Broad AI Timelines: Planning Under Uncertainty When will AI transform the world? Toby Ord argues we should accept uncertainty and adopt “broad AI timelines” to guide better decisionmaking and planning.

For more thoughts on broad timelines, take a look at the full essay:
forethought.org/research/bro...
22/22

But I want to stress that none of this implies we can slack off.
We’re in a race against AI timelines. It is just that we don’t know if that race is a sprint or a marathon. In either case, time is of the essence.
21/

Overall, we need a balanced portfolio of work aimed at short timelines and work aimed at long ones.
20/

These things are often highly leveraged, so are still worth doing even if there is a moderate chance that they pay off too late to help (an 80% chance of 5x is still great EV). As well as hedging against short timelines we need to make the most of long ones.
19/

5. There are many long-term plans of action that can pay off handsomely. Things like founding a research field or a company, building a movement, training promising students, and foundational research.
18/

China may have caught up with the West on chip production; Taiwan may have been invaded. There might be double-digit unemployment and an angry public.
So plans that are outside the current Overton window may be inside the Overton window that matters.
17/

4. We shouldn't over-index on current circumstances, as the world may look very different when transformative AI arrives. e.g. if it comes in 2035, the technologies and leading companies may look very different, as may the political and geopolitical landscapes.
16/

3. We should hedge against transformative AI coming very early, when we'd be least prepared. Such timelines are more important than their raw probability suggests.
15/

2. We can't just wait until the experts agree. That won't happen until it is too late. At that point, the only options are knee-jerk policy responses. So waiting for agreement is tantamount to committing to only the bluntest and least effective options.
14/

1. This isn't an excuse to go with your favourite timeline, just because experts find it credible. Decision making under uncertainty requires taking all credible timelines seriously (including the very short and very long).
13/

So that's the broad timelines claim: we should be thinking probabilistically about when transformative AI will arrive and the range of credible dates is very broad.
But what follows from this? How do we act under this severe uncertainty?
12/

For my forecast, both of these are true simultaneously: the distribution has an early peak, but also a long tail.
11/

Those concerned (or excited) about AI point to the substantial chance that it comes within just a few years — before anyone’s ready.
While others say it is more likely transformative AI arrives a decade or more from now.
10/

These distributions are so hard to summarise that I think part of the timelines debate stems from people describing different parts of the same elephant…
9/

Here's my own current forecast for when transformative AI will arrive (AI powerful enough to take over were it misaligned, and to double the global rate of scientific and tech progress). It is similarly shaped but even broader.
8/

And I think that is actually a pretty good summary of Kokotajlo's uncertainty and of the uncertainty of experts in general. "No idea" is going a bit too far, but it is a better summary than "4 years" (or any other number).
7/

A smart policy maker who understood this forecast might summarise it as:
"Oh, so you are saying you have no idea when it will happen — it could be next year, or it could be 6 presidential terms from now. And you’re saying there is a 1 in 5 chance it isn’t even in that range."
6/

It is very hard to summarise that forecast with a single year. If he were to simply say '4 years', that would suggest to most people that he thought it was about 3 to 5 years.
Any point estimate fails to do it justice.
5/

e.g. here is Daniel Kokotajlo's current forecast for AI that performs better at every topic than human experts in that topic.
His median is 4 year's time, but his 80% confidence interval ranges from 1 year to 25 years.
4/

Here are three expert AI forecasts (from 2023) addressing “In what year would AI systems be able to replace 99% of current fully remote jobs?”
This is very broad uncertainty between experts and a closer look shows that even each individual expert has very wide uncertainty.
3/

Here is the range of credible dates for AGI, across all forecasters at Metaculus.
This is a huge range of uncertainty. The median date is 2033, but their 80% confidence interval is from 2026 to 2067 — between 0.25 years and 41 years.
2/

B R O A D T I M E L I N E S
We should have neither short AI timelines, nor long timelines, but a broad probability distribution over when transformative AI will arrive.
My new essay explains why & explores the implications of such deep uncertainty.
🧵 1/

The median age in China has rapidly caught up with the United Kingdom. Line chart of median age for China and the United Kingdom from 1950 to 2025, with the vertical axis in years from 0 to 40 and the horizontal axis showing years 1950 to 2025. A line labeled United Kingdom stays around mid-30s in 1950, dips slightly to about 33 by the mid-1970s, then gradually rises to about 40 by 2025. A line labeled China starts around 22 in 1950, falls to about 18 to 19 in the mid-1960s and 1970s, then climbs steadily to meet the UK at about 40 in 2025. Annotated note: in the mid-1960s China’s median age was just under half that of the UK; another note states that today the median age in both countries is 40 years. Data source: UN, World Population Prospects (2024). License: CC BY.

The median age in China has rapidly caught up with the United Kingdom—

In 1965, the median age in the United Kingdom was almost twice that of China. Half of the people in the UK were younger than 34 years, and half were older. In China, this midpoint was just 18 years.

Birth Lottery If you were reborn today, where would you land? And how would that change your life?

Being born is a roll of the dice.

Most of us got insanely lucky.

Imagine you had to roll again, how would you want the world to look?

www.givingwhatwecan.org/birth-lottery

GRB 250314A Pull-out (NIRCam image) GRB 250314A Pull-out (NIRCam image)

And to find out more about the dark red dot itself (whose explosion was designated GRB 250314A), see ESA's report here: www.esa.int/ESA_Multimed...

You can find out much more about the affectable universe and the ultimate limits of causality in our universe in a paper I wrote called 'The Edges of Our Universe':
17/17
arxiv.org/pdf/2104.01191

Here's a diagram I made, showing our place in the affectable universe, and its place within even greater regions:
16/

So if beings from Earth and that distant galaxy set off towards each other at close to the speed of light, they could yet meet.
15/

Though interestingly, current events in that star's galaxy could still affect *us*. Our affectable universe doesn’t include it and its affectable universe doesn’t include the Earth. But these spheres do overlap...
14/

So this image shows a star which is so far away that it is outside the affectable universe. Nothing we do here and now could ever affect it, and nothing that happens there now could ever affect the Earth.
13/