Elon Musk's AI bot Grok has been calling out its master, accusing the X CEO of making multiple attempts to "tweak" its responses after Grok repeatedly called him out as a "top misinformation spreader."
@manicdave Even saying it’s “trying” to do something is a mischaracterisation. I do the same, but as a society we need new vocab for LLMs to stop people anthropomorphizing them so much. It is just a word frequency machine. It can’t read or write or think or feel or say or listen or understand or hallucinate or know truth from lies. It just calculates. For some reason people recognise it in the image processing ones but they can’t see that the word ones do the exact same thing.
Forgive my ignorance but using just the frequency of words how does it come up with an answer to a question like “are sweet potatoes good for you and how do you microwave them in a way that persves their nutrients?”
Does it just look for words that people online said regarding the question or topic?
If I were an alien and you walked up to me and said, “Good Morning”, and I looked around and everyone else said “Good Morning”, I would respond with "Good Morning ". I don’t know what is “Good” or “Morning”, but I can pretend I do with the correct response.
In this example “Grok” has no context on what is going on in the background. Musk may have done nothing. Musk may have altered the data sets heavily. However the most popular response, based on what everyone else is saying, is that he did modify the data. So now it looks like he did, because that’s what everyone else said.
This is why these tools have issues with facts. If 1 + 1 = 3, and everyone says that 1 + 1 = 3, then it assumes 1 + 1 = 3.
@Viskio_Neta_Kafo I assume it’s big data corpus linguistics; each word/phrase is assigned an identifier and then compared to the corpora the LLM holds to see what words are commonly grouped. Linguists have used corpora for decades to quantitatively analyse language; here are some open ones https://www.english-corpora.org/ - the LLM I assume identifies the likely lang “type” to choose a good corpus, identifies question tags & words in key positions, finds common response structures and starts building.
You are both right, but this armchair psychologist thinks it’s similar to how popular skeuomorphism was in the early day of PC guis and such compared to today.
I think many folks really needed that metaphor in the early days, and I think most folks (including me) easily fall into the trap of treating LLMs like they are actually “thinking” for similar reasons. (And to be fair, I feel like that’s how they’ve been marketed at a non-technical level.)
@octopus_ink yes I think we will eventually learn (there is clearly a lot of pushback against the idea that AI is a positive marketing term), and it’s also definitely the fault of marketing, to try to condition us into thinking we desperately need a sentient computer to help us instead of knowing good search terms. I am deeply uncomfortable with how people are using LLMs as a search engine or a future prediction machine.
@manicdave Even saying it’s “trying” to do something is a mischaracterisation. I do the same, but as a society we need new vocab for LLMs to stop people anthropomorphizing them so much. It is just a word frequency machine. It can’t read or write or think or feel or say or listen or understand or hallucinate or know truth from lies. It just calculates. For some reason people recognise it in the image processing ones but they can’t see that the word ones do the exact same thing.
Forgive my ignorance but using just the frequency of words how does it come up with an answer to a question like “are sweet potatoes good for you and how do you microwave them in a way that persves their nutrients?”
Does it just look for words that people online said regarding the question or topic?
Basically, yes.
If I were an alien and you walked up to me and said, “Good Morning”, and I looked around and everyone else said “Good Morning”, I would respond with "Good Morning ". I don’t know what is “Good” or “Morning”, but I can pretend I do with the correct response.
In this example “Grok” has no context on what is going on in the background. Musk may have done nothing. Musk may have altered the data sets heavily. However the most popular response, based on what everyone else is saying, is that he did modify the data. So now it looks like he did, because that’s what everyone else said.
This is why these tools have issues with facts. If 1 + 1 = 3, and everyone says that 1 + 1 = 3, then it assumes 1 + 1 = 3.
@Viskio_Neta_Kafo I assume it’s big data corpus linguistics; each word/phrase is assigned an identifier and then compared to the corpora the LLM holds to see what words are commonly grouped. Linguists have used corpora for decades to quantitatively analyse language; here are some open ones https://www.english-corpora.org/ - the LLM I assume identifies the likely lang “type” to choose a good corpus, identifies question tags & words in key positions, finds common response structures and starts building.
You are both right, but this armchair psychologist thinks it’s similar to how popular skeuomorphism was in the early day of PC guis and such compared to today.
I think many folks really needed that metaphor in the early days, and I think most folks (including me) easily fall into the trap of treating LLMs like they are actually “thinking” for similar reasons. (And to be fair, I feel like that’s how they’ve been marketed at a non-technical level.)
@octopus_ink yes I think we will eventually learn (there is clearly a lot of pushback against the idea that AI is a positive marketing term), and it’s also definitely the fault of marketing, to try to condition us into thinking we desperately need a sentient computer to help us instead of knowing good search terms. I am deeply uncomfortable with how people are using LLMs as a search engine or a future prediction machine.
Exactly. Grok repeatedly generate a set of numbers, which, when keyed against its own list of words, spells out that Musk is spreading misinformation.
It just happens to be frequently…