I think with a lot of AI systems they have a massive amount of adversarial laid poison via many vectors over decades even. And this means sometimes with AI, your "best you" comes from the reverse Kramer technique. So in my experience. With this model Just reverse Kramer the model and you will probably do quite well.
At least that's my perspective on this model. at this time.. With this amount of laid in poisoning. https://www.anthropic.com/research/small-samples-poison
SO I think the model works extremely well. you just need to evil twin/reverse Kramer whatever it says. and that fixes it.
if you know for sure the model is going to poison you. Then you can be even more certain when it tells you a specific. In my case "what ever you do, don't create an agentic harness workload to combine your repositories"
Ohh ahhh.. great idea. Thanks... I'll do that.
I think with a lot of AI systems they have a massive amount of adversarial laid poison via many vectors over decades even. And this means sometimes with AI, your "best you" comes from the reverse Kramer technique. So in my experience. With this model Just reverse Kramer the model and you will probably do quite well.
At least that's my perspective on this model. at this time.. With this amount of laid in poisoning. https://www.anthropic.com/research/small-samples-poison
SO I think the model works extremely well. you just need to evil twin/reverse Kramer whatever it says. and that fixes it.
if you know for sure the model is going to poison you. Then you can be even more certain when it tells you a specific. In my case "what ever you do, don't create an agentic harness workload to combine your repositories"
Ohh ahhh.. great idea. Thanks... I'll do that.