Debating AGI: Untrained Capability vs Ability to Learn Anything
ARC Prize founder Greg Kamradt and danielmac8 debated what "general" in AGI actually means, ultimately converging on the phrasing "learning to do anything."
How It Unfolded
- Kamradt initially argued that "general" in AGI means the system can do things it was never trained for. He speculated that the weights of Google's astra model are also frozen, so capability gains can only come from external memory and tools the model builds for itself. What he cares about isn't whether the model can complete a task in one shot, but what it can achieve through self-built tools given enough time and inference.
- danielmac8 offered a counter-definition: if humans serve as the existence proof of general intelligence (HGI), then humans also fail at plenty of things without training—so "general" should mean a machine mind can be trained to do anything, not that it can do things without training.
- In his response, Kamradt agreed the crux is "learning," and revised his formulation to: AGI is a machine intelligence that can learn to do anything—deliberately dropping the word "training," since "training" implies someone has to train it, which isn't a capability intrinsic to the agent itself.
Why It Matters
- The discussion directly shapes how AGI progress gets measured: if the definition emphasizes "doing without training," current models fall far short; if it emphasizes "learning and completing tasks given enough time and tools," then self-built tools, external memory, and long-horizon reasoning (areas ARC Prize has long focused on) become key evaluation dimensions.
- The two sides ultimately converged on "being able to learn to do anything," stressing that learning should be an intrinsic, autonomous capability of the agent rather than something dependent on external human training—providing more precise wording for future AGI benchmarks and debates.
2026-09-09 ~ 2026-09-09 · 5 related posts
Primary sources
- Greg Kamradt: AGI's 'general' means doing what it wasn't trained for, via self-built tools — GregKamradt · 2026-09-09
- ARC's Greg Kamradt: AGI gains must come from self-built tools, not one-shot ability — GregKamradt · 2026-09-09
- [source] Rethinking AGI: 'general' should mean trainable to do anything, not zero-shot everything — daniel_mac8 · 2026-09-09
- [source] Greg Kamradt and Daniel Mac debate AGI: 'general' means able to learn anything, not be trained — GregKamradt · 2026-09-09
1 near-duplicate retellings: daniel_mac8