Examining AGI Timelines

During a recent email exchange, a reader of the blog brought up the question of technology / AI timelines and what progress in that domain might look like. His question struck a chord, reminding me of the uneasy feeling I often get when reading the optimistic predictions of others (including futurist Ray Kurzweil and philosopher Nick Bostrom), some of which are summarized on the popular Wait but Why blog. This post will attempt to explore the rationale for that feeling further, with the intent of seeing if we can bring any more accuracy to the incredibly difficult domain of future prediction.

To start, it may be helpful to summarize the general sentiment espoused by Wait but Why and other future predictors (though I’d recommend instead reading the linked post above). The first part of that post is focused on the idea that progress has been accelerating and is best modeled as an exponential process, rather than a linear one (which is how we tend to think about things when imagining the future). The argument is that if you took someone from 250 years ago and showed them the world today, it would be vastly different from anything they knew (phones, computers, electricity, vehicles, and more would all be completely new). However, for the world of 1750 to elicit the same reaction, you’d have to go back far further – perhaps to pre-civilization times ~10,000 years prior. The claimed reason for this quickening is that more advanced societies have more tools / resources at their disposal, which allows them to continue to advance even faster – resulting in a curve like the below.

In the second part, the author spends some time reviewing the role that artificial intelligence plays in our world today, and then outlines some potential paths to artificial general intelligence (including copying the brain and leveraging evolutionary algorithms). He then makes the case that this jump to AGI could happen quickly because of the previously discussed exponential rate of progress, and so we might see it within the next 20 or 30 years, with superintelligent AI following shortly after (due to recursive self-improvement). 

This framing of the nature of progress and the short timelines until AGI are not unique to Wait but Why – as mentioned in the post, the author pulled together these ideas and timelines from the writings of AI scientists and researchers, and I’ve seen these views commonly espoused. It’s certainly an alluring idea that the near future may be nothing like the present; given a choice, I think many (especially most AI scientists and researchers) would want to exist where the little man is on the chart shared above, right before the progress explosion (I know I would). However, it’s not clear to me that the arguments hold up under further scrutiny.

Making predictions about the future is really hard. It’s extremely difficult (and oftentimes impossible) to understand how the current workings of the world will come together to form the world of tomorrow (or the world of a decade / century / millennium from now). Wait but Why simplifies this process by looking at the general “shape of the line” of past progress, and simply extrapolating forward. Progress has seemed fairly exponential, and so a prediction is made that it will continue to be fairly exponential. However, we can do better than that. Each advancement has its own unique character (e.g., the steam engine is distinctly different from the computer) and we can analyze those characters further to better understand the why behind past progress. Viewing history through this lens (rather than simply as an exponential line of progress) provides further clarity into some of the potential paths forward (though still not much – as mentioned, predicting the future is hard!). 

The key insight from looking more deeply at past progress is that it has not been uniformly exponential. There were periods of innovation in certain areas (e.g., energy production, transportation, economics, etc.) followed by the widespread adoption of these innovations (which generally looked exponential), followed by periods of more linear advancement (at least within that domain). For example, automobiles were invented in 1886, popularized in the early 1900’s (by Ford and others), and have enjoyed a relatively (slow) linear progression since (with advancements mainly in safety, comfort, and speed). Much of the exponential progress since ~1700 has been sustained by a continued stream of new innovations, with each seeing exponential initial adoption followed by gradual linear improvements.

However, the last ~75 years have been of a different nature. We have not been reliant on innovations in new domains; rather, progress has continued exponentially due (mainly) to improvements in the domain of computing (if we look outside this domain, progress seems more linear – for example, 10 of Hilbert’s 23 problems remain unsolved, physics still lacks a unification of quantum mechanics and general relativity, and nuclear fusion is still under development). The computer’s relation to innovation is different from that of the earlier advancements; unlike early cars (or steam engineers, or telephones, etc.), which could do most of what modern cars can do (i.e., get you place to place), early computers could do barely any of what modern computers can do. Computers operate in the world of information, and their utility in that world is dependent on their processing power. As we’ve found ways to increase that processing power (following the trend of Moore’s Law, with the number of transistors on a chip doubling every two years) we have identified new uses for computers. The 1950’s brought computers to the business world, the 1970’s to the public, the 1990’s to phones, and the 2000’s / 2010’s to nearly everything. Each year, computers gain the ability to generate, store, and transfer (exponentially) more information, leading to a feeling of quickening progress. 

However, it seems there may be a limit to the progress possible via this avenue alone. There’s only so much information generation, storage, and transfer that is useful to humans – and we seem to be close to reaching those levels. For example, in the last 10 years computers have gotten significantly faster, but the experience of using computers has remained largely unchanged (or at the very least, has progressed in a linear manner). 10 years ago, phones could browse the internet, multiplayer games could be played online, and business presentations could be given in PowerPoint (with the analysis conducted in Excel), just like today. To parallel the example from Wait but Why, a person from 2010 coming to 2022 would likely be less surprised than a person going from 2000 to 2012. We’re reaching the limits of the information processing power required for people to get “full use” out of their devices, and as such, this type of progress seems positioned to revert to a linear track.

One area where we have seen continued near-exponential advancement is in the field of AI research. Computers didn’t surpass humans at Go until 2015, and that achievement was followed by significant advancements in language models (e.g., GPT-3) and image generation (DALL-E), among other areas. Artificial intelligence, as currently structured, requires massive amounts of computing power, and so continued improvements in transistors per chip will be of use in this field. The key question is whether advancements in this domain will be sufficient to maintain the exponential rate of overall progress. I agree with Wait but Why that narrow AI will not be sufficient; its use cases are too (for lack of a better word) narrow, and it requires too much human oversight and data cleansing / manipulation to drive an exponential impact. For AI to sustain the next phase of exponential progress, we’ll need to get significantly closer to artificial general intelligence.

Wait but Why lays out several of the reasons AGI is hard, but then proceeds to assume we’ll get there soon anyway due to the exponential rate of progress. This seems to be putting the cart before the horse – AGI is a requirement for continued exponential progress, not a result of it. Scale up a language model 1,000,000x, and you’re still left with a (far more powerful) language model. Constructing generally intelligent systems will require far more innovation – and while we do have our brains as a working example, we seem to be far from understanding even simple things about them (like why we sleep or in what type of way information is coded). Barring some type of mass destruction event, it seems inevitable that we will achieve AGI eventually, but a timeline of 20 or 30 years (or even 50) puts no stock in our present lack of understanding regarding intelligence.

This lack of understanding is further highlighted in the below excerpt:

It takes decades for the first AI system to reach low-level general intelligence, but it finally happens. A computer is able to understand the world around it as well as a human four-year-old. Suddenly, within an hour of hitting that milestone, the system pumps out the grand theory of physics that unifies general relativity and quantum mechanics, something no human has been able to definitively do. 90 minutes after that, the AI has become an ASI, 170,000 times more intelligent than a human.

Wait but Why (https://waitbutwhy.com/2015/01/artificial-intelligence-revolution-1.html)

This is not how intelligence works. Every working example of intelligence involves some initial internal configuration (e.g., a newborn and their brain, or a neural network algorithm) which is then exposed to external stimuli, with the initial configuration structured in such a way as to “make sense” of the stimuli over time (e.g., through our synapses being potentiated or depotentiated, or through the neural network’s weights being updated via gradient descent). The process for making a system more intelligent is to change the initial configuration and subject it again to stimuli – not to directly change the end-state configuration (e.g., update the neural network’s weights directly). There’s still an opportunity for “scaffolding” (once AGIs are more intelligent than humans, they can design better initial configurations), but it’s certainly not as immediate as laid out here. Each new configuration would need to be trained (mostly) from scratch, and it may well be that extremely significant training times (i.e., on the order of years) are required for AGI systems to develop (particularly for them to learn about humans, as we’ll be the most complex part of their environment, and we operate on human timescales).

With a more complete understanding of the drivers of past progress and the requirements for future progress, it seems we may be on a linear path until new innovations (beyond computers) are made. AGI is among the most likely candidates, but numerous others could also drive a return to exponential progress – for example, advancements in gene editing or nuclear fusion. Determining when these might happen gets into the “really hard” part of future prediction – but at least for AGI, it’s clear we have a long way to go.