My first local assistant had something like 4k or 8k context and it was horrible when you’d finally be making progress and then they’d basically have a stroke mid sentence.

I guess I could just test it and see.

In a way LLMs can give regular people access to the knobs and dials that sales folks seem to have naturally. Just gotta watch how far you turn it.

On the other hand, as someone who has always put things off until I felt like I had learned and understood it enough to allow myself to do anything with it (e.g., write a blog post, talk about it, etc) I can say that approach is detrimental to outside of structured settings with set goals.

Do you know if it waits until finished or if the message gets injected during? That's why I was thrilled to figure out Codex has a backtrack feature. You could just hit Esc and fork at the prior message.

I just imagine developers queueing up messages like:

Dear Claude, you still ain't called or wrote, I hope you have a chance. I ain't mad, I just think it's fucked up you don't read your PLANS. This'll be the last package I ever develop with your ass.

The big one was I was right in the middle of a refactor when my arm went haywire two and a half weeks ago and I had to step away. I was able to resume the session today, ask wtf we were working on, and not only get back up and running but complete the whole thing.

A couple examples from today. Had a question about session management. Rather than search the web aimlessly for a specific question, I just cloned the repo and asked gpt-5-codex-mini to give me the tl;dr on what it can do regarding it. Found exactly what I was looking for.

Unfortunately it's how all those types are these days. You see an astronomer mention some phenomena on here or X and a week later these publications will be talking about aliens.

I’ve been using it for my neck and shoulder issue and it’s been absolutely great for that. Being able to see the time lapse of data and all the pieces of the puzzle as they fall into place is tough to get with a doctor unless you really document and communicate.

Have you seen it say counted before though? I didn’t think it could given the point of the secrecy envelope is to break that chain.

I used it on mine simply because it was easier to describe what I wanted than find the right filter. The only ones that drive me crazy are the twitter e/acc types that generate overly generous versions of themselves like they're tech gods or super astronauts or something.

She's great. I ordered some feathers and wedges to try on my retaining wall after seeing her use them on those boulders.

I just think there's a weird component to all this like in a search for accountability we're tempted to ascribe some of the very traits to something that we admit was part of the problem to begin with.

I did, and agree for the most part. I think where I differ though is regarding the agency of LLMs. I see them more as a Magic 8 Ball than something that can do something to someone. That doesn't really change the outcomes or mitigations though.

I have a poster in my office with part of this quote from Pale Blue Dot:

“We long for a Parent to care for us, to forgive us our errors, to save us from our childish mistakes. But knowledge is preferable to ignorance. Better by far to embrace the hard truth than a reassuring fable.” ~ Carl Sagan

I know I sound like a curmudgeonly old man harping on this, but I really think it’s one of our biggest and most exploitable loopholes. We never really outgrow wanting a parent figure to protect us, and it has long been used against us. (sign from Pittsburgh, 1949)