AIHas AI crossed the line into recursive self-improvement?
Berkeley RDI describes Oriol Vinyals’ dividing line: whether a system can rewrite its own harness and weights, not just its output.
TLDR
Berkeley RDI says it has seen plenty of self-improving AI, but not recursive self-improvement. It credits the distinction to Oriol Vinyals, Discovery Loop’s co-founder and CTO and a former Google DeepMind research VP: whether a system can rewrite its own harness—the software around the model—and its weights, or learned parameters, rather than just its output. Berkeley RDI asks whether anything has crossed that threshold yet and what it would take to get there.
Has AI crossed the line into recursive self-improvement?
Berkeley RDI describes Oriol Vinyals’ dividing line: whether a system can rewrite its own harness and weights, not just its output.
TLDR
Berkeley RDI says it has seen plenty of self-improving AI, but not recursive self-improvement. It credits the distinction to Oriol Vinyals, Discovery Loop’s co-founder and CTO and a former Google DeepMind research VP: whether a system can rewrite its own harness—the software around the model—and its weights, or learned parameters, rather than just its output. Berkeley RDI asks whether anything has crossed that threshold yet and what it would take to get there.
