DeepMind New Article: AGI Is Within Reach, Four Paths to ASI, $1 Billion in Compute Becomes $1,000
DeepMind Institute published an article stating that AGI is already within reach and proposed four paths to superintelligence. Effective compute grows at least 10x every year; the cost of $1 billion in compute will drop to $1,000 in ten years. Limiting training compute is useless; the power grid is the bottleneck.
Today, DeepMind Institute published a new article discussing what comes after AGI. In the article, author Tim Genewein directly says that AGI is "coming within reach" and lays out four paths to ASI (Artificial General Super Intelligence):
- larger models and more data
- new paradigms
- millions of coordinated agents
- recursive self-improvement
The article gives two key numbers:
Effective compute grows at least 10x every year. Ten years from now, the number of agent copies that can run simultaneously will exceed the Earth's population.
The cost curve is even more shocking: work that would cost $1 billion in compute today will cost roughly $1,000 by then. Compute capacity that is exclusive to big tech will become something individual developers can also access.
The article also rebuts a regulatory approach: limiting a single input, such as training compute, is useless. Other paths will keep moving forward regardless, and the process from AGI to ASI will not stop because of it.
The article also mentions physical bottlenecks: power plants, power grids, chip factories. Some commenters joked in replies that their electricity bills have something to say. Someone else added that the path is clear, but someone needs to build the substations first.
Commenters disagree on which of the four paths will go the farthest. Some guess it is millions of coordinated agents, because multi-agent systems are already showing early forms, but there is currently a lack of good benchmarks to measure them. Some think new paradigms are the real variable, but the term is too vague and can encompass anything. Another commenter who works on trading systems said that in their own practical use, AI is better suited as a filter above rules rather than directly as a strategy, and larger models and more data are the only path they can personally experience.
Commenters also weighed in on the significance of this statement: DeepMind's roadmap is more convincing than OpenAI's, because Google has recently been releasing products at a furious pace and is on a hot streak. Others remain cool-headed, saying they should first confirm the article's definition of AGI, since definitions often differ across labs.
The article link is in Shane Legg's tweet, titled What Comes After AGI? Those interested can read the original directly. Original link.
AGI has not landed yet, but the roadmap has arrived first. This time gap itself is information.
发布时间: 2026-11-20 01:41