That's the whole problem. We are developing software against LLMs in the same way we use SQL: we write our own instructions as a programmer, then we combine them with input from our users With SQL we can do that safely. With LLMs we can't
@simonw
-
Demanding Reliable Application Security Solutions
By
–
I disagree: when it comes to application security I think it's acceptable for us to demand solutions that reliably protect us
-
Prompt Injection Filter Risks vs Spam Filter Consequences
By
–
That only works if the damage caused by the occasional attack getting through the filter is acceptable A spam filter missing an email = you see one spam email in your inbox A prompt injection filter missing an attack could = now your private data has been stolen
-
Qwen 4bit and 6bit Model Memory Requirements Guide
By
–
The 4bit model should work comfortably on 32GB, the 6bit might just fit on that too If you have 16GB or less you're out of luck for this particular Qwen model
-
Claude 4 Best Model for Code According to Developer
By
–
Claude 4 is probably the best model for code right now – I suggest trying Gemini 2.5 Pro and OpenAI o3 if Claude 4 isn't working for you, but they are all about the same level as each other
-
Design Patterns Securing LLM Agents Against Prompt Injections
By
–
I like the way this "Design Patterns for Securing LLM Agents against Prompt Injections" paper puts it: https://
simonwillison.net/2025/Jun/13/pr
ompt-injection-design-patterns/#scope-of-the-problem
… -
LLM Security: Token Injection Risks and Tool Access Control
By
–
I have yet to see any truly credible protection for this, and I've been looking! You have to assume that anything that can get tokens into your LLM system will be able to trigger any tool that system has access to
-
Context Window Performance and RAM Requirements for AI Models
By
–
60 tokens a second, context window is supposedly up to 262,144 but that will use a ton more RAM, probably more than my Mac can handle (I've not tried though)
-
Qwen3-Coder-30B Demonstrates Strong Capabilities Without Thinking
By
–
For a non-thinking model I'm finding Qwen3-Coder-30B to be surprisingly capable Thinking models are a pretty new invention, up until about 8 months ago we didn't have any at all
-
Hosted AI Models Approaching Quality Parity with Alternatives
By
–
They're getting pretty close, if they're not there already – depends on how sensitive you are to differences in quality, the hosted models are still more capable