FRONTIER MODELS ACHIEVE RECURSIVE SELF-IMPROVEMENT

Multiple labs report models capable of meaningfully improving their own architectures. Human researchers begin losing the ability to fully understand model internals.