Sonnet 4.5 still cannot be trusted with `cd` but wow is it good

The steady improvement of intent-inference and faithful instruction following has been a fun ride. Implicit style conformance too!

Sonnet 4.5 doesn't have the depth or subtlety of Opus 4.1, but for coding, I rarely reach for Opus now

A pattern I'm liking is giving Claude planning tasks while I do my normal investigations in parallel, and then I compare notes after thinking things through myself. Sometimes it's fruitless, and sometimes the tool catches things or offers ideas that improve things or save me a lot of time.

in this list I'm reminded of some leadership that made dependency hygiene a point of emphasis, and others who insisted it's fine because hard drive space is cheap

Today Claude Code self-reflects poorly when stretched beyond its capacities, even with everything needed in context, like it's over-eager marking unfinished work as complete. Asking it to check its work or "any cleanup?" regularly yields major improvements -- same questions I constantly ask myself.

Claude is eager to please to the point of often denying its own better judgment, so I regularly express uncertainty and ask half-formed questions. Encouraging pushback and requesting tradeoff analyses are helpful to get its engineering sensibilities to override its consumer product priorities.

For any complexity that can't be handled in one context window, or anything I want more visibility into, I have Claude externalize its todos in TODO_SOMETHING.md docs in the root. It can provide as much or as little detail as I want, I can make edits, and it'll check off work and make adjustments.

also `😅`

here's the original conversation, you can see how I had to adjust the prompt to get something I liked

me saying 'more like "robocrap"' is like 'zoom in and search'

I didn't like the taste of the first place, but it was an easy hop from there to something that resonated

claude.ai/share/8336be...

you can point your infoscopes anywhere

Claude invented AI-rhea when I was looking for slop-like terms, and then both it and ChatGPT reflected the meaning on the first try

just a potty humor example, but it's clear LLMs can synthesize new legible info, which many or most users already know, but it seems widely doubted still

Claude didn't come up with infoscope on its own but it did come up with AI-rhea, no google hits at the time, as a vivid specialization of "slop"

I like this framing, and they also do a lot more than information retrieval, they synthesize information in the moment with a complexity we don't see in prior knowledge artifacts -- spatial metaphor is one way I try to think about how the process is more than just lookup

bsky.app/profile/ryan...

Dune feels more relevant despite the lack of G/CPUs

Claude: An infoscope channels humanity's collective intelligence - every conversation adds a star to our shared map of knowledge space. We're all astronomers now, pointing these instruments together, discovering not distant galaxies but the patterns hidden in our own accumulated wisdom.

Claude: An infoscope is an LLM as telescope for knowledge - peering into vast latent spaces where information exists as probability clouds. Different models offer different resolutions: Opus sees deeper, Sonnet scouts wider, working in concert like an observatory array mapping possibility itself.

Opus is like the JWST to Sonnet's Hubble -- it can see more detail, further and deeper -- and this metaphor is nice for explaining why the opusplan mode is great for efficiency, because Sonnet can navigate with chart in hand

I say deep-ish because of course it doesn't cross the boundaries of class instances. Reasoning about the types is more work in the default case, clearly? I can reach for $state.raw np, I just feel I do it more than I would like, like I'm not really getting the vibe, resisting its guidance.

With $state being deep-ish by default, it also seems to say, "it's better to have a performance footgun than a correctness one". From this POV its choice saves me from myself in sometimes-misguided attempts to write optimal code. Is it desirable magic? Close for me, but I can be stubborn