Com­put­ing Overhang

TagLast edit: 21 Sep 2020 23:00 UTC by Swimmer963 (Miranda Dixon-Luinenburg)

Computing overhang refers to a situation where new algorithms can exploit existing computing power far more efficiently than before. This can happen if previously used algorithms have been suboptimal.

In the context of Artificial General Intelligence, this signifies a situation where it becomes possible to create AGIs that can be run using only a small fraction of the easily available hardware resources. This could lead to an intelligence explosion, or to a massive increase in the number of AGIs, as they could be easily copied to run on countless computers. This could make AGIs much more powerful than before, and present an existential risk.


In 2010, the President’s Council of Advisors on Science and Technology reported on benchmark production planning model having become faster by a factor of 43 million between 1988 and 2003. Of this improvement, only a factor of roughly 1,000 was due to better hardware, while a factor of 43,000 came from algorithmic improvements. This clearly reflects a situation where new programming methods were able to use available computing power more efficiently.

As of today, enormous amounts of computing power is currently available in the form of supercomputers or distributed computing. Large AI projects can grow to fill these resources by using deeper and deeper search trees, such as high-powered chess programs, or by performing large amounts of parallel operations on extensive databases, such as IBM’s Watson playing Jeopardy. While the extra depth and breadth are helpful, it is likely that a simple brute-force extension of techniques is not the optimal use of the available computing resources. This leaves the need for improvement on the side of algorithmic implementations, where most work is currently focused on.

Though estimates of whole brain emulation place that level of computing power at least a decade away, it is very unlikely that the algorithms used by the human brain are the most computationally efficient for producing AI. This happens mainly because our brains evolved during a natural selection process and thus weren’t deliberatly created with the goal of being modeled by AI.

As Yudkoswky puts it, human intelligence, created by this “blind” evolutionary process, has only recently developed the ability for planning and forward thinking—deliberation. On the other hand, the rest and almost all our cognitive tools were the result of ancestral selection pressures, forming the roots of almost all our behavior. As such, when considering the design of complex systems where the designer—us—collaborates with the system being constructed, we are faced with a new signature and a different way to achieve AGI that’s completely different than the process that gave birth to our brains.


See also

Ta­boo “com­pute over­hang”

Zach Stein-Perlman1 Mar 2023 19:15 UTC
17 points
8 comments1 min readLW link

Are we in an AI over­hang?

Andy Jones27 Jul 2020 12:48 UTC
259 points
106 comments4 min readLW link

How Much Com­pu­ta­tional Power Does It Take to Match the Hu­man Brain?

habryka12 Sep 2020 6:38 UTC
44 points
1 comment1 min readLW link

GPT-2005: A con­ver­sa­tion with ChatGPT (fea­tur­ing semi-func­tional Wolfram Alpha plu­gin!)

Lone Pine24 Mar 2023 14:03 UTC
19 points
0 comments22 min readLW link

Mea­sur­ing hard­ware overhang

hippke5 Aug 2020 19:59 UTC
112 points
14 comments4 min readLW link

Rele­vant pre-AGI possibilities

Daniel Kokotajlo20 Jun 2020 10:52 UTC
38 points
7 comments19 min readLW link

A closer look at chess scal­ings (into the past)

hippke15 Jul 2021 8:13 UTC
49 points
14 comments4 min readLW link

Brain-in­spired AGI and the “life­time an­chor”

Steven Byrnes29 Sep 2021 13:09 UTC
65 points
16 comments13 min readLW link

AI over­hangs de­pend on whether al­gorithms, com­pute and data are sub­sti­tutes or complements

NathanBarnard16 Dec 2022 2:23 UTC
2 points
0 comments3 min readLW link

Thoughts on hard­ware /​ com­pute re­quire­ments for AGI

Steven Byrnes24 Jan 2023 14:03 UTC
52 points
30 comments22 min readLW link

Sam Alt­man’s Chip Am­bi­tions Un­der­cut OpenAI’s Safety Strategy

garrison10 Feb 2024 19:52 UTC
197 points
52 comments1 min readLW link

The 0.2 OOMs/​year target

Cleo Nardo30 Mar 2023 18:15 UTC
82 points
24 comments5 min readLW link

In­fer­ence cost limits the im­pact of ever larger models

SoerenMind23 Oct 2021 10:51 UTC
42 points
27 comments2 min readLW link

Are There Ex­am­ples of Over­hang for Other Tech­nolo­gies?

Jeffrey Heninger13 Dec 2023 21:48 UTC
59 points
50 comments11 min readLW link

How should Deep­Mind’s Chin­chilla re­vise our AI fore­casts?

Cleo Nardo15 Sep 2022 17:54 UTC
35 points
12 comments13 min readLW link
No comments.