT-style tools. But note that they are not doing any reasoning or evaluation. The most valuable parts are error correction (simple mistakes) and finding code snippets from a vast range of software that can be found on the WWW.
I recommend clicking on the link below for more information.
But caveat emptor: it can also insert bugs, especially for larger snippets that do almost what you want, but for a slightly different representation than what you're using.
These are the same issues as writing aids in English and other languages. They are good for correcting spelling, inserting useful phrases, reducing the amount of typing, etc. But when they insert larger amounts of text, you have to check the details -- because they can insert stuff that is not what you meant. The worst cases are sections that look good at first glance, but you (or somebody else) discovers misleading or just false passages.
Treat every passage they insert as if written by a clever assistant who does a lot of good stuff, but sometimes makes disastrous errors.
John
--------------------------------
From ACM technews. Information for computer professionals three times a week.
GitHub Survey Finds Nearly All Developers Use AI Coding Tools
Nearly all (97%) of the 2,000 developers, engineers, and programmers polled by GitHub across the U.S., Brazil, Germany, and India said they have used AI coding tools at work. Most respondents said they perceived a boost in code quality when using AI tools, and 60% to 71% of those polled said adopting a new programming language or understanding an existing codebase was "easy" with AI coding tools.
[https://www.infoworld.com/article/3489925/github-survey-finds-nearly-all-de… ]InfoWorld (August 21, 2024)
First-order logic is a universal format that has been used to define every digital device of any kind and anything that can be implemented on any digital computer. Furthermore, every logician in the world and everybody who has studied logic knows FOL.
Alex: Let me propose to keep it in a knowledge hub form of framework. Like for ugraph theory here. Today when we have GenAI as an alternative knowledge concentrator we at least know that it is possible to put ALL our theoretical knowledge in one "computer".
No, no, no absolutely NOT! Only a tiny fraction of logicians have ever seen ugraph.
Only until there is a formal translation to and from FOL, we have ZERO evidence that ugraph is precise and reliable. And if anybody can demonstrate that such a translation is possible, then that is a proof that we don't need ugraph. We can continue to use FOL.
But there is also a standard for the superset of FOL called DOL, which is an official standard of the Object Management
Group. For a summary of the DOL standard, see slides 8 to 12 of https://jfsowa.com/talks/eswc.pdf . That talk (with a subset of the slides) won the best presentation award at the 2020 European Semantic Web Conference.
For the complete specification of the Distributed Ontology, Modeling, and Specification
Language, see https://www.omg.org/spec/DOL/1.0
And GenAI is notorious for its errors and hallucinations. It is not a good recommendation for anything that requires absolute precision and reliability.
John
----------------------------------------
From: "Nadin, Mihai" <nadin(a)utdallas.edu>
Sent: 8/24/24 4:32 PM
Why not?
Just wondering.
Mihai Nadin
Get Outlook for iOSFrom: ontolog-forum(a)googlegroups.com <ontolog-forum(a)googlegroups.com> on behalf of Alex Shkotin <alex.shkotin(a)gmail.com>
Sent: Saturday, August 24, 2024 3:22:23 AM
Heinrich,
Following "We developed a new theory of time and space" is it possible to read it?
What is the form your theory
exists in?
Let me propose to keep it in a knowledge hub form of framework. Like for ugraph theory here.
Today when we have GenAI as an alternative knowledge concentrator we at least know that it is possible to put ALL our theoretical knowledge in one "computer". But this must be done through the formalization of accumulated verified theoretical and technological knowledge, and not through training✈️
Alex
пт, 23 авг. 2024 г. в 16:58, Heinrich Herre <heinrich.herre(a)uni-leipzig.de>:
Hallo together,
I know, of course, that my proposal is a bit provocative. I wanted to hint that the onto-axiomatic method allows an access to the understanding of this problem. This problem can be reduced to another problem: Quantum mechanics needs time quants (and space quants), on the other hand, general relativity assumes the continuity of time and space. How these completely different structures can be unified? We developed a new theory of time and space, based the ideas of Franz
Brentano, and hopefully these new models allows constructions of new type. I had some years ago contacts to top physicists, they never heard something about Brentano's approach. Sometimes scientists are living an a bubble of conventional standard notions. The onto-axiomatic method provides methods to break such bubbles. We are working on this problem, let us see what we get.
With best wishes
Heirich
Am 22.08.2024 um 22:14 schrieb John F Sowa:
Heinrich and Alex,
The goal Heinrich summarized is the ultimate goal of mathematical physics. But I realize that it is a goal that some of the greatest scientists in the world have been working on for the past century. And I would also like to quote two scientists from the past century:
Einstein: God does not play dice with the universe.
Niels Bohr: Stop telling God what to do.
If Einstein, Bohr, and their colleagues and students couldn't solve the problems. I doubt that Ontolog Forum and the other sites listed on this note
will do so
I agree with Alex that a smaller, but still vast project would be somewhat more manageable.
Alex: So let's formalize one or another existing theory. We have this effort in Math like Isabelle, Coq etc. libraries, but not so much in other sciences. Even Mechanics, just Statics, is not fully formalized...
However, Ontolog forum has over a thousand subscribers, and most of them are not mathematical physicists. Even those who do have a good background in math & physics have other work to do, and they won't be able to contribute on a daily basis to any such project.
Therefore, I recommend that somebody who does have the time, funding, and interest in this project should set up a mailing list dedicated to it. Then send a monthly summary of the status to Ontolog Forum and/or other email lists.
John
Phil,
That policy you suggest below would be hard to implement, and it would require a huge amount of government legislation by every country in the world.
But the social media giants could implement a solution if and when some company developed the HYBRID technology to detect bad stuff (or any kind of stuff that anybody might be searching for). As an example of the kind of technology required, see the VifoMind examples from 2010 and earlier: https://jfsowa.com/talks/cogmem.pdf .
Skip to slide 44, which describes three projects that scan large volumes of text to find patterns stated as English questions. Those questions could be English descriptions of patterns to be detected. In those days, the technology was implemented (by us) on a small server. Today, a cell phone has more power. Our customers ran the VivioMind software on large systems that were vastly more powerful than our server.
Those systems did not use LLMs. But our new Permion.ai company has developed a major upgrade to the VivoMind system. In effect, it supports a HYBRID that uses LLMs to support a natural language translator that maps English (or other NLs) to conceptual graphs, which support the analysis and reasoning for finding, analyzing, and evaluating data of any kind,
Unfortunately, we can't get it ready to search the social media for the 2024 election, but we could do that for the 2026 election and scale it up for the 2028 election.
John
----------------------------------------
From: "Philip Jackson" <philipcjacksonjr(a)hotmail.com>
Sent: 8/17/24 7:53 AM
Here is one simple way to greatly reduce the spam, scam, erroneous and evil (e.g. virus-containing) emails that are sent to and received each day by hundreds of millions of people: Make it so that sending a single email would have a nonzero cost, e.g. a nickel for each destination email address, which would need to be paid by the sender to the national post office. Without the sender paying such a cost, the email would go into the bit bucket, and not be delivered.
To be clear, although this is simple to describe and easy to understand, it would not be easy to implement: There would be a variety of technical, business and government challenges, and probably also new laws to create and implement. Yet we appear to have reached a point where something like this may be needed.
Phil
Just after I wrote that I was going to limit the number of notes that I would forward, I was gifted with the following note.
I believe that it is an important innovation that can make LLMs significantly better. But I still believe that some hybrid method for evaluating results is stll a requirement.
For example, I don't believe that it can detect and eliminate all the bad stuff on social media.
John
--------------------------------------------------------------------------------
Meta’s Multi-token Model, A New Beginning for AI?
A New Type of Faster… and Smarter LLMs?
https://medium.com/@ignacio.de.gregorio.noblejas/metas-multi-token-model-a-…
Meta offers this with new research: a model that predicts multiple tokens at once in every prediction, not just one, and unlike previous proposals, with no training overhead.
This not only speeds up the model's text generation but could also make it smarter, meaning we might be about to enter a new training paradigm for frontier AI.
Chris and Simon,
Your comments are consistent with widely used development tools that have been implemented.
I agree that frame notation can be quite user-friendly. But it's important to have a formal specification of exactly how it is mapped to and from FOL or some subset or superset. Many people throw up their hands when they see predicate calculus notation, but it can be translated to English-like notations with just nine words: AND, OR, NOT, IF, THEN, EITHER, OR, SOME, EVERY. (Actually, you only need three words -- AND, NOT, SOME -- but the other words make the sentences shorter and easier to read.)
However, there are many factors that can make implemented systems more user-friendly. For an overview of the technical issues and examples of various implementations, see https://jfsowa.com/talks/cnl4ss.pdf
Interesting point: TQA was a very usable English Query language developed by IBM research. They found that it was much easier to translate English to predicate calculus notation and then to SQL than to translate directly to SQL. Formal notations that are good for computers can be good intermediate notations to and from natural languages
In fact, that is an excellent application for LLMs (or more liikely SLMs -- Small Language Models): translate notations and diagrams with good human factors to and from computer systems..
As the cnl4ss slides show, the users loved TQA, but IBM canceled the project because the task of customizing TQA for each application was too difficult for most users and too expensive for IBM. But the current LLM or SLM technology could learn to do the translations very quickly and inexpensively.
I wrote cnl4ss long before LLM/SLM were available. But today that technology would be excellent for tailoring any of the systems discussed in that pdf.
John
----------------------------------------
From: "Chris Mungall" <cjmungall(a)lbl.gov>
Frames actually turn out to be quite useful for meta-modeling of OWL ontologies. I think we took the wrong fork in the path in 2006, as you suggest.
A lot of large OWL ontologies turn out to be quite unwieldy and difficult to maintain, leading to a lot of different approaches like templated ontology generation from spreadsheets or design pattern systems. See my keynote from the Ontology Pattern Workshop from 2020: doi.org/10.5281/zenodo.7655184
LinkML provides a language with frame-like semantics for a modern user base (YAML rather than S-expressions; compiling to Pydantic and JSON-Schema), and a more user-friendly way to incorporate IRIs for all elements.
We have a framework linkml-owl (https://linkml.io/linkml-owl/) that allows the (historically implicit) metaclasses in an OWL ontology to be modeled in LinkML/frames, with the OWL TBox being "compiled" from LinkML/Frame "instances". This kind of metamodeling is at best very awkward in OWL itself. See the tutorial.
On Thu, Aug 15, 2024 at 3:23 AM 'Polovina, Simon (BTE)' via ontolog-forum <ontolog-forum(a)googlegroups.com> wrote:
Hi John and all.
Protégé should have maintained its Frames version. At https://protege.stanford.edu/conference/2006/submissions/slides/7.2wang_pro…, there is an insightful presentation that compares Frames and OWL side by side. Notably, the leading industry-strength Enterprise Architecture (EA) tool The Essential Project | Enterprise Architecture Tool (enterprise-architecture.org) uses Protégé Frames under the hood, evidenced by its open-source version. OWL did not fit the bill, as Meta-modelling is important (highlighted in the above presentation link). John, you identified these benefits in your sowazach.pdf (jfsowa.com) 1982 paper with John Zachman, the ‘father’ of EA.
Hence, your remark about OWL’s limitations in commercial products is well-taken.
Simon
Protege is limited to OWL, which is more complex and more limited than first-order logic.
But I realize that many uses of a type hierarchy do not require the full power of FOL. My recommendation would be Aristotle's syllogisms for a type hierarchy, supplemented with FOL for a constraint language. This was the original intention for description logic before the decidability gang restricted its expressive power.
Unfortunately, the constraint of decidability had three results: (1) it made the language more complex; (2) it seriously limited its expressive power; (3) it made it unusable for a wide range of tools in AI, computer science, and commercial products. There are many reasoning tools that are more expressive, more powerful, and easier to use than Protege.
For a brief overview of Aristotle's syllogisms, see slides 25 to 30 of https://jfsowa.com/talks/patolog1.pdf
For more detail about Aristotle and modern logics, see all slides of patolog1.pdf and any references cited on any of those slides.
John
---------------------------------------------------------
From: "Mara Abel" <marabel(a)inf.ufrgs.br>
Colleagues
We are wondering here if we can use the reasoning of Protege to
automatically produce labels for the entities and instances of a domain
ontology.
Any idea about it?
Thank folks!
I received an offline note that said I was being too negative about the power of the LLM-based technology.
I never wanted to give that impression. I am enthusiastic about that technology and its potential.
But I am also enthusiastic about the achievements of the 70 years of AI and computer science. The most important new developments take advantage of both. I often criticize attempts to use LLMs for applications that they cannot support -- deductive reasoning is a prime example.
I am most enthusiastic about the hybrids, but I am critical of attempts to make LLMs do things that they cannot achieve by themselves. Therefore, I frequently comment on the failures of LLM applications that do not make a proper balance. There are many, many examples. Many of the ones I discuss were sent to me by other people -- often in offline notes by people who would prefer not to be mentioned.
If anybody thinks that I have not achieved a proper balance in one or more notes, please cite the examples. Some people do so on the lists and others send me offline notes.
John
