Korva is a scroll of the internet’s best ideas.
AGI may be the machinery
How close it feels
Even some real skeptics have called GPT-5.6, which has been out for about two weeks, very AGI-like. It is very hard to name something wanted from the model that it can't do.
What it can't do
Clear gaps remain. It can't yet be told to cure cancer and then cure it, or to do a complicated physical task with a robot. Although brilliant, it is still not learning continuously as it goes, which may not be a hard requirement for AGI but is certainly something worth wanting.
AGI as machinery
The counterargument is that AGI is not about any single model. It is the machinery that makes the models, and from model to model new things are being learned and new science figured out. That is working amazingly well, so people who say the genie is already here have a case.
Real AGI, by the standard that matters, is very close, and not much longer away.
Why can't models learn continuously today?
A trained model's weights are fixed once training ends, and what it picks up in a conversation lives only in its context window. Updating the weights on new experience risks catastrophic forgetting, a problem described by Michael McCloskey and Neal Cohen in 1989, in which a network trained on new material overwrites what it learned before. Humans avoid this in part by consolidating memories gradually during sleep. Research on continual learning tries to give networks the same stability.
Why does the definition of AGI keep shifting?
Alan Turing proposed in 1950 that a machine counts as intelligent if it can hold a conversation indistinguishable from a person's, a test he called the imitation game. Language models now converse fluently, yet few people treat that as settling the question. Researchers call this the AI effect, in which a capability stops seeming like real intelligence once machines can do it. Computer scientist Larry Tesler is often credited with the quip that intelligence is whatever machines have not done yet.
Why are physical tasks harder for AI?
In the 1980s, roboticist Hans Moravec observed that reasoning tasks such as chess are comparatively easy for computers while perception and movement are hard. Humans find the reverse, because sensing and handling objects rely on millions of years of evolved machinery that people use without noticing. Text for training language models is abundant online, while data of robots acting in physical settings is far scarcer.
More from Invest Like the Best with Patrick O'Shaughnessy
