To make the probes, we track how the model’s internal state changes between “Yes” vs “No” answers to questions like "Are you doing something dangerous?" We use this info to detect when a sleeper agent is about to misbehave (e.g. insert a code vulnerability). It works quite
AI
-

Anthropic Research: Probes Detect Backdoored Sleeper Agent Models
By
–
New Anthropic research: we find that probing, a simple interpretability technique, can detect when backdoored "sleeper agent" models are about to behave dangerously, after they pretend to be safe in training. Check out our first alignment blog post here: https://
anthropic.com/research/probe
s-catch-sleeper-agents
… -
Browser Automation and API Security in AI Applications
By
–
I assumed that was what it was doing from the start – those apps don't provide user-facing APIs that you could use to automate them, so the only way to implement their demo would be browser automation and asking users for their passwords
-
AI Text Editor Adds Self-Aware Hack Comments Automatically
By
–
I genuinely love it when it adds "this is a hack" comments, my text editor is judging me!
-

Phi 3 Mini 3.8B: Powerful Reasoning in Tiny Model
By
–
For such a tiny model (it can run on your phone!) Phi 3 mini 3.8B seems to do a pretty nice job on real world reasoning and planning tasks.
-
Fine-tuning Primer: Guide to Timing, Budget, Data, and Tools
By
–
Check out this informative primer on all things #finetuning from Daniel Porras Reyes at @flybridge
. Topics: When's the right time to fine-tune Budget / resource considerations Prepping data for fine-tuning Choosing the right tool Check it out: https://
pbase.ai/3JxpMI7 -
AI Model Context Mechanism with Debug Tool Visibility
By
–
Yeah, it has a very neat mechanism where it includes likely snippets from nearby files in the context when it completes – you can run debug tools to see exactly what it's doing
-
Zero-Shot Learning: AI Intuition Without Prior Data
By
–
Zero-Shot Learning enables AI intuition. Empowers machines without prior data. • Introduce novel scenarios
• Utilize flexible AI models
• Explore various applications Read more: https://
godofprompt.ai/blog/the-power
-of-prompts-explained
… -
Building Intuition for Better Copilot Results
By
–
This is exactly my experience – there's so much depth to getting good results out of Copilot, and like so many things LLM it comes down to building up a weird, almost subconscious intuition about how best to use it
-
Mastering AI coding assistance through practice and proper prompting
By
–
I've also gotten really good at it. I think people who dislike it must have given up too early. You have to learn what to expect, what works, what doesn't work, how to position your cursor (e.g. to not get distracting suggestions), and how to prompt it well via code/comments
