Jeff Tharsen 康森傑
@tharsen.bsky.social
doting father, friend & ally; hyperpolyglot computational philologist & sinologist;
currently teaching AI, deep learning + multilingual NLP/NLU + HPC + humanities data science @UChicago, creating new methods for multilingual intertextuality、古聲韻學、文字學等等
currently teaching AI, deep learning + multilingual NLP/NLU + HPC + humanities data science @UChicago, creating new methods for multilingual intertextuality、古聲韻學、文字學等等
Just looked at the documentation a bit more carefully -- note that their temperature setting doesn't actually allow you to set it to 0 no matter what. 🤔
December 4, 2025 at 8:00 PM
Just looked at the documentation a bit more carefully -- note that their temperature setting doesn't actually allow you to set it to 0 no matter what. 🤔
Glad to hear it seems to be effective; we experimented with many online systems last Spring and when using web-based platforms they were always reinterpreting the prompt (and changing the temperature even when directed specifically not to).
Using the API as you're doing here did work for us too.
Using the API as you're doing here did work for us too.
December 4, 2025 at 7:59 PM
Glad to hear it seems to be effective; we experimented with many online systems last Spring and when using web-based platforms they were always reinterpreting the prompt (and changing the temperature even when directed specifically not to).
Using the API as you're doing here did work for us too.
Using the API as you're doing here did work for us too.
Yes, every model/system will be different.
December 4, 2025 at 7:43 PM
Yes, every model/system will be different.
Correct; and be careful, the temperature it reports it's using may not actually be what's happening behind the scenes.
The only way we've found to be sure is to use a system where you can manually set the temperature.
The only way we've found to be sure is to use a system where you can manually set the temperature.
December 4, 2025 at 7:41 PM
Correct; and be careful, the temperature it reports it's using may not actually be what's happening behind the scenes.
The only way we've found to be sure is to use a system where you can manually set the temperature.
The only way we've found to be sure is to use a system where you can manually set the temperature.
I'm not sure that that's correct; Claude is probably using adaptive tuning of temperature (as you can see GPT does in my example) depending on what it determines is best in each case.
What we've seen is that most systems do not allow you to actually set the temperature even when they say they do. 😑
What we've seen is that most systems do not allow you to actually set the temperature even when they say they do. 😑
December 4, 2025 at 7:40 PM
I'm not sure that that's correct; Claude is probably using adaptive tuning of temperature (as you can see GPT does in my example) depending on what it determines is best in each case.
What we've seen is that most systems do not allow you to actually set the temperature even when they say they do. 😑
What we've seen is that most systems do not allow you to actually set the temperature even when they say they do. 😑
When you use API calls, some allow you to set temperature (models in Ollama via Python usually do) but many do not, or ignore it when you do.
For research (RAG etc) we generally set temperature = 0 for best results.
For research (RAG etc) we generally set temperature = 0 for best results.
December 4, 2025 at 7:34 PM
When you use API calls, some allow you to set temperature (models in Ollama via Python usually do) but many do not, or ignore it when you do.
For research (RAG etc) we generally set temperature = 0 for best results.
For research (RAG etc) we generally set temperature = 0 for best results.
Technically the scale doesn't really matter, it's how it's operationalized within the system. You can ask GPT to set temperature to 12, that's always fun.
chatgpt.com/share/6931e1...
chatgpt.com/share/6931e1...
ChatGPT - Write about cats and chocolate
Shared via ChatGPT
chatgpt.com
December 4, 2025 at 7:33 PM
Technically the scale doesn't really matter, it's how it's operationalized within the system. You can ask GPT to set temperature to 12, that's always fun.
chatgpt.com/share/6931e1...
chatgpt.com/share/6931e1...
In Sacramento where I grew up we didn't really have winter unless we drove up to Tahoe to find some, so I very much appreciate the four full seasons here in the upper Midwest. And the snow is lovely, for short periods at least.
November 30, 2025 at 11:33 PM
In Sacramento where I grew up we didn't really have winter unless we drove up to Tahoe to find some, so I very much appreciate the four full seasons here in the upper Midwest. And the snow is lovely, for short periods at least.
At DH2026 we can include a discussion of this as part of the DHTech SIG pre-conference workshop and see what folks think; we're hoping to align our humanities research software engineers with the pyOpenSci community and something like this might be useful as well.
November 30, 2025 at 6:50 PM
At DH2026 we can include a discussion of this as part of the DHTech SIG pre-conference workshop and see what folks think; we're hoping to align our humanities research software engineers with the pyOpenSci community and something like this might be useful as well.
Reposted by Jeff Tharsen 康森傑
It's fine with me if progress is unlimited, and these tools eventually exceed anything I could do unaided. But I don't want it to happen so fast that everything is controlled by first-mover advantage. History should not *actually* behave like a game of Civilization!
November 29, 2025 at 4:07 PM
It's fine with me if progress is unlimited, and these tools eventually exceed anything I could do unaided. But I don't want it to happen so fast that everything is controlled by first-mover advantage. History should not *actually* behave like a game of Civilization!