Interesting and thoughtful policy from Joe Kissell:
7 Likes
Thatās typically Joe; even-handed and level-headed. Iām taking a class (online) about using LLMs for practical purposes like file management, but itās more on the basis of understanding than planning to use them.
Most of the time, Iād rather use Perl, or an automation/shortcuts app.
What I want is a local LLM I train on early copyright free texts Iāve parsed, for textual analysis.
3 Likes
Thatās a post worth sharing. Cheers Joe and thanks Lisa for the steer.
2 Likes
I agree with Joe about āAIs are not I.ā I refer to it as āASā with the āSā meaning āstupidityā.
You can download, install and run many open source LLMs. I have never personally done this, but one good source (which includes a driver app compatible with lots of different models) is Ollama.
There are other ways as well, but Iām nothing even resembling an expert here.
For a very long time (including many posts here), I have refused to use the term āAIā when discussing this tech. I instead use non-emotionally-charged terminology like Machine Learning (ML), Computer Vision (CV) or Large Language Models (LLMs). Because, although these neural-network algorithms can do lots of useful things, there is nothing intelligent about it.
There are some research groups actually trying to create machine intelligence, but that research is far far away from being what todayās purveyors of chatbots are claiming about their products.
One really great book on the subject (from 1995) is Douglas Hofstadterās Fluid Concepts and Creative Analogies. This isnāt so much a book on AI as it is a collection of papers (from Hofstadterās research group) describing various kinds of algorithms intended to explore and understand what it means to be intelligent (whether human or machine).
5 Likes
Yep, it is very susceptible to the āGIGOā process.
Garbage in, gospel out?
3 Likes
Hofstadterās āGodel, Escher, Bachā was huge in my graduating year.
Close, but no cigar! Iām guessing you are just being facetious. 
What a well-written, informative piece by Joe.
He acknowledges his own use of the bots to code, debug and apply in so many ways that LLMs improve our lives. Yet he is spot on when he reasons that a Take Control book by ā or on ā A.I. is ridiculous.
1 Like
With all due respect to Joe, and acknowledging that I would have made exactly the same decision with respect to a book on AI, I am frustrated that itās so hard to keep up with the legitimately exciting things that are happening in the AI world. It seems that at least some developers are able to devote the necessary time to figuring out how to integrate Claude Code into their actual work, but I find that it takes me quite some timeāat least weeks, if not monthsāto figure out the best way to use any new tool for real work. With AI, thatās a tremendously annoying amount of time, since it means that just as Iām feeling as though Iāve wrapped my head around the utility of what Iāve been doing, some new development calls into question the entire scenario.
For instance, Iāve been experimenting with various AI dictation tools. I like the idea of them, and I use dictation heavily on the iPhone for messages and reminders and the like. But for actual writing, it has proven much more difficult since they all work by sending voice off to a server for transcription, which means thereās some delay. It turns out that I donāt actually know what word Iām going to write next until Iāve seen the previous ones in a lot of cases so Iāve had trouble working them into my workflow. Iām not certain if thatās a failing on my part or something to do with their design or something else, but each time I get frustrated with one, another comes out, making me restart the process to see if it somehow works better in my use case.
Ironically, Appleās built-in dictation has been easier for me to use because it works on a word by word basis, showing each word on screen as it recognizes it. The only problem is that itās much less capable than the new tools at understanding how to work with proper nouns and punctuation.
(And if you want to talk about dictation tools, Iāll move that to a new thread.)
2 Likes
I admire your persistence and for lack of better term, impersonal, or agnostic, approach to tech, @ace , and your tolearnce of the frustration you describe. Lots of things these days (
insert topic-hijacking snark here) seem to be developing faster than we can keep up with them (because 
insert random conspiracy theory here)!
The appleOSes/services for example, keep getting more and more complicated, intertwined and imho buggy. So since I am not making my living from tech I am actively moving back to paper and pen, investigating Linux and ādumb phonesā, keeping in mind that I still need to interact with an increasingly digital society.
Still, as a former journalist, I enjoy reading about things Iām not acquainted with and learning how people approach them, so please go ahead with the dictation Topic!
(
remove parenthetical distractions)
1 Like
Slight tangent, but if you havenāt used Agents and Skills directives, Iād recommend looking into them. These are files which many (all?) AI coding agents (Claude Code and Codex for sure) look for and read before working on your request to āBuild me an app thatā¦ā. The best Iāve seen so far for SwiftUI coding in particular comes from Paul Hudson, and is freely available for both platforms.
@LisaS ās reply mentioning file management reminded me of an excellent blog written by Howard Oakly (eclecticlight.co) on the dangers of using AI when youāre not in a position to judge the validity/appropriateness of AI āanswersā: How online search and AI can install malware ā The Eclectic Light Company . Very scary.
And with Claudeās latest advances in discovering zero-day exploits the dangers are not getting any less.
1 Like