So, exactly how big will the intelligence explosion be?
…Ten years of AI progress in a year? In a month?
Our new paper tackles this question head-on.
I've researched AI takeoff speeds for many years. This is my best stab at an answer.🧵
…Ten years of AI progress in a year? In a month?
Our new paper tackles this question head-on.
I've researched AI takeoff speeds for many years. This is my best stab at an answer.🧵
An intelligence explosion is where AI makes smarter AI, which quickly makes even smarter AI, etc.
Our scenario: AI fully replaces humans at improving AI “software” (algorithms and data).
(We conservatively assume that the amount of compute remains constant.)
Our scenario: AI fully replaces humans at improving AI “software” (algorithms and data).
(We conservatively assume that the amount of compute remains constant.)
Our model has three main parameters:
1) Initial speed-up in software progress from AI automating AI research
2) After the initial speed-up, does progress accelerate or decelerate?
3) How far until AI software reaches fundamental limits on compute efficiency?
1) Initial speed-up in software progress from AI automating AI research
2) After the initial speed-up, does progress accelerate or decelerate?
3) How far until AI software reaches fundamental limits on compute efficiency?
We estimate these three parameters through a mix of empirical evidence and guesswork.
One interesting insight:
Compressing >10 years of total AI progress into <1 year via software improvements is tough.
It would require >10 OOMs of efficiency improvements. (Effective compute for developing AI has recently risen by 10X/year.)
That's a 10 billion-fold increase!
Compressing >10 years of total AI progress into <1 year via software improvements is tough.
It would require >10 OOMs of efficiency improvements. (Effective compute for developing AI has recently risen by 10X/year.)
That's a 10 billion-fold increase!
We put probability distributions over the params and run a Monte Carlo simulation.
The model spits out the probability of compressing multiple years of total AI progress into a few months from software improvements alone.
The model spits out the probability of compressing multiple years of total AI progress into a few months from software improvements alone.
I roughly estimate that our model assigns a ~20% probability to takeoff being faster than in AI-2027
@DKokotajlo @eli_lifland
@DKokotajlo @eli_lifland
https://x.com/DKokotajlo/status/1907826614186209524
How scary would this be?
6 years of progress might take us from 30,000 expert-level AIs thinking 30x human speed to 30 million superintelligent AIs thinking 120X human speed (h/t @Ryan)
If that happens in <1 year, that's scarily fast just when we should proceed cautiously
6 years of progress might take us from 30,000 expert-level AIs thinking 30x human speed to 30 million superintelligent AIs thinking 120X human speed (h/t @Ryan)
If that happens in <1 year, that's scarily fast just when we should proceed cautiously
It goes without saying: the model is very basic and has many big limitations.
E.g. we assume AI progress will follow smooth trends.
But if there’s a big paradigm shift, AI progress could be much more dramatic. Alternatively, the current paradigm could fizzle out.
E.g. we assume AI progress will follow smooth trends.
But if there’s a big paradigm shift, AI progress could be much more dramatic. Alternatively, the current paradigm could fizzle out.
See the full paper for more:
forethought.org/research/how-q…
forethought.org/research/how-q…
This builds on previous work by myself and @daniel_271828
https://x.com/TomDavidsonX/status/1904956748995182704
• • •
Missing some Tweet in this thread? You can try to
force a refresh
