Grokking “Semi-informative priors over AI timelines”

1JBlack

2gwern

1JBlack

2gwern

1JBlack

1anson.ho

1JBlack

New Comment

7 comments, sorted by Click to highlight new comments since: Today at 3:58 PM

Sorry, but I don't think AGI development as a sequence of Bernoulli trials with constant probability is anything like a *remotely sane* model, and any conclusions drawn from it are worthless.

For one, it implies that after any "trial" after which AGI is not developed, we are in exactly the same state for future development of AGI as after any other "trial". Even excluding all the ridiculous post-hoc dickering over what constitutes a "trial", this is obvious nonsense.

Suppose that you were told by a trusted oracle that this model was absolutely correct and p=0.5. What would this *actually mean* for your understanding of reality?

For one, it implies that after any "trial" after which AGI is not developed, we are in exactly the same state for future development of AGI as after any other "trial". Even excluding all the ridiculous post-hoc dickering over what constitutes a "trial", this is obvious nonsense.

That obvious nonsense is a common feature of many models of technological development or combinatorial innovation and also scientific publishing (equal-odds rule): each paper, patent, or experiment is a lottery ticket with a similar probability of success.

"Success" in that context means something very different from "success" here: it refers to *any impact whatsoever* beyond the mere fact of being published, patented, or developed. For example, citations in future papers, an invention producing a profit, and other such measures.

I have no problem at all in supposing that (for example) any given AI researcher will have *p* chance this year of publishing a paper that is cited in some threshold of papers in the future, that the same chance *p* will apply to a randomly selected scientist's paper next year of doing the same thing, and also applied to a researcher 30 years ago. That is not what this post's model is saying.

This post refers to serial attempts to produce *one specific thing* that we preselected. The independent Bernoulli trial model *does not* work for that, except over the very shortest timescales, far shorter than the span of all attempts to make a thing.

For every *specific* invention, in hindsight it is clear that the earliest attempts had essentially zero chance of complete success, subsequent attempts had not much more, until eventually a relatively sharp threshold was reached due to meeting increasingly better understood prerequisites, and it became nearly inevitable.

This is not at all the same as a model that has constant *p* per trial right from the start. The only common factor is that if you *retrospectively* randomly sample a time point T between the start and the eventual success, you get E[T - start date] = E[success date - T]. However this is completely uninformative from a prediction point of view: it is true *no matter where you set the start date*.

This post refers to serial attempts to produce one specific thing that we preselected.

'AI' is a problem and a goal, not a specific thing. It's what we label the thing which turns out to do what we want it to. "AI" is whatever works. We can no more 'preselect' the specific thing which will be the invention of 'AI' than Douglas Lenat could preselect Cyc as 'AI'. (Because that didn't work.)

For every specific invention, in hindsight it is clear that the earliest attempts had essentially zero chance of complete success, subsequent attempts had not much more, until eventually a relatively sharp threshold was reached due to meeting increasingly better understood prerequisites, and it became nearly inevitable. This is not at all the same as a model that has constant p per trial right from the start.

No it's not. Innovation happens all the time by brute force, accident, trial-and-error, and serendipity. Literally the first example Clancy gives from Weitzman, of Edison testing lightbulb filaments, disproves that (an example one should remember from middle school history class). There was no 'sharp threshold' or 'increasingly better understood prerequisites'. If Edison had tried the carbonized cotton thread first after his proof-of-concept using a metal filament, it would have worked; if he hadn't eventually tried cotton after literally reaching in to grab thousands of other candidate plant materials, it wouldn't've, and he would've gone on grabbing thousands of more materials with no greater chance of success in each trial until eventually he gave up or stumbled across another rare material which would work (bamboo apparently would've worked).

Attempts to produce a light bulb are exactly an example of what I mean.

Inventors in Ancient Greece would not have produced an electric light bulb, at all, even if thousands of them worked for hundreds of years. Early 18th century inventors could possibly have produced one if they had reason to put in immense effort. So rather than starting at Edison, *this* is where we should start that story. This is about the time period where *p* > 0 for a "trial" for the first time.

By the early 19th century the conditions were becoming ripe for the possibility of electric lighting, and some expensive and short-lived forms of electric light bulb were already being invented (Davy's incandescent bulb and later arc lights). The techniques and materials still weren't reliable or cheap enough at that time, and the theory and practice of electricity was still lacking, but people were attempting the task of making a long-lived, cheap and moderately efficient bulb and making progress with increasing degrees of success.

By the later 19th century there were dozens of successful incandescent light bulb designs, with the first patent granted about 40 years before Edison. Later versions were developed that had some narrow commercial use, and Edison's bulb was an economically better incremental improvement that was sufficiently cheaper and longer-lived to go into mass production nearly 100 years after the first research prototype for incandescent lighting was developed.

I stated that the independent Bernoulli trial model does not work, *except over the very shortest timescales*, far shorter than the span of all attempts to make a thing. Edison's attempts were indeed far shorter than the span of all attempts to make the thing.

The original post is talking about the span of all attempts to make the thing, and the constant-p model is not at all reasonable for that.

To make sure I'm understanding you correctly, do you think the largest problem comes from (1) thinking of AGI development as a sequence of Bernoulli trials, or (2) each Bernoulli trial having constant probability, or (3) both?

It's not obvious to me that (1) is hugely problematic - isn't Laplace's rule of succession commonly applied to forecasting previously unseen events? Are you perhaps arguing that there's something particular to AGI development such that thinking of it as a series of Bernoulli trials is completely invalid?

I'm more sympathetic to your criticism of (2), but I'll note that Davidson actually relaxes this assumption in his model extensions, and further argues (in appendix 12) that the effect of the assumption is actually pretty small - most of the load of the model is carried by the first-trial probability and the reference classes used to generate it.

(1) seems reasonable as a model at this level of abstraction, absent quibbles about whether some outcome *really is* AGI or not, instead of *some degree* of AGI-ness.

(2) seems utterly wrong, and I don't think it even makes sense to talk about a "first trial" as being a clear-cut thing, let alone having a sensible probability, and definitely not as something related to success of all future trials. I contend that it is not even "semi-informative", it is useless.

Notes:I give visual explanations for Tom Davidson’s report,Semi-informative priors over AI timelines, and summarise the key assumptions and intuitionsThe diagrams can be foundhere– you can click on the boxes to get linked to the part of the report that you’re interested in^{[1]}Thanks to theEpochteam for feedback and support! Thanks especially to Jaime Sevilla and Tom Davidson for providing detailed feedback.## Executive Summary

The framework in

assumes a model ofSemi-informative priors over AI timelinesAGIdevelopment which consists of a sequence ofBernoulli trials, i.e. it treats each calendar year as a “trial” at building AGI with constant probability p of succeeding.However, we don’t know what this value of p is, so we use a generalisation of Laplace’s

rule of successionto estimate P(AGI next year | no AGI yet). This is done by specifying afirst-trial probability, the probability of successfully building AGI in the first year of AI research, together with thenumber of virtual successes, which tells us how quickly we should update our estimate for P(AGI next year | no AGI yet) based on evidence. The framework leans very heavily on the first-trial probability, which is determined using a subjective selection of reference classes (more here).How much evidence we get depends on the number of trials that we see, which depends on theyou can think of this as the time before which failure to develop AGI doesn’t tell us anything useful about the probability of success in later trials. For instance, we might think that 1956 (the year of the Dartmouth Conference) was the first year where people seriously started trying to build AGI, so the absence of AGI before 1956 isn’t very informative. If we think of each trial as a calendar year, then there have been 2021-1956 = 65 trials since the regime start-time, and we still haven’t developed AGI, so that’s 65 failed trials which we use to update P(AGI next year | no AGI yet), where “next year” now corresponds to 2022 rather than 1957.

regime start-time–But why should a trial correspond to a calendar year? The answer is that it doesn’t have to! In total, Davidson considers three candidate

trial definitions:Calendar-year trials:1 trial = 1 calendar yearCompute trials:1 trial = a 1% increase in the largest amount of compute used to develop an AI system to dateResearcher-year trials:1 trial = a 1% increase in the total researcher-years so far

P(AGI by 2036 | no AGI yet)=1−P(no AGI by 2036 | no AGI yet)=1−P(no AGI in 2022 | no AGI by 2021)...P(no AGI in 2036 | no AGI by 2035)If we extend this reasoning, then we can predict the probability that AGI is built X years into the future. Davidson does this to predict P(AGI by 2036 | no AGI yet) as follows:

The idea is that this framework only incorporates a small amount of information based on observational evidence, giving “

semi-informative priors”over AI timelines. This framework is shown in more detail below:Since Davidson uses three different trial definitions, we actually get three of these diagrams!

All in all, Davidson uses this to get a central estimate of P(AGI by 2036 | no AGI yet)=8%, with the following cumulative probability function:

## Motivation

One way of forecasting

AI Timelinesis to consider the inner workings of AI, guess what kinds of developments are the most important, and then generate a probability distribution over whenwill be developed. This is the approach taken by Ajeya Cotra inArtificial General Intelligence (AGI), a really detailed draft report that draws analogy to the human brain to forecast whenForecasting TAI with biological anchorswill first be developed.Transformative AI (TAI)^{[2]}Tom Davidson’s report,

, is also a detailed report forecasting AI timelines, but it takes a different approach to Cotra’s report. Rather than thinking about the details of AI development, it assumes we knowSemi-informative priors over AI timelinesalmost nothingabout it^{[3]}!The goal of this post is to explain the model through the liberal use of diagrams, so that you can get high-level intuitions about how it works, hopefully informing your research or understanding of AI forecasting.

## Laplace’s Rule of Succession

Suppose we’re trying to determine when AGI will first be developed, without knowing anything about the world except that there have been N years so far, and AGI has not been developed in any of these years. How would you determine the probability that AGI is developed in the next year

^{[4]}?A naive approach we might take is to think of each year as a

P(AGI next year | no AGI yet)=successessuccesses + failures=successestotal trials“trial” with two possible outcomes– (1)successful trials, where AGI is successfully built in the year of interest, and (2)failed trials, where AGI is not built in the year of interest. We then assume that the probability of building AGI in the next year is given by the total successful trials divided by the total trials:Since AGI hasn't been built in any of the last N years, there have been zero successes out of N trials. We thus conclude that the probability of AGI in the next year is zero… but clearly there’s something wrong with this!

The problem is that this approach doesn’t even account for the possibility that AGI might ever be developed, and simply counting the number of successes isn’t going to be very helpful for a technology that hasn’t been invented yet. How can we modify this approach so that both the possibility of success and failure are considered?

One clever way of doing this is to consider “virtual trials”. If you know that it’s possible for each trial to be either a success or a failure, then it’s as if you had previously observed one “virtual success” and one “virtual failure”, which we can add to the total observed successes and failures respectively. We can then modify the equation to:

P(AGI next year | no AGI yet)=successes + 1(successes + 1) + (failures + 1)=successes + 1total trials + 2This equation is called

Laplace's rule of succession, which is one approach to estimating the probabilities of events that have never been observed in the past. In particular, it assumes that we knownothingabout the world except for the number of trials and the number of successes or failures.If we apply this method, then we find that the probability of building AGI in the next year is 1/(N+2). Assuming that the field of AI was formed in

1956 at the famous Dartmouth Conference, then this suggests that N=2021−1956=65 and P(AGI is built in 2022)=1/67, or a probability of around 1.5%.If we extend this reasoning, then we can predict the probability that AGI is built X years into the future. Davidson does this to predict P(AGI by 2036 | no AGI yet) as follows:

P(AGI by 2036 | no AGI yet)=1−P(no AGI by 2036 | no AGI yet)=1−P(no AGI in 2022 | no AGI by 2021)...P(no AGI in 2036 | no AGI by 2035)This seems a lot more reasonable than the naive approach, but there’s still some serious problems with it, like the following:

It’s extremely aggressive before considering evidence: For instance, according to Laplace’s rule the attendants of the1956 Dartmouth Conferenceshould have predicted a 50% probability of developing AGI in the first year of AI research, and 91% probability within the first ten years!It’s sensitive to the definition of a “trial”: If we had chosen each trial to be “one day” instead of a year, our conclusions would be drastically different.What’s going on here (among other things) is that the rule of succession makes very few prior assumptions – i.e. it’s an

uninformative prior. In fact, it’s so uninformative that it doesn’t even capture the intuition that building a transformative technology in the first year of R&D is not commonplace! Clearly, we still need something better if we’re going to make predictions about AGI timelines.## Making the priors less uninformative

The solution that Davidson proposes is to make this prior less uninformative, by incorporating certain pieces of common sense intuition and evidence about AI R&D. Looking more closely at the framework given by Laplace’s rule of succession, we see that it depends on several factors:

Regime start-time:You can think of this as the time before which failure to develop AGI doesn’t tell us anything useful about the probability of success in later trials. We’ve been assuming this to be 1956, but this doesn’t have to be the case!First-trial probability: The odds of success on the first “trial” from the regime start-time onwardsTrial definition: Why are we using “one year” as a single trial, and what are some alternatives?We can also add an additional modification, in the form of the

number of virtual successes. This affects how quickly you update away from the first-trial probability given new evidence – the more virtual successes, the smaller your uncertainty about how difficult it is to build AGI, and thus the less you update based on observing more failed trials. For example, suppose that your initial P(AGI next year | no AGI yet) is 1/100:So far, we’ve been thinking about predicting whether or not AGI will be developed in the next year, but what we’re really interested in is

when it will be developed, if at all. Davidson tries to answer this by assuming a simple model of development, consisting of a sequence of trials, where each trial has a constant probability p of succeeding.^{[5]}Note that this probability is not the same as P(AGI next year | no AGI yet) - the latter corresponds to ourbeliefabout the value of p; it isn't the same as p itself.When the four inputs to the distribution P(AGI in year X | no AGI yet) are determined using common sense and some relevant reference classes, Davidson calls this distribution a “

semi-informative prior” over AGI timelines. Rather than considering tons of gnarly factors that could in principle influence progress towards AGI, we only look at a few select inputs that seem most relevant.The diagram above shows how the framework is pieced together. The first trial probability and number of virtual successes are used to generate an initial distribution for the probability of AGI in the next year. We then update this distribution with 2020 evidence based on the trials we’ve observed, depending on our specified regime start-time. This gives us the 2020 distribution for P(AGI next year (i.e. 2021) | no AGI yet). We combine this with the number of trials between 2020 and the year X that we're interested in, to get the final distribution over P(AGI by year X | no AGI in 2020). Note that this actually also depends on the trial definition – we’ll discuss how this fits into the diagram later.

## Semi-informative priors demystified

Now that we have the basic framework established, we just need to figure out what values we should assign to the input variables (i.e. first-trial probability, number of virtual successes, regime start-time, and trial definition). Davidson considers the first-trial probability to be the most significant out of these four input factors (via a

sensitivity analysis), although all are based on fairly subjective judgements.Let’s take a look at each of these in turn.

## First-trial probability

The

first-trial probabilityasks, “what is the probability of successfully building AGI on the first ‘trial’?”. This is very hard to determine just on the surface, and so Davidson turns to several historical examples from a fewreference classes. In particular, he looks at:transformative impact on societyDavidson uses these reference classes to derive constraints on the first-trial probability – this can be done by obtaining a base rate of successful trials from the past examples. Most of these don’t succeed in the first trial

^{[6]}, so one approach he uses is to look at how many successes there are after X trials, then works backwards using Laplace’s rule. He ultimately settles on abest guess first-trial probability of 4%.It’s worth noting that these reference classes and upward adjustments from the other trial definitions are the most important part of the framework, and the choice of these reference classes makes a really big difference to the final conclusions.

## Number of virtual successes

The

number of virtual successeschanges how quickly we should update based on our observation of failed trials.^{[7]}We want the size of this update to be reasonable, so we don’t want this number to be too large or too small. Davidson ultimately settles on1 virtual successfor most of the report, based on a combination of pragmatism, the plausibility of the prior^{[8]}, and the plausibility about the update size given new evidence.^{[9]}Different choices of the number of virtual successes matter less when the first-trial probability is lower, because making a big update (in proportion) from the prior distribution matters less in an absolute sense when the initial priors are already small.

## Regime start time

The

regime start-timeis the time for which “the failure to develop AGI before that time tells us very little about the probability of success after that time”, and affects the number of failed trials that we observe. While we previously considered the Dartmouth Conference in 1956 as the natural start of AI research, other alternatives (e.g. 1945, when the first digital computer was built) also seem reasonable.A problem with assuming a constant probability p of AGI being developed in any year becomes especially salient if we consider

veryearly start-times. Suppose we argue that people have beentrying to automate parts of their work since ancient times, and choose a start-time correspondingly. Then the framework would suggest the odds of building AGI in any year in ancient times is the same as that today!Davidson addresses this problem by down-weighting

the number of trialsoccurring in ancient times relative to modern times, by multiplying (with normalisation!) each year by the global population or the economic growth in that year.^{[10]}Overall, he places the most emphasis on a start-time of 1956, but does a sensitivity analysis with several alternatives, which do not significantly change the conclusions when appropriate down-weighting is applied.## Trial definition

The final input to the framework is the

trial definition, which specifies what exactly constitutes a single “trial” at building AGI. The initial approach we considered was in terms of calendar years, but there are reasonable alternatives, for example:Compute trials:Trials based on compute, e.g. 1 trial = “a 1% increase in the largest amount of compute used to develop an AI system to date”. These trials implicitly assume that increases in training compute are a key driver of AI progress^{[11]}Researcher-year trials:Trials that are defined in terms of the number of researcher-years performed so far, e.g. 1 trial = “a 1% increase in the total researcher-years so far”. We’re in effect assuming that each 1% increase in the “level of AI technological development” has a constant probability of developing AGI.^{[12]}Davidson considers both of these possible trial definitions, together with the calendar-year definition, finding that the resulting probabilities can vary a little depending on the chosen trial definition. In effect, we now have three separate frameworks based on the trial definition:

If we change the trial definition, then presumably we’ll also change the first-trial probability, so how do we calculate this? One approach that Davidson takes is to compute the first-trial probability for compute-years and researcher-years from the first-trial probability for calendar years – I’ll not go into this here, but I suggest looking at

thesesectionsof the report to find out more.Assuming 1 virtual success and a regime start-time of 1956, here’s what we get:

P(AGI by 2036)

Trial definition

Low-end

Central estimate

High-end

Calendar-year

1.5%

4%

9%

Researcher-year

2%

8%

15%

Compute trial

2%

15%

25%

Importantly, we can choose our first-trial probability such that our predictions remain the same for trivial changes in the trial definition, helping solve one of the aforementioned problems with applying Laplace’s rule of succession.

^{[13]}Overall, Davidson assigns⅓ weight to each of the three trial definitionsconsidered.## Putting things together: Final distribution

## Model Extensions

The framework also considers three extensions to the stuff outlined above:

Conjunctive model of AGI: It considers treating AGI development as the conjunction of multipleindependenttasksHyperpriorsover update rules: Updating a prior over what weight to assign to different update rules, which are themselves determined by the four inputs^{[14]}Allow some probability that AGI is impossibleFor the most part, these extensions don’t have a particularly large effect on the final numbers and conclusions.

## Final Distribution

If we combine everything from above then we end up with the following distribution and predicted numbers

^{[15]}:P(AGI by 2030)

P(AGI by 2050)

P(AGI by 2100)

~6%

~11%

~20%

10%

50%

90%

~2044

>2100

>2100

Davidson highlights three main strengths of his framework:

It quantifies the size of the update toP(AGI next year | no AGI yet)based on observed failuresIt highlights the significance of intuitive parameters,e.g. the first-trial probability, regime start-time, and the trial definitionIt’s arguably appropriate for expressing deep uncertainty about AGI timelines,e.g. by avoiding claims about “what fraction of the research we’ve completed towards AGI”He also points out some main weaknesses of the framework:

It incorporates limited kinds of evidence which could be really informative, e.g. how close we are to AGIIts near term predictions are too high,because current AI systems are not nearly as capable as AGI, and the framework doesn’t account for this evidence^{[16]}It’s insensitive to small changes in the definition of AGIIt assumes a constant chance of success in each trial(although the conjunctive model of AGI proposed in the extension relaxes this assumption)There are also some situations where it doesn’t make sense to use this framework – for instance, when we know what “fraction of progress” we’ve made towards achieving a particular goal. This can be hard to quantify for AGI development, but it’s actually closely related to an approach that the

Median group has previously attempted.## Conclusion

I think this model suggests that developing AGI within this century is

at leastplausible – we shouldn’t dismiss the possibility of developing AGI in the near term, and that the failure to develop AGI to date is not strong evidence for low P(AGI by 2036).I personally found the approach taken in this report really interesting, particularly in terms of the solutions Davidson proposes to the problems posed by the rule of succession. This seems possibly very valuable for other work on forecasting. I encourage you to look at the report’s

blog post^{[17]}, and to trymaking your own predictions using the framework.You can play with the diagramshere, where the boxes link to the corresponding part of the report.^{^}Green boxes correspond to inputs, red boxes are assumptions or limitations, and blue boxes are classed as “other”.

^{^}I’ve written a

summary of the reportas part ofthis sequence, if you’re interested!^{^}One way to think about this is as a distinction between

“inside view” and “outside view”approaches (however see alsothis post). Cotra’s bioanchors report takes an inside view, roughly based on the assumption that training compute is the biggest bottleneck to building TAI, and quantifying how much we’ll need to be able to train a transformative model. Davidson’s semi-informative priors report instead specifies very little about how AI development works, leaning more heavily on reference classes from similar technologies and a general Bayesian framework.^{^}This is a variation of the

sunrise problem, which was the original problem thatPierre-Simon Laplacewas trying to solve.^{^}This is of a course a somewhat dubious assumption, and we’ll come back to this later on.

^{^}Indeed, looking only at the base rate of successful first trials alone would have a big problem of sparsity – there’s just not enough historical data!

^{^}We could also think about the number of virtual

trialsrather than virtualsuccesses, but Davidson decides against this. Loosely speaking, if we use virtual trials, then it’s not as easy to separate out the effects of the first-trial probability and the effects from observed failed trials (more).^{^}The prior is defined using a

Beta distributionparameterised by (1) the number of virtual successes, and (2) the inverse of the first-trial probability. Seeherefor more information.^{^}The “plausibility of the prior” focuses on the shape of the

Beta distribution, e.g. whether or not you should expect the probability density to be larger in the interval [0, 1/1000] or [1/1000, 2/1000]. On the other hand, the “plausibility of the update” looks at your expected probability of building AGI next year should change given the outcomes of newly observed trials. For example (borrowing from the report), “If you initially thought the annual chance of developing AGI was 1/100, 50 years of failure is not that surprising and it should not reduce your estimate down as low as 1/600”.^{^}This approach also applies to researcher-years and compute years, and is described more

here.^{^}Incidentally, this is a claim that’s central to another of

Open Philanthropy’s Worldview Investigations,, whichForecasting TAI with biological anchorsI’ve discussed in another post.^{^}Note that this doesn’t imply that there’s an infinite probability of developing AGI in the first researcher-year of effort, because it’s not true that we’re starting from the “zero” level of AI technological development. Essentially, the regime start-time is

notabout “when the level AI technological development started increasing” –see this footnotefor more on discussion.^{^}For example, we would like our prediction for P(AGI within 10 years) to remain the same even if we use a trial definition of 1 month instead of 1 year. Although using a trial definition of 1 month would ordinarily lead to more total observed trials and thus more updating, this effect is cancelled out by choosing a different first-trial probability.

^{^}More concretely, suppose you think that several different updates rules (corresponding to e.g. different numbers of virtual successes) all seem reasonable, and you’re uncertain what to do. One approach is to weight the results for the different choices of update rules, and use these rules to update the forecasts based on evidence. But we might also be interested in

updating how we weight the update rules, which is where the hyper prior comes in (more).^{^}These numbers were extracted using

WebPlotDigitizer.^{^}Depending on your point of view, this may not be very compelling evidence – e.g. you might think that the ramp up to AGI would be extremely fast due to the discovery of a “

secret sauce”.^{^}You can also have a look at the

full reportif you want to get into the details!