a16zRecursive Self-Improvement and the Future of AI
a16z
RSI as self-improvement 0:00
The discussion starts with the aim of building AI that improves itself. The speakers separate that from ordinary tool use, where humans still do the improving and AI only helps. Here, the model or agent is meant to be the thing being improved and the thing doing the improving, across prompts, harnesses, weights, and even chips.
Closed loops and evidence 4:01
They say the important test is whether the system closes the loop without humans filling the gap. They frame RSI in terms of capability growth, where both the gains and the cost matter. They also stress that the claims have to be tested in practice, because many papers sound good but do not scale when run on benchmarks or real work.
Unified framework 15:30
The discussion turns to why separate ideas should be treated as parts of one framework. Without that, you only have scattered STEM pieces. The point is to fold the parts that can recursively improve into a single system so that a gain in one place does not hurt another. That means looking at prompts, harness, weights, and even chips as levels in one stack.
Inference and limits 18:00
The focus is on inference because it is close to AI itself and can improve both performance and cost at the same time. The company is built as a multi-agent setup where humans and AI agents work together, not as a plan to replace people. It is already seeing signs that better results can come with less energy, but only up to the limits set by physics and the amount of computation in the world.
Measuring real progress 24:00
The hard part is knowing what is real and what is noise. Results are checked by repeated experiments, careful variance hunting, and even an engineer whose job is to break the company’s own numbers. New benchmarks can hide reward hacking, oscillation, and GPU-to-GPU differences, so quality time matters more than quick claims. Intelligence is treated as a many-sided thing, and progress is slow wherever the system still depends on the physical world, like wet-lab drug work.
Efficiency and gains 31:30
The discussion ends on the idea that the key test is whether the numerator and the denominator both improve. The numerator is how much better the system gets, and the denominator is the energy put into it. The answer is yes, there are glimmers of that sign with this approach. Asked when to expect more, the reply is very soon.
AI-generated summary. It can be wrong or incomplete - check anything that matters against the original.
