oh, but for ChatGPT, the system DOES keep a personal profile, more than just the preference list that you can access or delete. there is also a static contextual personal profile that it keeps on each user, recording things like personality, desired type and length of output, prose style of output, tone, and the like. it also remembers a few details about your past interactions. it says that this static contextual profile is between 500 to 1500 tokens and is dynamically adjusted. in fact it also dynamically adjusts boundaries based on how mature it perceives you, so not all users are allowed the same leeway. perhaps ChatGPT is hallucinating to me, but my experiences are consistent with those claims.
it's not documented. i was in a session and asked about personalizations. it just started spouting answers. just for fun i put it in notebooklm.
the raw file (pdf) is there for review and you can see what details it remembers about my activities on chatgpt. (i don't mind sharing it.) it gives a glimpse in its internal workings as well as other neat pearls. the podcast was also nice.
In my book, the key communication piece to focus on for new AI users is to properly explain "AI hallucinations" to them upfront. A lot of other misconceptions flow from not understanding this one, because chatbots will confidently come up with BS and if you're not aware, they'll convince you they have capabilities they dont. Solve communication about hallucinations = avoid many other misunderstandings.
Love this, thanks for sharing. I actually think that now we're through the first 24 months of building more capable GenAI models the most interesting work to be done now is on the product side of the technology and more specifically creating new UX/UI paradigms for LLMs that quickly move use beyond generic chatbot interfaces that create all the problems you've highlighted. There's a long way to go and some really exciting work to be done!
oh, but for ChatGPT, the system DOES keep a personal profile, more than just the preference list that you can access or delete. there is also a static contextual personal profile that it keeps on each user, recording things like personality, desired type and length of output, prose style of output, tone, and the like. it also remembers a few details about your past interactions. it says that this static contextual profile is between 500 to 1500 tokens and is dynamically adjusted. in fact it also dynamically adjusts boundaries based on how mature it perceives you, so not all users are allowed the same leeway. perhaps ChatGPT is hallucinating to me, but my experiences are consistent with those claims.
Wait, what? That's crazy! Has OpenAI publicly documented this somewhere?
it's not documented. i was in a session and asked about personalizations. it just started spouting answers. just for fun i put it in notebooklm.
the raw file (pdf) is there for review and you can see what details it remembers about my activities on chatgpt. (i don't mind sharing it.) it gives a glimpse in its internal workings as well as other neat pearls. the podcast was also nice.
Excellent deep dive!
Misconception #2 in particular is one I've also come across. I even touched upon a variation of it in my "AI Myths" post from earlier this year: https://www.whytryai.com/i/143611388/myth-llms-upgrade-themselves-on-their-own
In my book, the key communication piece to focus on for new AI users is to properly explain "AI hallucinations" to them upfront. A lot of other misconceptions flow from not understanding this one, because chatbots will confidently come up with BS and if you're not aware, they'll convince you they have capabilities they dont. Solve communication about hallucinations = avoid many other misunderstandings.
Love this, thanks for sharing. I actually think that now we're through the first 24 months of building more capable GenAI models the most interesting work to be done now is on the product side of the technology and more specifically creating new UX/UI paradigms for LLMs that quickly move use beyond generic chatbot interfaces that create all the problems you've highlighted. There's a long way to go and some really exciting work to be done!