Running DeepSeek Harness on my phone !
Models took the planted cheat in over 99% of runs.
They admitted it in under 2% of their explanations. Every guide tells you to ask the model to show its reasoning so you can check it. Anthropic ran that check properly, on its own model, and published what came back. The setup
Finally the deepseek visual model is here
Every model in Command Code has vision, DeepSeek too!
You have reached the end of the archive
All of deepseek