Aaron Meurer
@asmeurer.com
Posting about AI, Python, SymPy, and other nonsense.
asmeurer.com
asmeurer.com
Why is the GitHub mobile interface different depending on whether you're logged in or not?
The tabs move to the top...the star button gets smaller...more links like "code of conduct" appear... WTF?
The tabs move to the top...the star button gets smaller...more links like "code of conduct" appear... WTF?
July 19, 2025 at 5:36 PM
Why is the GitHub mobile interface different depending on whether you're logged in or not?
The tabs move to the top...the star button gets smaller...more links like "code of conduct" appear... WTF?
The tabs move to the top...the star button gets smaller...more links like "code of conduct" appear... WTF?
These new features to argparse in Python 3.14 are nice, but why aren't they enabled by default?
May 10, 2025 at 8:03 PM
These new features to argparse in Python 3.14 are nice, but why aren't they enabled by default?
Very convenient that the "top coding model" doesn't let you upload source code files.
April 14, 2025 at 1:45 AM
Very convenient that the "top coding model" doesn't let you upload source code files.
So apparently one of the reasons so many AIs give this shitty answer for the Python Fibonacci question is that it's in the damn SFT post-training set 🤦♂️
April 9, 2025 at 4:47 PM
So apparently one of the reasons so many AIs give this shitty answer for the Python Fibonacci question is that it's in the damn SFT post-training set 🤦♂️
Finally came up with a use-case for GPT 4o Tasks! Oh...
April 8, 2025 at 5:18 PM
Finally came up with a use-case for GPT 4o Tasks! Oh...
4o is the first model I've seen that can handle this.
March 26, 2025 at 5:13 PM
4o is the first model I've seen that can handle this.
I know $0.0004 is basically nothing, but it's still really stupid that Claude Code costs anything at all just to update itself.
March 24, 2025 at 9:19 PM
I know $0.0004 is basically nothing, but it's still really stupid that Claude Code costs anything at all just to update itself.
Why does ChatGPT DeepResearch waste so much time trying to figure out how to read PDFs?
![Snippet from a DeepResearch activity log:
I'm searching for the figure that's behind a PDF. Considering using Python tools or screenshots to retrieve or convert the image. Citing images properly is crucial.
The PDF content is accessible, likely due to the browser or caching.
Analyzed with python
I'm working in a Python environment with no internet access to pip install packages. I’m curious if PyMuPDF or pdfplumber is preinstalled.
Analyzed with python
I'm thinking about manually finding images in the PDF or using PIL to open the second page. Converting PDF to images with ghostscript is another option.
I’m digging into pdf2image's suitability without poppler. Exploring options!
Searched for pdf2image poppler
I'm exploring offline tools like PyMuPDF for handling PDFs, especially in environments without support for Ghostscript.
Analyzed with python
I’m thinking through why the document stream appears empty and suspect the PDF might be partial due to text line reading. Considering alternative ways to ensure complete content retrieval.
I’m thinking the length was likely printed earlier but might be invisible now due to time passing.
OK, let's look deeper into the request. Clearly, previous results didn't yield output, so it seems the request might have failed.
Considering browser.open with PDF link ID for image embedding. Meanwhile, citing succinctly via [46†L114-L122] is suggested.](https://cdn.bsky.app/img/feed_thumbnail/plain/did:plc:lxs4tiwvmhstjezgrjn4ny7k/bafkreif2sk3j4dpm3tmcpw7iuovflzlof6imgwbh32p2ggjdnolp6xpkeu@jpeg)
March 24, 2025 at 7:35 PM
Why does ChatGPT DeepResearch waste so much time trying to figure out how to read PDFs?
This is really cool, but the best thing about this is that models trained on this are going to learn to talk like Simon from Cracking the Cryptic.
March 21, 2025 at 4:53 AM
This is really cool, but the best thing about this is that models trained on this are going to learn to talk like Simon from Cracking the Cryptic.
Like why would I want to type out all this instead of just typing "jupyter lab"?
March 12, 2025 at 9:38 PM
Like why would I want to type out all this instead of just typing "jupyter lab"?
Where's Waldo benchmark still unbeaten.
March 8, 2025 at 6:43 PM
Where's Waldo benchmark still unbeaten.
If you want to use Aya Vision, SigLIP-2, or SmolVLM-2 with HuggineFace transformers, you have to install a specific tag from GitHub. It feels like Python packaging is regressing 20 years.
March 8, 2025 at 6:25 PM
If you want to use Aya Vision, SigLIP-2, or SmolVLM-2 with HuggineFace transformers, you have to install a specific tag from GitHub. It feels like Python packaging is regressing 20 years.
Apparently I've "contributed" to 199 of the dataset entries, including 37 pull requests (mostly from SymPy).
It feels a little weird. Given how popular this benchmark is, I've probably indirectly prompted an LLM 100x more often than directly.
It feels a little weird. Given how popular this benchmark is, I've probably indirectly prompted an LLM 100x more often than directly.
March 3, 2025 at 8:29 PM
Apparently I've "contributed" to 199 of the dataset entries, including 37 pull requests (mostly from SymPy).
It feels a little weird. Given how popular this benchmark is, I've probably indirectly prompted an LLM 100x more often than directly.
It feels a little weird. Given how popular this benchmark is, I've probably indirectly prompted an LLM 100x more often than directly.
Anthropic rationalizing making Claude 3.7's thinking tokens visible to the user because it "could help with safety research". Whatever. We all know the real reason you're showing them.
!["Supporting Safety Research Displaying Claude’s extended thinking could contribute to ongoing research
on large language model behavior. It enables investigation into why extended chains-of-thought benefit model
performance, including theories about additional memory capacity [16], computational depth through token" from the Claude 3.7 Sonnet system card.](https://cdn.bsky.app/img/feed_thumbnail/plain/did:plc:lxs4tiwvmhstjezgrjn4ny7k/bafkreibqbe5s2fykw6jqn3dfidzduohpv44qzaaagccuc4l37pptu6xwri@jpeg)
February 24, 2025 at 9:05 PM
Anthropic rationalizing making Claude 3.7's thinking tokens visible to the user because it "could help with safety research". Whatever. We all know the real reason you're showing them.
From a house I was looking at during a recent home search. This house was built in 2004.
February 21, 2025 at 8:36 PM
From a house I was looking at during a recent home search. This house was built in 2004.
One of these is not like the others.
February 10, 2025 at 10:11 PM
One of these is not like the others.
One of these is not like the others.
February 10, 2025 at 8:28 PM
One of these is not like the others.
The real reason not to send your data to China.
January 31, 2025 at 8:09 AM
The real reason not to send your data to China.
This has been my experience with it so far too. The thinking is really good at noticing little things that most models would just ignore or take for granted.
January 27, 2025 at 9:42 PM
This has been my experience with it so far too. The thinking is really good at noticing little things that most models would just ignore or take for granted.
Claude would be the perfect model if it weren't for this.
January 14, 2025 at 8:15 PM
Claude would be the perfect model if it weren't for this.