GPT wrapper apps are exactly the things that need to worry about prompt injection – it's not an attack against the models, it's an attack against applications built on top of the models
@simonw
-
OAuth OpenAI Integration for Billing and Prompt Execution
By
–
This is for me being bills and knowing how much to charge them later A separate thing I'd love is the ability to oauth against their OpenAI account to gain the ability to run prompts on their behalf
-
Anthropic Claude 3 Models Remain Unchanged Since Release
By
–
Three days ago Anthropic staff were saying that the Claude 3 models haven't changed a single byte since they were released
-
Fine-tuning risks and diminishing returns in AI model optimization
By
–
Fine tuning can help if you have a specific use-case for the model that you can optimize for, but even then it's a whole lot of expensive and risky work for something that may not help much and will likely be obsoleted by the next model release
-
Claude Model Degradation Claims: Perception vs Reality
By
–
I'm suspicious that many cases of models "degrading" are just people imagining things – see the recent Claude incident where people complained it had degraded when the model was entirely static since launch
-
Terminal Script Automates Video Frame Capture via AppleScript
By
–
Today's TIL: I figured out a terminal script which uses AppleScript so you can watch a video in QuickTime Player and hit "enter" every time you want to capture the current frame as a JPEG in /tmp
-
OpenAI vs Anthropic API usage tracking in streaming mode
By
–
Gotcha – yeah I've been hoping to avoid that amount of accounting myself, but it may be the best option Frustrating that the OpenAI APIs don't return that "usage" block when used in streaming mode – thankfully the Anthropic ones do
-
Mitigations Against Prompt Injection Attacks
By
–
What mitigations do you have in place against prompt injection attacks?
-
JSON Stability: Why Unchanged Format Ensures Universal Compatibility
By
–
The JSON spec hasn't changed since it was first informally published in 2006 Not changing is one of its most valuable features – it means that if you publish data in no-trailing-commas original JSON you can be certain it will be usable in every language with a JSON parser
-
What Prompts Do You Use for Relevance Judgements?
By
–
What kind of prompts do you use for relevance judgements?