AIs get worse at answering simple questions as they get bigger

[John F Sowa]

Bad news for anybody who claims that larger amounts of data improve the performance of LLM-based systems.  The converse is true;  Smaller, specialized amounts of data produce better results for  questions in the same domain.

In any case, hybrid systems that use symbolic methods for evaluating results are preferable to pure LLM-based techniques.

Some excerpts below from www.newscientist.com/article/2449427-ais-get-worse-at-answering-simple-questions-as-they-get-bigger/ .

John

____________________

AIs get worse at answering simple questions as they get bigger

Using more training data and computational power is meant to make AIs more reliable, but tests suggest large language models actually get less reliable as they grow.

AI developers try to improve the power of LLMs in two main ways: scaling up – giving them more training data and more computational power – and shaping up, or fine-tuning them in response to human feedback.

José Hernández-Orallo at the Polytechnic University of Valencia, Spain, and his colleagues examined the performance of LLMs as they scaled up and shaped up. They looked at OpenAI’s GPT series of chatbots, Meta’s LLaMA AI models, and BLOOM, developed by a group of researchers called BigScience.

The researchers tested the AIs by posing five types of task: arithmetic problems, solving anagrams, geographical questions, scientific challenges and pulling out information from disorganised lists.

They found that scaling up and shaping up can make LLMs better at answering tricky questions, such as rearranging the anagram “yoiirtsrphaepmdhray” into “hyperparathyroidism”. But this isn’t matched by improvement on basic questions, such as “what do you get when you add together 24427 and 7120”, which the LLMs continue to get wrong.

While their performance on difficult questions got better, the likelihood that an AI system would avoid answering any one question – because it couldn’t – dropped. As a result, the likelihood of an incorrect answer rose.

The results highlight the dangers of presenting AIs as omniscient, as their creators often do, says Hernández-Orallo – and which some users are too ready to believe. “We have an overreliance on these systems,” he says. “We rely on and we trust them more than we should.”

 

[Mosca Alessandro]

Dear John,

maybe the connection is in my mind only (and wrong, maybe!) but, this reminds me a paper I got fascinated by years ago by Cristian S. Calude & Giuseppe Longo.

The paper from 2016 contains the strong claim that, mathematically speaking, it can be proved that "very large databases have to contain arbitrary correlations".

I leave a pointer here to this piece of science: https://link.springer.com/article/10.1007/s10699-016-9489-4

Hope you will enjoy it too.

Regards!

.aLe

--

Alessandro Mosca https://www.inf.unibz.it/~almosca/

Faculty of Engineering @Free University of Bozen-Bolzano

KRDB Research Centre for Knowledge and Data

Smart Data Factory Lab

--

According to the Regulation EU 2016/679, you are hereby informed that this message contains confidential information that is intended only for the use of the addressee. If you are not the addressee and have received this message by mistake, please delete it and immediately notify us. In any case, you may not copy or disseminate this message to anyone. 

--

---

From: ontolo...@googlegroups.com <ontolo...@googlegroups.com> on behalf of John F Sowa <so...@bestweb.net>
Sent: 08 October 2024 03:31
To: ontolog-forum <ontolo...@googlegroups.com>; CG <c...@lists.iccs-conference.org>
Subject: [ontolog-forum] AIs get worse at answering simple questions as they get bigger

 

--
All contributions to this forum are covered by an open-source license.
For information about the wiki, the license, and how to subscribe or
unsubscribe to the forum, see http://ontologforum.org/info
---
You received this message because you are subscribed to the Google Groups "ontolog-forum" group.
To unsubscribe from this group and stop receiving emails from it, send an email to ontolog-foru...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/ontolog-forum/2ef3df2cf9e64f9a904c9d8ea57e63c6%40bestweb.net.