3 ms·
At the end of the day, I guess I just don't really see what the issue is with AI hallucinations or how they make them less intelligent. My brother in law who i
by anon291 2y ago
At the end of the day, I guess I just don't really see what the issue is with AI hallucinations or how they make them less intelligent.
My brother in law who is extremely smart, and very well trained -- MIT PhD -- and exposed to orders of magnitude more data than ChatGPT, given that the human senses ingest gigabytes of data per day, has -- like everyone else -- persistent errors in his thinking. I single him out because of his credentials, but everyone has these.
As a particularly funny example, we were talking about steers and bulls and he was like "Oh, right, I remember, some male cattle are born bulls and some are born steers". My wife, who is similarly educated, believed for a very long time that watermelon's grew on trees. At my undergrad school, a woman who was getting her physics degree and then who eventually went on to get a PhD, fundamentally believed that yeast would not rise if you talked around it -- I guess her parents had told her that and she never questioned it.
That is to say, humans lack full knowledge. We're trained on significantly more data than ChatGPT et al. We still have persistent inaccuracies despite all of that. We're not perfect.
That doesn't make us less useful. Half the time we just rely on other humans trained on similar, but slightly different data, to iron those out, and indeed AI systems that employ those methods get better accuracy.
Now again, my brother in law has fewer errors than most people since he is very smart. However, the vast majority of people are really not that bright (rumor has it that fifty percent have below average intelligence), and yet, simply through verification with other humans, they are able to accomplish useful things.
Sometimes even very capable people have persistent thinking errors. This doesn't make them less smart. No one can claim perfect intelligence. That's the only thing that's bullshit.
- nikvaes 2y agoI just want to point out that I think ChatGPT has seen more data than a single human being. With 45 TB of text data [1], it would cost a human around 70k years of continuous reading (assuming 1 GB is 200M words, and humans read 225 words per minute) to see all the training data. [1] https://www.businessinsider.com/google-researchers-openai-chatgpt-to-reveal-its-training-data-study-2023-12?international=true&r=US&IR=T https://www.businessinsider.com/google-researchers-openai-ch...
- anon291 2y agoHumans have a much richer IO system than any GPT model. Estimates say humans ingest about 75GB of data per day. That means ChatGPT has seen about 600 days worth of human data. Or about two years. So if it's smarter than a two year old, it's safe to say that it surpasses human ability. The mistake is thinking that 'data' means only text.
- alephxyz 2y ago> We're trained on significantly more data than ChatGPT et al. I don't think that's true anymore. GPT-4 is rumored to have been trained on a dataset with >1 trillion tokens. An average human would take a few thousand years to read through that dataset once.
- anon291 2y ago1 trillion tokens is about 1024 GB (assuming 1 byte per token). Humans ingest about 74GB of data per day: https://kids.frontiersin.org/articles/10.3389/frym.2017.00023 https://kids.frontiersin.org/articles/10.3389/frym.2017.0002... That means for a human baby to be as smart as GPT they would have to go from zero to speaking English within two weeks of birth. Your error is mistaking training data for text only. Humans are particularly inefficient with text because it takes a lot of visual data to get to the thought of a particular word or character due to how human senses work (we have no 'token' sense).
- justsomehnguy 2y ago> At the end of the day, I guess I just don't really see what the issue is with AI hallucinations or how they make them less intelligent. At the end of the day, you wouldn't be bothered if Average Joe would add arsenic to your favourite pie, because AI told it would enhance it's taste, right? Thing is, Average Joe wouldn't be the one asking your brother about bulls and steers, or your wife about watermellons. They would ask ChatGPT. And 99% it would be all fine, until some orphanage cook would try to improve some dish.
- anon291 2y agoI mean sure... I think we should be careful to avoid thinking any system that is attempting human like intelligence to be perfect. Because human intelligence is not so any system that models it would also not be.