Levels of Autonomy in AGI

Trying to understand the categorization of AGI from Level 0-5 in a recent paper by Google DeepMind

Dec 19, 2023

Google DeepMind has proposed a framework for “classifying the capabilities and behavior of Artificial General Intelligence (AGI) models”. In this post, I will share some thoughts regarding the different levels of AI autonomy.

Google DeepMind’s framework is inspired by the six levels of autonomous driving used by the U.S. Department of Transportation to assess the level of autonomy in self-driving cars - from 0 (fully manual) to 5 (fully autonomous).

Google DeepMind’s corresponding framework for assessing the level of intelligence in AGI models is divided into three parameters: “performance”, “generality”, and “autonomy” in AGI models. Below is the table for AGI autonomy.

Level 1 & 2 Autonomy

Currently, AI’s level of autonomy in modern work offices is typically somewhere between Level 1 and Level 2 in my estimation.

At this interim stage, knowledge workers may use generative AI tools like ChatGPT to explain concepts, translate sentences, or do grammar checks on writing. Many larger or just slightly more tech-savvy offices are developing custom GPTs or synthetic videos for internal use and employee training. This is what I would characterize as Level 1 autonomy.

At Level 2 autonomy, AI tools are used a bit more extensively. Perhaps to summarize documents or to brainstorm ideas. Most developers use GitHub CoPilot or other AI coding tools in their daily work, which is another Level 2 use case.

As I see it, the step from Level 1 to Level 2 is fairly substantial. The difference is that workers are still in full control of the work they are doing at Level 1, but from Level 2 and onwards, workers delegate smaller and increasingly larger tasks and responsibilities to the AI – work that they would normally do but lack the time, energy, or mental bandwidth to deal with. At Level 2, AI tools shine as productivity boosters but on the downside, the “AI consultant” can make mistakes that are easy to overlook for humans and impossible to understand, explain, or learn from in hindsight.

To exemplify what a typical Level 2 mistake could look like, we can look at the magazine Sports Illustrated which was recently exposed by Futurism for relying on AI writers from an external company. In one of the magazine’s articles, the deep fake author, Drew Oritz, gave recommendations on the best volleyballs to buy. He warned that “volleyball can be a little tricky to get into, especially without an actual ball to practice with." A strange, alien-like statement that a human writer would likely never make.

Level 3 Autonomy

As AI becomes more entrenched in organizations at Level 3, the stakes get higher, and the risk landscape gets more complex.

In a Harvard Business School paper that I analyzed last month, the authors imagine GPT-4 as a collaborator, corresponding to Level 3 autonomy. At this level, you could theoretically “hire” GPT-4 instead of a “median human” to perform a task. AI optimists like Ethan Mollick would likely claim that generative AI at this level of autonomy should be used for office work already today. There are of course many problems with this. For example, that AI:

cannot easily adapt to new knowledge and circumstances
struggles to catch subtleties
cannot understand “tacit” knowledge that humans acquire from living and experiencing the world
have problems with long-term memory
cannot provide satisfying explanations for decisions or outputs it makes
lacks a coherent understanding of past, present, and future

AI models’ understanding of the world is limited to 0s and 1s while the human brain is vastly, more complex. I never understood when experts such as Yann LeCun compare AI to human or animal intelligence. I don't see how it's comparable.

Inflection CEO and DeepMind co-founder, Mustafa Suleyman, has an insightful quote about the difference between human and AI intelligence in his book “The Coming Wave”:

“A big part of what makes humans intelligent is that we look at the past to predict what might happen in the future. In this sense, intelligence can be understood as the ability to generate a range of plausible scenarios about how the world around you may unfold and then base sensible actions on those predictions. Back in 2017, a small group of researchers at Google was focused on a narrower version of this problem: how to get an AI system to focus only on the most important parts of a data series in order to make accurate and efficient predictions about what comes next.”

Suleyman is referring to Google’s “Attention is All You Need” paper that led to the ChatGPT revolution. I think it's an interesting idea that human intelligence is marked by our ability to predict what might happen in the future based on our past. Large language models cannot do the same as they don’t have a past to remember or a future to experience. They don’t have any intrinsic motivation that guides them in their actions like humans do. For these reasons, it is arguably wrong to treat AI models as co-working equals with Level 3 autonomy.

Level 4 & Level 5 Autonomy

A survey of US-based developers by GitHub shows that 9 out of 10 developers use AI coding tools like GitHub Copilot. At the same time, coding is one professional field that faces a severe extinction threat from full automation. Google DeepMind’s newly released coding tool AlphaCode 2 could outperform 86% of humans in competitive coding (technical report). As we get to Level 3 autonomy and AI’s capabilities rapidly advance, it may only be a matter of short time before we reach Level 4.

At Level 4, massive job displacement is bound to occur. People will struggle to find meaning in life as they see their hard-earned skills being replaced by AI systems. The roles suddenly shift. At Level 3, workers collaborate with AI but before they know it, the AI is no longer a useful productivity-boosting tool but doing their jobs and humans have an assisting role at best. Even this assisting role will of course diminish in importance over time and eventually, AI will take over completely as it reaches Level 5 autonomy.

At Level 5, it will be clear that human qualities we up until this point in history have cherished as unique and significant, are not really so. We will realize that many of the skills, achievements, and emotions we in the past have derived a sense of identity from can be replicated and outdone by autonomous agents. It's wild to think about.

My main point is: automation may have a rolling-snow ball effect. As I said earlier, I think that the move from Level 1 to Level 2 autonomy is a substantial shift. Level 3 is more of a temporary stage before we get to Level 4 and this is where the roles switch, AI tools are no longer assisting humans but vice versa. Level 5 AGI is a great unknown, but if we rush too fast to get there, the consequences of creating AI with such an extremely high level of autonomy will likely be unmanageable.

Closing Thoughts

For the last couple of decades, industry leaders and tech wizards have claimed that fully autonomous cars were right around the corner. Today, not a single fully self-driving car is on the roads anywhere. We are so close, and yet so far away. In contrast to autonomous cars, we cannot even conceptualize what Level 5 AGI looks like or predict how it will impact the world. Hopefully, AGI will follow a similar trajectory to autonomous cars and hit a stumbling block. Too much change too fast is not good.

Reads of the Week

Excuse me, but the industries AI is disrupting are not lucrative - Erik Hoel, December 8, 2023 (The Intrinsic Perspective).

Albania to speed up EU accession using ChatGPT - Alice Taylor, December 13, 2023 (Euractiv).

ChatGPT is winning the future — but what future is that? - David Pierce, Nomber 30, 2023 (The Verge).

Judges Given the OK to Use ChatGPT in Legal Rulings - Thomas Germain, December 12, 2023 (Gizmodo).