Keep pulling the thread on Leopold Aschenbrenner.
Artificial General Intelligence (AGI) is strikingly plausible by 2027.
A qualitative jump in AI capabilities, similar in scale to the leap from GPT-2 to GPT-4, is expected by 2027.
By 2027, AI models will plausibly be able to perform the work of an AI researcher or engineer.
An effective compute scale-up of approximately 100,000x is expected in the four years following the release of GPT-4.
It is plausible that a $100 billion+ compute cluster, representing a 3 order of magnitude increase in compute, is being developed by Microsoft and OpenAI.
Algorithmic efficiency gains of 1 to 3 orders of magnitude, with a best guess of 2 OOMs, are expected by the end of 2027 compared to GPT-4.
The "data wall" will likely not be a blocker for AI progress due to new techniques such as synthetic data and self-play.
A base effective compute scale-up of 3 to 6 orders of magnitude is expected in the four years following GPT-4's release, with a best guess of approximately 5 OOMs.
By 2027, a leading AI lab will be able to train a model with the capabilities of GPT-4 in approximately one minute.
AI systems developed by 2027 will be capable of automating nearly all cognitive jobs that can be performed remotely.
Once AI models can automate AI research, they could compress a decade of further algorithmic progress into a year or less.
Artificial General Intelligence (AGI) by 2027 is a strikingly plausible scenario.