Google DeepMind's CEO Demis Hassabis has listed a number of research breakthroughs or new capabilities he believes are needed for artificial general intelligence (AGI). I caught at least two that he mentioned when I listened to a recent episode of Google DeepMind's podcast:

• Continual learning, a longstanding problem in AI, the ability to learn in every new moment, as in humans do, rather than only learning every few months when a new training run happens (sometimes Hassabis seems to use the term "long-term memory" interchangeably with continual learning, but in the recent podcast he explicitly says continual learning)
• World models, which, as I understand the term, is somewhere between a philosophical concept and a technical term, referring to some general ability to understand, predict, and model how the real world works, such as the way humans and other mammals have an intuitive understanding of physics

In a January interview, he also mentioned the following missing capabilities:

• Reasoning, which he also discusses on the podcast, noting a contrast between chatbots' impressive performance on some advanced math problems and a preponderance of elementary mistakes^[1]
• Hierarchical planning, the ability to plan actions that are composed of sub-actions, which are composed of sub-sub-actions, and so on, in a nested hierarchy
• The ability to creatively generate novel hypotheses or conjectures

Here's the quote from that interview:

The models today are pretty capable, but there are still some missing attributes: things like reasoning, hierarchical planning, long-term memory. There's quite a few capabilities that the current systems don't have. They're also not consistent across the board. They're very strong in some things, but they're still surprisingly weak and flawed in other areas. You'd want an AGI to have pretty consistent, robust behavior across the board for all cognitive tasks.

One thing that's clearly missing, and I always had as a benchmark for AGI, was the ability for these systems to invent their own hypotheses or conjectures about science, not just prove existing ones. They can play a game of Go at a world champion level. But could a system invent Go? Could it come up with relativity back in the days that Einstein did with the information that he had? I think today's systems are still pretty far away from having that kind of creative, inventive capability.

Remarkably, immediately following this, Hassabis says he thinks AGI is "probably three to five years away."

I'm skeptical of such predictions for (at least) two reasons:

1. Many such predictions have come and gone. At very beginning of the field of artificial intelligence in the 1950s, a similarly ambitious goal was set for the first summer of AI research. Anthropic's CEO Dario Amodei predicted in March that AI would take over 90% of coding by September and, well, here we are. (It's not even true at Anthropic!)

2. Hassabis's discussion of the remaining research breakthroughs or missing capabilities seems incongruent with a prediction that those breakthroughs will be made or those capabilities will be developed in such a short time. Hassabis gives the impression he's not even sure he knows the full list yet of what is still missing. And he really thinks the longstanding problems in AI research that he did list will be solved in such a short term? I don't understand what could possibly justify such confidence.

I think it could be an interesting exercise to, rather than (or in addition to) forecasting AGI as a single idea, to decompose AGI into research problems or capabilities like continual learning, world models, reasoning, hierarchical planning, and creative idea generation, and then ask people to forecast those things individually. My guess is that for many or most people, reframing the forecasting question this way would lead to longer timelines for AGI.

I think before venturing a guess, people who want to try forecasting when these research problems will be solved should look into how long AI researchers have been working on them and how much research has already been published. Many of them, in fact I think all of them, are decades old. For instance, I searched Google Scholar and found several papers on hierarchical reinforcement learning from the early 1990s. (Searching for the exact phrase filters out unrelated stuff, but also misses some relevant stuff.) There is more funding now, but it seems like the vast majority of it is being spent on scaling large language models (LLMs) and AI models that generate images and videos, and very little on discovering the new science that Hassabis says is necessary to get to AGI.