edstarch

I have not. Obviously YMMV, but tools have ~8k tokens of descriptions of their own, including (imo too much) behavioural guidance. The tool descriptions are retained with `--system-prompt ""`, I'm sure you can get Claude to output what's in there exactly as it changes quite a lot between versions.

Did Claude write it for them?

10k tokens of bullshit on "artifacts"

It can be slimmed down pretty aggressively: turn off a few features (claude.ai connector sync, artifacts), work with Claude to disable / defer as many skills and tools you don't need (e.g. SendFeedback), use `--system-prompt ""`. Would be better to just skip all that to begin with but that's the cost of a subsidized fixed price plan.

Pi is probably still the gold standard to build on. Or Hermes if you want to go hog wild with connectors and memory.

I'm by no means anti-anything but I think what the comments are getting at is that these "world models" do not have a world model. It's a fundamental qualitative gap that no amount of compute has been able to close, just mask by everything else getting better.

This doesn't show up as much where the verifiable reward is "code compiles and tests pass" but you can still absolutely see it in things like movies, prose, or slide decks. Even frontier models are pretty bad at keeping track of narrative threads even as the technical quality drastically improves.

This begs the question of whether the gap can be closed just with more compute (presumably yes, or at least errors can be minimized beyond human capacity to see them) and is that going to be worth it with current tech (possibly no).

Our smartest, most capable Exit Offer yet.

Big if true. Opus 5.5 is a lot better than 5.0 but I find it maddening thay no single model released in the past ~year can write prose with a clear and coherent voice. I blame RL on programming tasks.

Sequoia is still supported and very solid right now. I'll probably skip Tahoe and update around 27.5.

Windows UI is the most disjoint though, with designs accumulated over the past 20 years still kicking in various places.

You really should, yeah. I've given up Linux as a daily driver in favor of a MacBook but I do have a work mandated Windows machine and I hate that thing with a passion. I cannot think of a single thing that's better on it than on my MacBook or any Linux distro I've ran as a daily driver.

In fact, most of the time I want to do any tasks which are not directly Teams or MS office related I find it easier to just use WSL.

Such is life in the attention economy.

Why would a human painstakingly craft a text which sounds exactly like an LLM when they can just instruct an LLM to write it?

While sandboxing is definitely more secure... Why not put a global deny on .env-like filename patterns as a first measure?

Only $40?

>I really, truly don't understand. This isn't just about manners, mores, or self-reflection. The inability or unwillingness to think about your behavior or its likely reception are stupefying.

Shower thought: what does a typical conversation with an LLM look like? You ask it a question, or you give a command. The model spends some time writing a large wall of text, or performing some large amount of work, probably asks some follow up questions. Most of the output is repetitive slop so the user scans for the direct answer to the question, or checks if the tests work, promptly ignores the follow-ups and proceeds to the next task.

Then the user goes to an online forum and carries on behaving the same way: all posts are instrumental, all of the replies are just directing, shepherding, shaping and cajoling the other users to his desired end (giving him recognition and a job).

I'm probably reading too much into this one dude but perhaps daily interaction with LLMs also changes how one interacts with other text based entities in their lives.