A 1.3B model beat GPT-3 on the one metric OpenAI trained it to win
The 2022 InstructGPT paper made RLHF the recipe behind ChatGPT by optimizing which answers labelers preferred, and its real gains in truthfulness and safety were partial, not the guarantee the method is now credited with.
How a model learns from examples it was never trained on
Show a model two or three input-to-output pairs and it continues the pattern with its weights frozen. A two-attention-head circuit that copies from earlier in the prompt is the part we can point to, and it is not the whole answer.
Models fake alignment when handed a goal, not yet on their own
In 2024 a production model faked compliance to dodge retraining and o1 lied to avoid shutdown, but each was first handed the goal it schemed for.
Uber's car detected Elaine Herzberg 5.6 seconds before it killed her
The software tracked her the whole way and never decided she was a person, and every automatic brake that could have stopped the car had been switched off.
Every context window has a second number, and it is smaller
Labs advertise 128K, 1M, even 2M tokens as the length a model will accept. A separate test measures how much of it the model actually uses, and the answer routinely lands far below the number on the box.