DeepMind essay on what follows AGI draws skepticism on HN, with the page itself unreadable
Summarized from the Hacker News discussion because the article couldn't be read
- Commenters quote the essay saying AGI is within reach and that AI could move toward artificial general superintelligence (ASI) at its current pace or faster.
- A commenter argues that if frontier labs were near AGI, they would use it to run their own businesses instead of selling model access.
- Commenters say AGI has no agreed definition, and that what earlier writers called AGI is now often called ASI.
- One commenter asks whether DeepMind will fire its engineers once AGI arrives, and a reply links a Sam Altman blog post titled Three Observations.
Hacker News opinions
If these labs were really close to AGI, they'd be using it to run their own businesses instead of selling API access. Why sell the shovel if the gold mine is right there?
Stateless LLMs are the blocker for me. Without real memory between sessions, AGI feels like a daydream, and a memories.md file doesn't fix that.
Maybe the human brain is also just sampling from a probability distribution of learned behavior. We don't really know what intelligence is, so calling LLM output fake feels early.
Scaling current approaches won't get to human-level intelligence, I think. They're way less efficient than brains and split training from inference, and the new architectures are still unproven.
I don't buy that intelligence is unbounded. Human-level might be enough for most physical tasks, and more of it mostly helps with coding and other intelligence-shaped stuff.
Has anyone asked DeepMind when they'll fire their engineers? If they still need people, is it really AGI? That seems like the obvious question for a frontier lab CEO.
Nobody agreed on what AGI means in the first place. What people called AGI in the singularity books sounds like what everyone now calls ASI.
We actually have a decent mathematical handle on the upper limits of general intelligence. Funny to use today's math and physics to bound something that would understand them better than we do.