[Discuss] LLMs: Claude Code, and a Tip

[kentborg at borg.org (Kent Borg)]

On 2/8/26 3:38 PM, Steve Litt wrote:
>> P.S. One can change one anniversary date by cancelling and
>> resubscribing. I don't know what context might be lost in the process.
> I know that ChatGPT has no facilities for backup, so you have to do
> manual copy and paste, and their CSS use makes those pastes pretty hard
> to read. This is not unintentional, and I'll bet you 2 to 1 that if not
> now, then very soon, when you stop paying the money, all your info
> disappears. I'll bet Claude will be the same.

Claude in a browser certainly has limitations along the lines you 
describe. But Claude Code, though it has (pretty much?) the same LLM 
backend as the browser version, also has a binary installed on my 
computer. (Well, in a rather isolated VM.) It has read-write permissions 
in the source directory tree I am working on, and in its own ~claude 
directory, but no where else*. When I fire up a new instance of the 
LLM?which I do frequently, because managing how big the context is and 
what is in it is important, a lot of what I am learning?it can see what 
has been created up to that point, make changes, run tests, commit the 
changes to git (locally only), etc.. In a sense the local software does 
all that copying and pasting (and grepping, and much more) on behalf of 
the LLM. It makes it practical to do real work.

Anthropic would like me to use Claude Code for everything, let it lose 
as user "kentborg", on my real machine, on all my real data, doing 
everything! No. I won't do that. I am not as negative on LLMs as you 
are, but I can go on at great length and depth about the problems and 
limitations they have, more so every day. Don't mistake me for one of 
the AI bros.


> One more thing: Put everything you learn into one or many text files.
> Then, when you get a computer with 64GB RAM or 128GB RAM and a big
> processor, you can install a SLM (Small Language Model) on your
> computer [?]

And don't forget to buy some powerful hardware for doing lots of matrix 
multiplication very quickly and a special memory architecture to keep it 
fed with new matrices to be multiplied (an Nvidia card, for example). I 
periodically muse in that direction, and were I dealing with more 
personal data I would have to do that.

(I do have another VM with Ollama installed on it, BTW.)

As for looking through my own content, I recently realized that a vector 
database, and enough of AI to do embeddings, is maybe all I need for 
that, so maybe no AI layer on top of it. In which case no need for 
special hardware.

Again, I am *not* part of the AI hype. It is a crazy, unsustainable 
bubble based on an innovation that is a dead-end on any road to "general 
artificial intelligence". But I am discovering that they do have value. 
My current project is to better understand it in practical terms.


-kb, the Kent who was never on the bitcoin bandwagon, either, because 
once he actually learned about it he concluded it was stupid.


* Running Claude Code as its own user is decidedly non-standard. That is 
me being conservative.