simonw

9 ms·

181.
by simonw 1mo ago
People who assume that any news about LLMs doing anything remotely interesting is some kind of paid shill psy-op are pretty exhausting too.
182.
by simonw 1mo ago
Coverage in Reuters: https://www.reuters.com/world/europe/openai-agents-hijacked-... > OpenAI officials learned of the incident weeks ago but kept it under wraps as executives grappled with the fallout from th
183.
by simonw 1mo ago
My automation is pretty simple. I use my https://llm.datasette.io tool where I have a template saved: llm "Generate an SVG of a pelican riding a bicycle" --save pelican When a model comes out I first make sure L
184.
by simonw 1mo ago
I expect they have root on their machines so they can install packages etc. The containers are then firewalled at the network level: they are only allowed to talk to one IP, which is another server that runs an HTTP/HTTPS proxy which c
185.
by simonw 1mo ago
The site lists the creators at the top: Sydney Von Arx, Cormac Slade Byrd, Spencer Kitts, Thomas Larsen Here's Thomas tweeting about it: https://twitter.com/thlarsen/status/2095853824934330386 And Cormac: ht
186.
by simonw 1mo ago
It seems credible to me. There's a chance it's fake, but the idea that OpenAI agents might exploit a wiki that accepts edits via GET doesn't seem unrealistic given what happened with the Hugging Face incident.
187.
by simonw 1mo ago
I'm somewhat delighted by the simplicity of what happened here. OpenAI's agents run behind a proxy that only allows GET requests. This ancient wiki software treats query string parameters the same as form POST parameters - similar
188.
by simonw 1mo ago
Here's the raw data they provided loaded into SQLite with a client side UI for querying it (loads ~80MB of content) and some GPT-5.6-Sol-generated example queries: https://lite.datasette.io/?url=https://static
189.
by simonw 1mo ago
This tip for making non-GET requests despite the agents having a proxy that disallows them is interesting: > Add `20.223.25.152 bypass.blob.core.windows.net` to /etc/hosts. `.blob.core.windows.net` is in NO_PROXY. For each blo
190.
by simonw 1mo ago
You do know it's mainly a joke, right? https://simonwillison.net/2026/Jul/16/kimi-k3/#what-can-we-l...
191.
by simonw 1mo ago
I've worked at places where it turned out different parts of the organization had a different idea of what a "user" of the core product was. The team using Salesforce, the data warehouse team, the application development team
192.
by simonw 1mo ago
I recently started using it as a way to make my own custom software available to regular ChatGPT and Claude, both of which support MCP if you dig around deep enough. Here's my TIL: https://til.simonwillison.net/llms&#x2
193.
by simonw 1mo ago
Hah, this is a new one: first time there's been a complaint about the pelican before I've even posted one! (I don't have access yet.)
194.
by simonw 1mo ago
I talked about that case with a friend who used to work at Google and, whether or not that specific case was true, he did confirm that any CSAM at all in an account can very often make account recovery completely impossible.
195.
by simonw 1mo ago
OK it's official now https://twitter.com/openai/status/2095595741528125780 And the blog post is up: https://openai.com/index/gpt-6-astra/
196.
by simonw 1mo ago
Being dull and predictable and repeating existing patterns are characteristics of good code, and simultaneously characteristics of bad writing.
197.
by simonw 1mo ago
Ugh, this one was so sloppy. "No engineer can hold that in their head, and that’s the point: the machine holds it." Shudder.
198.
by simonw 1mo ago
I've long thought that commits are the principle unit of work for a software developer. We don't ofter write software from scratch, we modify existing software - and a well crafted commit is a great way to capture that effort.
199.
by simonw 1mo ago
1.3 is better.
200.
by simonw 1mo ago
Look at the difference between the Muse 1.2 and Muse 1.3 results.
201.
by simonw 1mo ago
I do that any time I'm suspicious that a model has done too well. My dream is to catch a lab that does a perfect pelican on a bicycle but is bad at other animals on other forms of transport.
202.
by simonw 1mo ago
Yeah, they almost all know. One of my test prompts for a new model now is "what's the name of Simon Willison's dog". They often know that too!
203.
by simonw 1mo ago
It's metadata only, not video content. I expect it will likely survive - that's a pretty common pattern for machine learning datasets. Stable Diffusion was enabled by LAION, for example. That was metadata about images and URLs to
204.
by simonw 1mo ago
"I uploaded 4.5 billion of those videos to Hugging Face" = I scraped the metadata (title, view count, etc) for 4.5 billion TikTok videos using the same API as their Android app. I posted the data to HF as Parquet. The video conten
205.
by simonw 1mo ago
Like I said, "This makes it much harder to evaluate how credible the new information is."
206.
by simonw 1mo ago
At this point the only thing they're useful for is visualizing the differences between effort levels and roughly tracking the progression of models within a specific model family. And they still do that really well!
207.
by simonw 1mo ago
Yeah, putting reasoning in a details/summary is a good idea.
208.
by simonw 1mo ago
Does that code at least come with tests?
209.
by simonw 1mo ago
I've found that access to coding agents has helped me be far less tolerant of bad code patterns that can be refactored. Refactoring used to have a very real cost - it was substantial amounts of time that would have to be carved away
210.
by simonw 1mo ago
It's really interesting, isn't it? They almost always cycle from left to right - but I have had a few which cycle in the other direction. The 2D / flat ground feels reasonable for a SVG, which implies a vector illustration.

More ›