#LLMs are not #ExpertSystems.
There, I said it.
#ExpertSystems
#Silicium logithèque 💾
Avant #ChatGPT, il y avait GURU (MDBS ~1986)
Système expert PC mêlant règles, base de faits et SGBD. « Introducing software with a mind of its own »
➡️Un jalon de l’IA symbolique
#IA #ExpertSystems #RetroComputing #HistoireTech
#heritage #geek #toulouse
Rule-Based Expert Systems: The Mycin Experiments (1984)
https://www.shortliffe.net/Buchanan-Shortliffe-1984/MYCIN%20Book.htm
#HackerNews #RuleBasedExpertSystems #Mycin #Experiments #AIHistory #ExpertSystems #1984Tech
Expert systems were the bleeding edge of AI in 1970s - 1980s.
Currently experts systems are being reincarnated as reasoning models. Experts transform their know-how to Chain-of-thought (CoT) models.
Copying is the fastest way to learn.
AI groups spend to replace low-cost ‘data labellers’ with high-paid experts
https://www.ft.com/content/e17647f0-4c3b-49b4-a031-b56158bbb3b8
❝Top artificial intelligence groups are replacing low-cost “data labellers” in Africa and Asia with highly paid industry specialists, in the latest push to build “smarter” and more powerful models.
❝Companies such as Scale AI, Turing and Toloka are hiring experts in fields such as biology and finance to help AI groups create more sophisticated training data that is crucial for developing the next generation of AI systems.
❝“What these models now need is data of a real human using the models to do knowledge work, and getting feedback on when the model is failing,”
In a lot of ways contemporary AI resembles expert systems like ELIZA. Mostly what you get out of it is as a conversation partner to bring out yr own good ideas. Maybe it'd be better to undertrain it like the original ELIZA?
We need to be thinking about zen systems rather than literal systems. Might also save a fuck of a lot of energy.
#AI #LLM #ExpertSystems #eliza
"The problem facing the development of expert systems, that is, systems that enable a computer to simulate expert performance ... is that an important part of the expert knowledge is tacit. If experts try to articulate the knowledge they apply in their performance, they normally regress to a lower level. Therefore, according to Hubert and Stuart Dreyfus, expert systems are not able to capture the skills of an expert performer."
#RagnarFjelland, 2020
MYCIN for identifying bacterial infections, and Dipmeter Advisor for analyzing oil exploration data, were expert systems developed with Interlisp as well as early AI success stories. Learn more:
MYCIN
https://en.wikipedia.org/wiki/Mycin
https://archive.org/details/rulebasedexperts00buch
Dipmeter Advisor
https://en.wikipedia.org/wiki/Dipmeter_Advisor
#ExpertSystems in general, and #Cyc in particular, were once the coolest things in #ArtificialIntelligence 😐:
“Cyc: History’s Forgotten AI Project”, I. A. Fisher (https://outsiderart.substack.com/p/cyc-historys-forgotten-ai-project).
» However while research funding dried up and the term #AI became less used, many green shoots where planted and continued more quietly under discipline specific names:
cognitive systems,
machine learning,
intelligent systems,
knowledge representation and reasoning.
Offshoots of these then made their way into commercial systems, such as #ExpertSystems in the #BusinessRules Management System #BRMS market. «
https://blog.kie.org/2012/05/drools-5-4-artificial-intelligence-a-little-history.html
Agora com a "moda" da Inteligencia Artificial muita gente me vem perguntar sobre o assunto.
Respondo apenas o basico que sei e evito botar faladura sobre o assunto.
Foi uma area que nao acompanhei/acompanho desde 1990, altura em que tive as ultimas cadeiras nessa area (4 se bem me recordo), pelo que sou ignorante no tema, em particular as evolucoes que teve nos ultimos 30 anos.
#lisp #scheme #prolog #expertsystems #neuralnetworks #turingtest #bayesianinference
Reviewing some of that noisy AI hype, I sometimes tend to think of all the oh so unfullfilled promises of "expert systems" in the late 1980s and the dear wasted money that went down the drain together with the promises.
#AI #hype #expertSystems #moreHype #evenMoreHype
An intro to Truth Maintenance Systems
https://ntrs.nasa.gov/citations/19930006101
#tms #expertsystems #FaultTolerance #gofai
This was an EXTREMELY interesting paper on what modern LLMs can learn from older older symbolic AI / expert system approaches to improve the validity of what those statistical models generate on their own, using the Cyc system as the contrast to the modern systems.
https://arxiv.org/abs/2308.04445
#Cyc #DougLenat #SymbolicAI #ExpertSystems #LLMs #ChatGPT #Bard #AI #KnowledgeRepresentation
I have not thought of Prolog much since college but it was the language we developed #ExpertSystems in our #AI computer science class. https://techhub.social/@pragprog/110950377314338218
I feel like this current push of #LargeLanguageModels and #MachineLearning to apply #AI to practical problems has caused people to ignore research into rule-based #ExpertSystems, which are the better fit for a lot of situations (e.g. recommender systems, inference engines, etc).
Take this with a pound of salt because I am not an AI practitioner, know relatively little about LLMs, and have written inference engines for money.
I asked Bing, "How can I paddle a canoe with chopsticks?" How do you think an expert canoeist would have responded? #LLMs #ai #artificialintelligence #expertsystems #wisdom
@asj had an ex boss who worked quite a bit on the Symbolics Lisp Machine. I also recently discovered a site by Alumni: https://smbx.org
Very good look at the variety of problems with automated decision making for social services, credit ratings, predicted recidivism, etc.. Products exist doing these things right now, and there are no regulations or checks.
The systems are flawed and dangerous; resulting in deaths, false accusations, imprisonment, children being taken away, and more.
#ML #ExpertSystems #SocialJustice
Against Predictive Optimization https://predictive-optimization.cs.princeton.edu/
@gothick @pandion This is probably far closer to the truth than you might think. Here is a true story:
This was back in the mid 1980’s, the early days of expert systems, before they had that name. At my institution (gov), if you wanted a computer you had to write, submit and get approval of an Automated Data Processing (ADP) plan before you could buy it. The plans were complicated with many rules, cases, etc., and consumed tremendous amount of time to prepare and review, often with many iterations.
Fresh out of grad school, I could not let this stand. A couple of us worked for about a month and created an expert system to write perfect ADP plans. The system was built in HyperCard! It asked users a series of necessary and sufficient questions and filled in the ADP Office’s template plan. It was hugely popular and people made plans and finally started ordering the computers they wanted. Success! But wait, it gets better.
About two months later, I got a call from the ADP Office asking for help. They were being inundated with ADP plans and couldn’t review them in a timely way (as if they ever did). Could I write them a program to help review the plans? I was astonished that they were clueless about my ADP-Writer program, and I did not enlighten them.
It was trivial (but I took plenty of time) to create ADP-Reader. So ADP-Reader was “reading” plans created by ADP-Writer! Users were happy, the ADP Office was happy, and I was a hero.
After six months or so of this, I couldn’t let the charade continue. I pulled back the curtain and … it wasn’t fast, but the ADP Office was defunded and ADP plan requirements were eliminated.
On #SymbolicAI vs #NeuralNetworks in #ArtificialIntelligence and #MachineLearning:
"Artificial Intelligence Is A House Divided", The Chronicle Of Higher Education (https://www.chronicle.com/article/artificial-intelligence-is-a-house-divided).
Client Info
Server: https://mastodon.social
Version: 2025.07
Repository: https://github.com/cyevgeniy/lmst