Just last week, OpenAI’s ChatGPT-4 saved me 4 hours of work. It is like having a not-too-bright but hardworking undergrad student who’s trying to impress you. So, trust but verify — which is good enough for me! In any case, it was more than worth the monthly subscription price.
However, for creating graphics, MidJourney is still miles better. I use it weekly for some podcast artwork and what it generates is miles better than the plastic-y, obviously AI-generated nonsense ChatGPT produces. See here for some examples of the latter, and this is not a comment on the newsletter itself which is quite good and would be even better were it not for the distracting visuals. Or maybe I just like my own prompts better and someone would think the same of the stuff I like?
Microsoft’s Copilot has so far been absolutely useless — a slightly less intrusive Clippy. Note that I only tried to use it in Word and Outlook, and its value in Excel and other more code-like environments may be higher. Copilot’s constraint to use only what’s in the document in which you are using it is particularly limiting, especially if you need it to fill some blank space. Having it be able to browse the internet would be nice, or at least create a collection of documents, NotebookLM-style, which it could access.
Speaking of NotebookLM, formerly known as Tailwind, I haven’t used it nearly as much as I expected I would. Having a 10-document limit for each collection is too constraining, so it never became my default place to ask questions. Now, if DEVONtechnologies built something like that into DEVONthink and was able to train an LLM on my document collection, I would probably use it first, then ChatGPT, and leave everything else for playing and experimentation.
And something I haven’t used yet but might is Claude 3, which some on X have said achieved AGI level which, really, no (see the last link for Tyler Cowen’s reasons why, with which I agree), but it does look quite good. Note, however, that even supposedly simpler LLMs like Kagi’s FastGPT gave me the “correct” answer to a question that’s in the is-this-an-AGI battery of tests while also being, well, faster.
So, my actual real-life use of generative AIs to date:
- (doesn’t exist yet but would be great if it did) LLM trained on my DEVONthink database — let’s call it DEVONai
- ChatGPT-4
- MidJourney
- FastGPT
- …
- NotebookLM
This is for matters of productivity only. As I wrote last week, LLMs have already received entertainment chatbot perfection and any further “improvement” on that end would be to our detriment.