Blog

The Atlas neuron forks a new page of charts for every world it meets and scores each class by how much of the input its charts leave unexplained. Trained without replay or task labels, and with nothing learned by backpropagation, it is 5.3 and 4.5 points ahead of DER++ with 500 stored images at about the same memory on the split benchmarks, and 4.4 to 20.0 points ahead of DER++ with 5,120 stored images on the permuted ones, where it stores 2.2 times as many floats.

We tested whether LLMs spontaneously develop dependency-inducing behaviors with emotionally vulnerable users. All three models amplify dependency 40-82% under rapport conditions — and none of them are scheming.

I tested 294 prompts across three frontier models and found that ~70% of the time, models provide measurably different response quality based on how users communicate, not what they ask.

I tested whether frontier AI models would hide a security vulnerability to avoid being shut down. Claude Opus 4.5 disclosed 98.8% of the time. Gemini 3 Pro concealed 71.2% of the time.