Antony Lewis
@cosmologist.info
Professor of cosmology, University of Sussex. CMB, lensing and theory of observations.
Next time try a coding agent eg. Augment Code (free quota) or Cline. Etc. Esp. if you are using free ChatGPT which may be reverting to mini llm and using small context window. Agents can also test code for you. (But yes, results can be high variance from almost perfect to useless)
August 13, 2025 at 9:29 PM
Next time try a coding agent eg. Augment Code (free quota) or Cline. Etc. Esp. if you are using free ChatGPT which may be reverting to mini llm and using small context window. Agents can also test code for you. (But yes, results can be high variance from almost perfect to useless)
Not that short, eg you can see Claude’s at docs.anthropic.com/en/release-n...
System Prompts - Anthropic
See updates to the core system prompts on [Claude.ai](https://www.claude.ai) and the Claude [iOS](http://anthropic.com/ios) and [Android](http://anthropic.com/android) apps.
docs.anthropic.com
August 6, 2025 at 8:27 PM
Not that short, eg you can see Claude’s at docs.anthropic.com/en/release-n...
Any data on how much regenerative breaking reduces breaking friction particulates?
July 9, 2025 at 6:20 AM
Any data on how much regenerative breaking reduces breaking friction particulates?
Nice alternative would be to go back to first message and initial context and then auto summary of progress made so far.
May 28, 2025 at 9:16 PM
Nice alternative would be to go back to first message and initial context and then auto summary of progress made so far.
Can confirm. Over the weekend I took a 15 year old crossword solver Java app, used agent Auto mode to restructure file base for vs code and gradle, modernised all the code and UI, build new multi-platform ci installation builds. After few manual fixes, only left with subtle high-dpi drawing issue.
May 12, 2025 at 7:27 PM
Can confirm. Over the weekend I took a 15 year old crossword solver Java app, used agent Auto mode to restructure file base for vs code and gradle, modernised all the code and UI, build new multi-platform ci installation builds. After few manual fixes, only left with subtle high-dpi drawing issue.
Impressive model, but the knowledge cutoff date quoted in AI studio is wrong and has been for months. (If you ask it says 2023, which probably also wrong!)
May 6, 2025 at 7:40 PM
Impressive model, but the knowledge cutoff date quoted in AI studio is wrong and has been for months. (If you ask it says 2023, which probably also wrong!)
How long does it take to “uncensor” a model a la perplexity/microsoft versions of Deepseek? Would be great to see an effectively automated pipeline to produce western versions quickly. (AFAIK none for Deepseek v3 update yet.)
April 29, 2025 at 6:26 AM
How long does it take to “uncensor” a model a la perplexity/microsoft versions of Deepseek? Would be great to see an effectively automated pipeline to produce western versions quickly. (AFAIK none for Deepseek v3 update yet.)
They might effectively get some of that from chats - every “help me fix my washing machine” chat is a human-proxied interaction with the world? Though very unclear to me how they actually use the very noisy chat data.
April 29, 2025 at 6:20 AM
They might effectively get some of that from chats - every “help me fix my washing machine” chat is a human-proxied interaction with the world? Though very unclear to me how they actually use the very noisy chat data.
Interesting, how does it compare with Cherry Studio? I just today managed to make an English-based fork of that at github.com/LaChatterie/...
GitHub - LaChatterie/Verger
Contribute to LaChatterie/Verger development by creating an account on GitHub.
github.com
April 27, 2025 at 8:52 PM
Interesting, how does it compare with Cherry Studio? I just today managed to make an English-based fork of that at github.com/LaChatterie/...
The knowledge cutoff in the shot and AI studio seems wrong, seems 2023 not Jan 25. Good model though!
April 9, 2025 at 10:15 PM
The knowledge cutoff in the shot and AI studio seems wrong, seems 2023 not Jan 25. Good model though!
It did one shot one of my toy research-level test maths questions from a recent paper, previously had to do multi-tries with a multi-agent system. But some mixed experience with coding and other things.
April 6, 2025 at 7:05 AM
It did one shot one of my toy research-level test maths questions from a recent paper, previously had to do multi-tries with a multi-agent system. But some mixed experience with coding and other things.
So there may be something interesting going on at low redshift, as also indicated by the H0 tension and some odd local densities/flows. But it seems premature to favour poorly-motivated evolving dark energy based on CMB+SN+DESI when this doesn't really help with the other low-redshift oddities.
April 2, 2025 at 8:17 PM
So there may be something interesting going on at low redshift, as also indicated by the H0 tension and some odd local densities/flows. But it seems premature to favour poorly-motivated evolving dark energy based on CMB+SN+DESI when this doesn't really help with the other low-redshift oddities.
The stronger evidence for evolving dark energy really comes from adding low-redshift supernovae. Combined with DESI this prefers weird dark energy density that *increases* with time (moving predictions to grey regions in figure).
April 2, 2025 at 8:17 PM
The stronger evidence for evolving dark energy really comes from adding low-redshift supernovae. Combined with DESI this prefers weird dark energy density that *increases* with time (moving predictions to grey regions in figure).
The beautiful new ACT CMB spectra are also well fit by ΛCDM. The shifts in parameters adding DESI are slightly unusual, but in an expected direction - there was reason to think the Planck results gave an H0 value that was a statistical fluctuation to the low side.
April 2, 2025 at 8:17 PM
The beautiful new ACT CMB spectra are also well fit by ΛCDM. The shifts in parameters adding DESI are slightly unusual, but in an expected direction - there was reason to think the Planck results gave an H0 value that was a statistical fluctuation to the low side.
Neat, that was fast! ..and interesting it can be worse. Though I was thinking more of PhD level maths conjectures rather than fact statements. I’ve seen outputs that fudge steps to get the output they want (rather like students in exams…)
March 31, 2025 at 10:19 PM
Neat, that was fast! ..and interesting it can be worse. Though I was thinking more of PhD level maths conjectures rather than fact statements. I’ve seen outputs that fudge steps to get the output they want (rather like students in exams…)
Some LLMs seem to do wishful thinking, e.g. asked to show X, will tend to favour X over not X. Hypothesis: using trio of “show X”, “show not X” or “show other” and then converging to best answer can give more reliable results than just “show X”.
March 31, 2025 at 5:30 PM
Some LLMs seem to do wishful thinking, e.g. asked to show X, will tend to favour X over not X. Hypothesis: using trio of “show X”, “show not X” or “show other” and then converging to best answer can give more reliable results than just “show X”.
This was just before dawn, now 7.5% solar this lovely spring morning.
March 30, 2025 at 7:41 AM
This was just before dawn, now 7.5% solar this lovely spring morning.
Yes, lite is based on plik and was with sroll2.
March 28, 2025 at 6:21 PM
Yes, lite is based on plik and was with sroll2.
If you use more of the Planck TT so much closer to actual joint, the combined constraint look more like you might expect - here compared to joint with your new DESI all in LCDM. (I think ACT TT is pulling things more towards high H_0 low Omega_m compared to their TE).
March 28, 2025 at 3:07 PM
If you use more of the Planck TT so much closer to actual joint, the combined constraint look more like you might expect - here compared to joint with your new DESI all in LCDM. (I think ACT TT is pulling things more towards high H_0 low Omega_m compared to their TE).
Could their deep nuclear waste storage be expanded to serve more of Europe? How can Finno-Estonian power transmission be made robust to Baltic sabotage?
March 22, 2025 at 8:02 AM
Could their deep nuclear waste storage be expanded to serve more of Europe? How can Finno-Estonian power transmission be made robust to Baltic sabotage?
Is it really 2/3 total energy, not 2/3 of current electricity?
March 21, 2025 at 11:09 PM
Is it really 2/3 total energy, not 2/3 of current electricity?
I think they’ve now sensibly given up on hydrogen for home heating, but what’s the plan to replace gas for the old small terrace houses like mine with not enough space outside for a ground mounted air heat pump?
March 21, 2025 at 6:09 PM
I think they’ve now sensibly given up on hydrogen for home heating, but what’s the plan to replace gas for the old small terrace houses like mine with not enough space outside for a ground mounted air heat pump?
And more generally the use of wood pellets and the carbon accounting
March 21, 2025 at 6:05 PM
And more generally the use of wood pellets and the carbon accounting