"we consider the separation as a fundamental problem that should be disentangled, conceptually and technically, from other safety training measures, such as rejecting harmful prompt"
@simonw
-
Paper Analyzes Prompt Injection: Separating Instructions from Input
By
–
Saltiness aside, it's a good paper! Digs into the question at the heart of prompt injection: can we safely separate instructions from inout in an LLM?
-
Using LLMs to Convert Blog Content to Academic LaTeX Format
By
–
Casually wondering if I should use an LLM to convert my blog to LaTeX and then email pages from it to arXiv in order to increase my changes of getting cited in papers like this one https://
arxiv.org/html/2403.0683
3v1
… -
GPTs and Fine-tuning: Understanding AI Model Customization
By
–
Gotcha, thanks Yeah, I know about GPTs – I wrote one of those tutorials myself! https://
simonwillison.net/2023/Nov/15/gp
ts/
… I think of fine-tuning as this kind of thing https://
openai.com/blog/gpt-3-5-t
urbo-fine-tuning-and-api-updates
… -
Hybrid Mamba-Transformer Model Addresses Architecture Limitations
By
–
This one looks like a big deal. It's not a pure transformer model – it's a combination of both Mamba and transformer, which appears to address different limitations present in each architecture
-
GPT-4 Fine-tuning Capability Release Discussion
By
–
I didn't think GPT-4 had a fine-tuning capability released yet, how did you do that?
-
GPT prompt workarounds: disabling code interpreter solutions
By
–
Yeah, that's really frustrating. You can prompt it to not do that sometimes, or you can try in a GPT with code interpreter disabled
-
GPT-Vision API Limitations Without Code Interpreter Access
By
–
You can beat that with different prompting, but it's really annoying I've been using GPT-Vision via the API which doesn't have access to Code Interpreter
-
Gemini Pro 1.5 API Video Input Requires Manual Image Conversion
By
–
I thought I might need video input too but it looks like the Gemini Pro 1.5 API currently expects you to video into images yourself
-
Developer discusses branching image processing feature implementation
By
–
Oh interesting! I've been brewing a -i image.png feature in a branch, are there file types I should be considering other than images?