1 of 19

What does a compute centric perspective say about AI takeoff speeds?

2 of 19

What is takeoff speed?

Capabilities takeoff speed: How quickly will AI improve as we approach and surpass human-level AI?

Impact takeoff speed: How quickly will AI’s impact on the world increase as we approach and surpass human-level AI?

3 of 19

A “compute centric” framework

Capability of AI determined by the “effective compute” used to train it

Effective compute = compute * quality of AI algorithms

The more capable an AI, the more tasks it can automate

Better AI → automates more tasks → more economic impact → more investment → better AI
Better AI → automates more tasks → faster hardware progress → cheaper compute → better AI
Better AI → automates more tasks → faster algorithmic progress → better algorithms → better AI

Measure takeoff as:

Time from [AI that can automate 20% of economic tasks] to [AI that can automate 100% of tasks]
Tasks are weighted by their importance to the economy in 2022 (by their “share” of output)

Takeoff speed with this definition depends on:

Difficulty gap: How much more effective compute to train [AI that can automate 100% of tasks] than to train [AI that can automate 20% of tasks]?
Speed crossing difficulty gap: How quickly will the effective compute used in training increase?

4 of 19

Theoretical framework

Difficulty gap

Speed crossing difficulty gap

AI can automate 20% of tasks

Effective compute used in training

More compute, better algorithms

AI can automate 100% of tasks

5 of 19

Theoretical framework

Difficulty gap

Speed crossing difficulty gap

AI can automate 20% of tasks

Effective compute used in training

More compute, better algorithms

AI can automate 100% of tasks

6 of 19

What is the meaning of the difficulty gap?

Hans Moravec's "rising tide of AI capacity" can illuminate the meaning of the difficulty gap

How much more effective compute to train [AI that can automate 100% of tasks] than to train [AI that can automate 20% of tasks]?

7 of 19

What is the meaning of the difficulty gap?

How wide is the hump?

Effective compute used to train AI

AI can automate 100% of tasks

AI can automate 20% of tasks

8 of 19

How big is the difficulty gap?

Arguments for a small difficulty gap (~1 - 2 OOMs):

Large effect of brain size on intelligence in humans and animals
It’s hard in practice to automate just 20% of a job
We’re close to training AGI!

Arguments for a big difficulty gap (~4 - 8 OOMs):

Huge variety of tasks in the economy
AI seems much better suited to some tasks than others
Some types of AI are much more expensive to train than others
We’re far from training AGI!

My median: ~3-4 OOMs.

9 of 19

Theoretical framework: fitting the framework

Difficulty gap

Speed crossing difficulty gap

AI can automate 20% of tasks

Effective compute used in training

More compute, better algorithms

AI can automate 100% of tasks

10 of 19

How fast will we cross the difficulty gap?

In the last decade (h/t Epoch):

Compute in the largest training runs increased by ~0.6 OOM/year

~0.1 OOM/year from compute getting cheaper
~0.5 OOM/year from more spending on compute

Algorithms reduced compute requirements by ~0.4 OOM/year

Progress towards AGI might be different

→ Effective compute on largest training run increased by ~1 OOM/year

Speed crossing difficulty gap

More spending, cheaper compute, better algorithms

11 of 19

How fast will we cross the difficulty gap?

What might change while we cross the difficulty gap?

Spending could increase faster or slower

Faster if certain actors “wake up” to the economic+strategic implications of TAI
Slower once we hit limits on budgets or limits on the # AI chips in the world

AI will speed up AI R&D progress

Faster reduction the price of compute
Faster algorithmic progress
~2-3X as fast is plausible

	Speed crossing the gap (OOM/year)
	More spending	Cheaper compute	Better algorithms	Total
Last decade	~0.5	~0.1	~0.4	~1
Crossing the difficulty gap	~0.3	~0.2	~0.8	~1.3

Speed crossing difficulty gap

More spending, cheaper compute, better algorithms

12 of 19

Very tentative conclusions

	Difficulty gap How much more effective compute to train [AI that can automate 100% of tasks] than to train [AI that can automate 20% of tasks]?	Speed crossing difficulty gap How quickly will the effective compute used in training increase?	Takeoff speed Time from [AI that can automate 20% of tasks] to [AI that can automate 100% of tasks]
Median	~3.5 OOMs	~1.3 OOMs/year	~3 years
Aggressive	~1 OOM	~2 OOMs/year	~0.5 years
Conservative	~8 OOMs	~0.5 OOMs/year	~16 years

13 of 19

14 of 19

So what?

This is pretty fast, and no discontinuities are baked into the model.

A faster takeoff means:

Less time to study “AIs similar to those that will pose x-risk”
Fewer warning shots for AI capabilities → fewer actors involved (labs, govs)
Fewer warning shots for AI risks → less consensus, worse coordination
Easier for power imbalances to emerge
Less chance for AI to transform the world before it poses x-risk → “nearcasts” look better

15 of 19

“20% is an arbitrary startpoint”

16 of 19

“20% is an arbitrary startpoint”

	Doubling times (median) X years before AI can readily automate 100% of tasks
Quantity	1 year	2 years	5 years	10 years
FLOP/$	0.3	1.2	2.7	3.1
Software	0.2	1.2	2.4	2.9
GWP	0.7	2.4	8.4	21.1

17 of 19

Time from 100% automation to superintelligence?

Reasons for “very fast”:

Huge amounts of AI labour working on AI R&D.
Naive analysis suggests a “software-only singularity” possible
Physical compute growing very quickly

Reasons against:

Software progress may be bottlenecked by compute for experiments
Hardware progress may be bottlenecked by limits to Moore’s law and time to build new fabs

My guess: more likely than not this happens in <12 months.

18 of 19

Limitations

Ignores data/environment
Assumes no lag between developing and deploying AI
Assumes AI capabilities improve continuously
Assumption: AI capability = training compute * quality of algorithms

19 of 19

Questions