[Question] Is there a public tracker depicting at what dates AI has been able to automate x% of cognitive tasks (weighted by 2020 economic value)?

[Mitchell Laughlin🔸]

In the next few decades we may develop AI that can automate ~all cognitive tasks and dramatically transform the world. By contrast, today the capabilities and impact of AI are much more limited. Once we have AI that could readily automate 20% of cognitive tasks (weighted by 2020 economic value), how much longer until it can automate 100%? This is what I refer to as the question of AI takeoff speeds; this report develops a compute-centric framework for answering it. - What a Compute-Centric Framework Says About Takeoff Speeds, Tom Davidson of Open Philanthropy, June 27, 2023.

I recently listened to Tom Davidson’s 80,000 Hours episode on AI takeoff and was interested to know if anyone had any estimates of where we currently were (by his measure of per cent of cognitive tasks by 2020 economic value), where we’ve previously been at certain dates, and potentially where people thought we would be going forward.

There’s a good chance that this was in the report (that I haven’t properly read), so if it is feel free to point that out!

Thanks a bunch all

[Vasco Grilo🔸]

Hi Mitchel,

In Epoch’s implementation of Tom’s model, you can plot the “fraction of all cognitive tasks automated (goods and services and R&D)”. With the default parameters, it is currently negligible, and only reaches 1 % in around 2029:

[Mitchell Laughlin🔸]

This was helpful, thanks Vasco.

[Mo Putera]

In Tom’s report it’s an open question:

• To inform the size of the effective FLOP gap

  • ...

  • What is the current $ value-add of AI? How is it changes over time, or with model size?

    • Various ways of operationalising this: investment, revenues, effect on GDP.

    • Relevant for when AI will first be capable enough to readily add $trillions /​ year to GDP.

The closest the report gets to answering your question seems to be in the Evidence about the size of the effective FLOP gap subsection, where he says (I put footnotes in square brackets)

• As of today the largest training run is ~3e24 FLOP. [I believe these were the requirements for PaLM.] …

• In my opinion, today’s AI systems are not close to being able to readily perform 20% of all cognitive tasks done by human workers. [Actually automating these tasks would add ~$10tr/​year to GDP.]

• If today’s systems could readily add $500b/​year to the economy, that would correspond to automating ~1% of cognitive tasks. [World GDP is ~$100tr, about half of which is paid to human labour. If AI automates 1% of that work, that’s worth ~$500b/​year.]

That last assumption bullet is what seems to have gone into the https://​​takeoffspeeds.com/​​ model referenced in Vasco’s answer.

[Mitchell Laughlin🔸]

Super helpful addition, thanks Mo.