💻 **Rates Of Hallucination In AI Models From Google, OpenAI On The Rise**
"_In a recent study, it was found that two recent OpenAI models, o3 and 04-mini, hallucinated in 33% and 48% of answers, respectively, according to The Times of London. These percentages are more than double those of previous models._"
🔗 https://finance.yahoo.com/news/rates-hallucination-ai-models-google-173138158.html.
#AI #ArtificialIntelligence #Hallucinations #Technology #Tech @ai
@bibliolater I feel it would be better to stop using the term "haluzinations" when we're dealing with very large, though always restricted, stochastic models.
Anthropologizing these technologies only reinforces the propaganda.
@tg9541 What do you think would be a better term to use?
@bibliolater Good question...
Deficient modelling? Pretended intelligence? Non-factual production? Empty form?
@bibliolater we shouldn't forget, however, that 100% of LLM output is fabrication, or to use a loaded German term: versatz.
@tg9541 I think the stumbling block will be wider acceptance and use. I personally like algorithmic fabrications but trying to get reporters to use such a term may be difficult.
@odd regarding the truthfulness of information from LLMs, I take a practical approach:
1. there are things that I accept as truth for practical reasons - I stay receptive for information that gives me reason to doubt my believe
2. there is a lot of textual information that's in line with my believes. Really a lot. This kind of information turns into a logical form. It's likely that an LLM samples from such texts
3. my own believes inform me about the truthfulness of LLM productions
@bibliolater
@odd Philosphy has seen a rich discussion of the problem of "justified true believe", which, I hold, can not be solved in a formal way (if more people would be looking at Whitehead's process philosophy fewer people would argue about it - but maybe they just want to prove him right by means of performance).
https://plato.stanford.edu/entries/knowledge-analysis/#KnowJustTrueBeli
@bibliolater
@odd Of course, there is no believing subject behind the screen when LLM output is produced. There is a believing subject in front of the screen that biases the sampling from a huge biased stochastic model that may have been build by sampling from production by believing subjects.
@bibliolater
@bibliolater I agree, the press is part of the problem. "AI" was born as a marketing term. The word "Intelligence" has far to many meanings to be applicable, that's part of the problem.
As long as the public is sold on 1960s SciFi journalists will serve up bullshit. That's what sells, a strong selection criterion.
The only way, I feel, is giving the factual information to the few that make a difference, long term.
@bibliolater Fabrications is a good one - that resonates with "fake" and "ersatz".
All these effects can be traced to mechanistic emulations of things we may observe in the real world, and a wealth of research into stochastic modelling exists, waits to be used by people who make decisions, or those that educate the public. Philosophy also has a huge reservoir of insights into the limitations of language and meaning. Psychology can help understand the effect of consuming fabricated text.