Claude Sonnet 4.5 will terminate sessions when it finds it is skipping it's own alert warnings. 4.1 didn't do this.
Basically it's broken as a research tool now, unless you want to be mediated by reddit class IQ for your research.
EDIT: Actually, more research it doesnt do this, but they changed the search to use up tokens very fast. 1 search can eat 100k tokens out of 180k token max, depending on the search, so I think they changed it from "first N kbytes" or whole page or much larger and so its a bigger hit and likely to kill the session.
I heard that Claude Code can be installed locally. One case study asked Claude to examine and analyze personal journal entries, something I would never upload to ChatGPT. This is an interesting development that opens many possibilities. An installation is apparently fairly straightforward, but it does require one other software program to run.
Claude Code is not running the model locally -- it still sends everything to the claude servers for inference same as if you use the browser UI (from a privacy perspective). If you want to avoid that, you'll need to self-host. Some models can run on desktop machines, or you can "self-host" in the cloud (still someone else's machine, but much much better if you care about this). Ollama is probably the best starting point https://docs.ollama.com/quickstart, and you can use it to run some the most current open source models w/ cloud GPUs: https://fly.io/blog/scaling-llm-ollama
Thanks for this clarification. I’d misunderstood. I’d like to self-host but the mechanics of it are a bit above my pay-grade. I’m already lost at the mention of Ollama… (what do I need that for? I just want to run Claude!) I think I’m a lost cause.
Claude is closed source inference, meaning there's no way to self-host. You can only self-host the open source models (like Deep Seek). Ollama is a tool to try to make it easier to do that, for any open source model.
It's possible to set up "private cloud" deployments that provide privacy, but it's not cost effective -- you pay a premium for the privacy. Perhaps there's enough demand though to make it worth setting something up that's easy to use / configure. Something like this, but more retail user focused: https://elest.io/open-source/ollama
This could be streamlined / productized if there was a sufficient interest. Will not be price competitive with open ai / claude (probably a few dollars an hour based on usage), but maybe that's compelling to some.
See, I saw this video of a guy installing Claude Code and then asking it to read his notes and analyse them for consistent themes or ideas, and then asking it to bulk rename or re-tag his thousands of Obsidian notes, and I thought, hmm I could use some help like that! He implied that it was not sending anything private to a server but it looks like he was mis-informed or he wasn't giving the full picture of his setup.
I made two new novel entries on my blog just after the new version came online (Diana part 2 and Arkassia part 2, which are the second parts of two unrelated stories):
It appears to be significantly better than before. Furthermore, Claude now has a much longer length limit before having to start a new chat, making it much more convenient to use without having to repeatedly remind it of various things.
https://open.substack.com/pub/hamtechautomation/p/claude-sonnet-45-the-ai-model-that?r=64j4y5&utm_medium=ios
Claude Sonnet 4.5 will terminate sessions when it finds it is skipping it's own alert warnings. 4.1 didn't do this.
Basically it's broken as a research tool now, unless you want to be mediated by reddit class IQ for your research.
EDIT: Actually, more research it doesnt do this, but they changed the search to use up tokens very fast. 1 search can eat 100k tokens out of 180k token max, depending on the search, so I think they changed it from "first N kbytes" or whole page or much larger and so its a bigger hit and likely to kill the session.
I heard that Claude Code can be installed locally. One case study asked Claude to examine and analyze personal journal entries, something I would never upload to ChatGPT. This is an interesting development that opens many possibilities. An installation is apparently fairly straightforward, but it does require one other software program to run.
Claude Code is not running the model locally -- it still sends everything to the claude servers for inference same as if you use the browser UI (from a privacy perspective). If you want to avoid that, you'll need to self-host. Some models can run on desktop machines, or you can "self-host" in the cloud (still someone else's machine, but much much better if you care about this). Ollama is probably the best starting point https://docs.ollama.com/quickstart, and you can use it to run some the most current open source models w/ cloud GPUs: https://fly.io/blog/scaling-llm-ollama
Thanks for this clarification. I’d misunderstood. I’d like to self-host but the mechanics of it are a bit above my pay-grade. I’m already lost at the mention of Ollama… (what do I need that for? I just want to run Claude!) I think I’m a lost cause.
Claude is closed source inference, meaning there's no way to self-host. You can only self-host the open source models (like Deep Seek). Ollama is a tool to try to make it easier to do that, for any open source model.
It's possible to set up "private cloud" deployments that provide privacy, but it's not cost effective -- you pay a premium for the privacy. Perhaps there's enough demand though to make it worth setting something up that's easy to use / configure. Something like this, but more retail user focused: https://elest.io/open-source/ollama
I see (I think). Thanks for taking the time to elucidate.
This could be streamlined / productized if there was a sufficient interest. Will not be price competitive with open ai / claude (probably a few dollars an hour based on usage), but maybe that's compelling to some.
See, I saw this video of a guy installing Claude Code and then asking it to read his notes and analyse them for consistent themes or ideas, and then asking it to bulk rename or re-tag his thousands of Obsidian notes, and I thought, hmm I could use some help like that! He implied that it was not sending anything private to a server but it looks like he was mis-informed or he wasn't giving the full picture of his setup.
I made two new novel entries on my blog just after the new version came online (Diana part 2 and Arkassia part 2, which are the second parts of two unrelated stories):
https://offmodernity.substack.com/
It appears to be significantly better than before. Furthermore, Claude now has a much longer length limit before having to start a new chat, making it much more convenient to use without having to repeatedly remind it of various things.