simon.fedi.simonwillison.net.ap.brid.gy
Open source developer building tools to help journalists, archivists, librarians and others analyze, explore and publish their data. https://datasette.io […]
[bridged from https://fedi.simonwillison.net/@simon on the fediverse by https://fed.brid.gy/ ]
985 posts
8,659 followers
2 following
Regular Contributor
Active Commenter
comment in response to
post
@virtuous_sloth the most credible potential solution I have seen so far has echoes of taint, it's still extremely difficult to implement effectively though https://simonwillison.net/2025/Apr/11/camel/
comment in response to
post
We have seen the exact same bug previously in ChatGPT, Google Bard, Writer.com, Amazon Q, Google NotebookLM, Slack, Google AI Studio, Microsoft Copilot, Mistral LeChat, xAI Grok, Claude and now GitLab - I've been collecting examples here https://simonwillison.net/tags/exfiltration-attacks/
comment in response to
post
Anthropic's Hannah Moran finally addressed the elephant in the room at this conference when she subtly dropped "Agents are models using tools in a loop" during the intro to the "Prompting for Agents" workshop https://simonwillison.net/2025/May/22/tools-in-a-loop/
comment in response to
post
I picked up some more details on Claude 4 from a dive through the Anthropic documentation
The training cut-off date is March 2025! Input limits are still stuck at 200,000 tokens. Unlike 3.7 Sonnet the thinking trace is now summarized by a separate model […]
comment in response to
post
I just released llm-anthropic 0.16 (and a tool-enabled 0.16a1 alpha) with support for the two new Claude models, Claude Opus 4 and Claude Sonnet 4: https://simonwillison.net/2025/May/22/llm-anthropic-016/
comment in response to
post
There is an extraordinary amount of detail in there - I included a prompt you can use that reveals the details of what it's gathered about you so far https://simonwillison.net/2025/May/21/chatgpt-new-memory/#how-this-actually-works
comment in response to
post
My first hint at this was when I told it to "dress this dog in a pelican costume" and it added a Half Moon Bay sign in the background without me asking for one
comment in response to
post
Google released a new gemini-2.5-flash-preview-05-20 model today - I added it to my llm-gemini plugin and used it to draw some pelicans riding bicycles, here's the animation I got out of that one https://simonwillison.net/2025/May/20/gemini-25/
comment in response to
post
@troed with AGPL do the rules change if I have "import PyMuPDF as the top line in my Flask app as opposed to the top line in my CLI tool?
comment in response to
post
@troed This is where AGPL makes everything even more complicated: it's different from the GPL (most specifically the network server software part) which means it's unclear how GPL advice applies to it
My hunch is that Python library imports don't trigger it but I don't like sharing my work with […]
comment in response to
post
@Reshirams_Rad_Slam VL is Qwen's vision model series - this is Qwen2.5 VL which came out in January, what's new is that Ollama added support for it a few days ago which makes it easier to try out
Qwen 3 doesn't yet have any vision capacities but I expect we will see a Qwen 3 VL at some point soon
comment in response to
post
@simon Having been through a sufficiently large set of similarly structured problems, even an expert lawyer will in the end tell you that “it’s a question of how much risk you’re willing to accept“. They might even help you translate the risk into financial terms.
If you want to be really sure […]
comment in response to
post
(I have, in the past, made the mistake of trying to get a clear yes or no answer from a lawyer and learned that's not what they do)
comment in response to
post
@sriku.org interesting! I don't see how they could do that since they wouldn't have permission to relicense my code
Where does the barrier sit? I'm fine licensing llm-pdf-to-images under AGPL but I'm not OK doing the same for llm itself - that loads functionality from llm-to-pdf-images via […]
comment in response to
post
@sriku.org Right but the tricky angle here is that I'm not distributing that library as part of my code
I'm distributing a single line, "PyMuPDF" a https://github.com/simonw/llm-pdf-to-images/blob/0ce38f1e0e5effc1ecff0aafd009417d7d78a745/pyproject.toml#L12 - which causes the user's system to […]
comment in response to
post
@sriku.org do you know how that affects library usage?
comment in response to
post
... which inspired me to create and then back-fill a new "hallucinations" tag on my blog https://simonwillison.net/tags/hallucinations/
comment in response to
post
@mschfr oh interesting, I hadn't heard about that one. Do you know if there are clear instructions anywhere for getting it to work on a Mac?
comment in response to
post
... it looks like this may be my bug: using a URL gave me bad results, downloading the file first and passing the filename did a lot better. Still not perfect - it got stuck in an output loop! Results here: https://gist.github.com/simonw/2b46e932a16c92e673ea09dfc0186ec2#response
comment in response to
post
I built a tool to help create these which lets you drop in the slide images, OCR the initial alt text and then edit the alt text and annotations
I wrote it with GPT-4 a couple of years ago, today I gave it a design refresh with Claude 3.7 Sonnet (thinking) […]
comment in response to
post
@jscholes there's some documentation here https://llm.datasette.io/en/latest/tools.html
comment in response to
post
@jscholes in that example the "tools" are Python functions that run in-process - but the system is based around plugins so in the future you'll be able to e.g. "llm install llm-tools-playwright" to get tools that you can start working with directly
comment in response to
post
I hadn't surely tried Gemma 3 4B for vision before this and I'm very impressed at how much it can do for just a 3.2GB model download
comment in response to
post
Got some strange editorial choices for illustrative photos for my prompt "Provide a vegetarian recipe for butter chicken but with chickpeas not chicken and include many inline illustrations along the way"
comment in response to
post
... and the one time I forget to run my "Generate an SVG of a pelican riding a bicycle" eval is the time it produces one of the best results I've seen yet https://simonwillison.net/2025/May/6/gemini-25-pro-preview/#gemini-pelican
comment in response to
post
@Reshirams_Rad_Slam Which article? I missed that one
comment in response to
post
Here's a follow-up by someone with a whole lot more credibility than myself or Andy: Hannah Ritchie is "Head of Research at Our World in Data, and a Senior Researcher at the University of Oxford" and runs a high quality sustainable energy blog […]
comment in response to
post
I had o4-mini write most of the plugin for me, based on an example existing plugin and this prompt. Transcript here: https://gist.github.com/simonw/4f545ecb347884d1d923dbc49550b8b0a