shy's journal

Artificial General Intelligence (AGI): What Is It?

In my circles, I’ve been hearing the term “P(doom)” a lot. It estimates the probability of existential catastrophes due to artificial (general) intelligence. I’ve heard hype and doom. I know people leaving their jobs to join AI companies, and others who have decided the probability of doom was so large that it was best to live a hedonistic lifestyle.

I wrote off their responses as dramatic. Artificial general intelligence (AGI), as known by insiders, is an ill-defined term. Why upend one's life choices so callously? Andrew Ng, pioneering AI scientist and co-founder of Google Brain, said in a January 2024 newsletter:

AGI is one of the tech world’s hottest buzzwords, yet it has had no clear definition… This lack of specificity makes it hard to talk about related technology, regulation, and other topics… OpenAI defines AGI as “a highly autonomous system that outperforms humans at most economically valuable work.” This definition, had it been formulated in the early 1900s, when agriculture accounted for 70 percent of work globally, would have described the internal combustion engine. - The Batch

General intelligence

The working definition of AGI in this essay, and somewhat broadly accepted in the industry, is that AGI is artificial intelligence (AI) that possesses general capabilities, rather than narrow skills.1 For example, ChatGPT demonstrates general competencies across many domains—law, coding, writing, science, mathematics—which is different from AlphaGo, an AI that is specifically good at playing Go.2 3

I now claim that large language models like ChatGPT are already AGI, and should be considered as such; they are just not very high-performing ones. They clearly have general skills, even if weaker in "reasoning" and mathematics. My claim may be unpopular, but it is important for conceptual understanding, so we can start seeing AGI as a spectrum rather than a singularity. The point at which we decide something is “AGI” is unimportant; the point is that we are getting there. By increasing the breadth of competencies a system like ChatGPT can have, and by increasing the depth of these competencies like improving accuracy and performance, we are building a more and more powerful AGI.

“General intelligence” must be thought of in terms of a multidimensional scorecard, not a single yes/no proposition. — Noema

Now, why is it that AI research labs hesitate to call language models today “AGI”? One reason is the models' clear deficiency in some other areas (e.g. counting the number of G's in "strawberry"). Another reason is funding. Research labs like OpenAI and Anthropic use AGI for fundraising; if you've reached AGI, why need more funding? I don’t want to oversimplify or over-index on this, but “AGI” is also a fundraising gimmick. (The people from whom you hear "AGI" a lot are probably in the business of fundraising or garnering public support.) It captures our imagination, as a fill-in-the-blank for whatever you imagine a powerful AI to be.

Superintelligence

In fact, a lot of what people imagine as “AGI” is more precisely captured by artificial superintelligence (ASI), which is an AGI that surpasses humans across its general competencies.4 When AlphaGo beat top Go player Lee Sedol, we can claim it's “superintelligent” because it beat the best human at Go. But, as mentioned before, it’s not AGI, let alone ASI, because it does not outperform humans in other tasks (e.g. writing, math). We can build AGI towards ASI, when AGI attains superintelligence in more and more domains.

When labs like Anthropic envision "an entire nation of geniuses," they really allude to AGI possessing superintelligence in multiple domains. "Models will outsmart Nobel Prize winners... Bye-bye cancer, infectious diseases, depression; hello lifespans of up to 1,200 years." If AGI were to solve these problems that humans have yet to crack, it would imply that it surpasses human intelligence, becoming superintelligent.

When AGI becomes "AGI?"

By calling large language models like ChatGPT "AGI," I am making a claim that we're already at a class of AI that is not just incrementally better than past AI, but breaking into a new category of AI. We've never had such powerful AI that can be reasonably deemed "competent" in multiple general domains. They are AGI, artificial general intelligence. They will become more and more powerful AGI.

The varying thresholds AI labs use to classify their models as "AGI" underscores the inherent arbitrariness of their definitions. The point at which AI labs might call their models "AGI" might be when AGI reaches superintelligence. It might also be when labs deem AGI capable in replacing humans as AI "agents" in most industries. The latter seems to be OpenAI's approach, given their emphasis on "economically valuable work" and this recent job posting where they were trying to answer the question "how close are we to AGI" by deploying their models to "strategically-important domains."5 These "AGI" milestones are made up by AI labs. However, regardless of your favorite definition, notice that the path to get to "AGI" is by expanding the breadth and depth of competencies of AGI we already have. Speaking in terms of the latter helps us stay grounded amidst hype and branding.

Specificity Matters

Words shape our understanding of AGI, and specificity matters if we want to navigate change effectively. I propose that current large language models are already AGI. It help us start seeing AGI as a class of artificial intelligence, and stop seeing it as a singularity. This nuanced understanding already prevails among many AI researchers, and it's time we incorporate these insights into our ordinary discourse. As we discuss the future in terms of AGI—safety, hype, or doom—let us shift toward examining specific outcomes tied to concrete capabilities (e.g. AGI improving bio-weapon synthesis) rather than fixating on an arbitrary tipping point. When someone presents vague concerns without identifying specific outcomes, we should respectfully challenge them to provide greater clarity and precision.

Now that we have a working vocabulary and common understanding of today's AI developments, in my next piece, I will write about how far we are from "AGI". In other words, what breadth and depth of competencies are we missing?


  1. This paper lays out the clearest definitions for AGI that I know. It also classifies AGI into 5 levels (that are also applicable to narrow AI). Borrowing from Andrew Ng’s discussion of the paper, “Level 1 (“emerging”) matches or slightly exceeds unskilled humans. Levels 2 (“competent”), 3 (“expert”), and 4 (“virtuoso”) systems surpass the 50th, 90th and 99th percentiles of skilled human performance, respectively. Level 5 (“superhuman” or “artificial superintelligence”) outperforms 100 percent of skilled humans.”↩

  2. AGI is not the same as narrow AI trained on multiple tasks. Narrow AI systems excel at specific tasks they've been explicitly trained for, while general AI demonstrates emergent capabilities beyond the predefined parameters established by researchers. So, multi-task AI that live within the confines of tasks envisioned by the engineers are still narrow AI.↩

  3. I want to write about why current large language models are a step-function or discontinuity from previous narrow AI models, without breaking the flow of the main article since I've already finished it. (This is an add-on after publishing.) Reason 1: These AI models excel at tasks generally. Reason 2: These AI models excel at a wide variety of tasks without being explicitly trained on each one. Reason 3: Large language models can perform the "meta-task" of being "instructable," so they can learn in-context and extend their capabilities beyond any training corpus, making them truly general.↩

  4. People might define ASI as competence across all human capabilities, but "all human capabilities" is not well-defined. You can presumably keep finding more granularity and exceptions to the original set that you find relevant. For example, consider an AI that's the best at balancing chestnuts on its toes. Or, best at talking to my mother as her child. Or, best at being me. It’s also not clear that some of these competencies have clear evaluations, especially around art and aesthetics. That is why I find a domain-by-domain, grounded, and precise discussion of AGI more productive and useful.↩

  5. Their goal should maybe terrify you! But I don't think the world of economic valuable world is exhaustible, because—surprisingly—I have confidence in human intelligence. I don't think I could have that confidence if not for my spiritual practices, since it's in vogue and even rational to underestimate human intelligence to correct for hubris and bias. But some things, you just know. Or, if God is willing, I can just know.↩