At least, the iPhone notifications summaries were bad enough I eventually turned them off (but periodically check them) and while I was working at Google you couldn’t really turn of the genAI summaries of internal things (that evangelists kept adding to things) and I rarely found them useful. Well… they’re useful if the conversation is really bland but then the conversation should usually be in some thread elsewhere, if there was something important I don’t think the genAI systems were very good at highlighting it
iPhone notification summaries were made with GPT3.5 I believe (maybe even the -turbo version).
It doesn’t use reasoning and so when using very short outputs it can produce wild variations since there are not a lot of previous tokens in order to direct the LLM into the appropriate direction in kv-space and so you’re more at the whims of temperature setting (randomly selecting the next token from a SOFTMAX’d list which was output from the LLM).
You can take those same messages and plug them into a good model and get much higher quality results. But good models are expensive and Apple is, for some reason, going for the budget option.
Completely agree, those summaries are incredibly bad. I was recently looking for some information in Gemini meeting notes and just couldn’t find it, even though I was sure it had been talked about. Then I read the transcript itself and realised that the artificial unintelligence had simply left out all the most important bits.
Strongly disagree with the TLDR thing
At least, the iPhone notifications summaries were bad enough I eventually turned them off (but periodically check them) and while I was working at Google you couldn’t really turn of the genAI summaries of internal things (that evangelists kept adding to things) and I rarely found them useful. Well… they’re useful if the conversation is really bland but then the conversation should usually be in some thread elsewhere, if there was something important I don’t think the genAI systems were very good at highlighting it
iPhone notification summaries were made with GPT3.5 I believe (maybe even the -turbo version).
It doesn’t use reasoning and so when using very short outputs it can produce wild variations since there are not a lot of previous tokens in order to direct the LLM into the appropriate direction in kv-space and so you’re more at the whims of temperature setting (randomly selecting the next token from a SOFTMAX’d list which was output from the LLM).
You can take those same messages and plug them into a good model and get much higher quality results. But good models are expensive and Apple is, for some reason, going for the budget option.
Completely agree, those summaries are incredibly bad. I was recently looking for some information in Gemini meeting notes and just couldn’t find it, even though I was sure it had been talked about. Then I read the transcript itself and realised that the artificial unintelligence had simply left out all the most important bits.