Every few months, somebody at a frontier AI lab thinks the rest of us ought to worry about a new acronym. It began with AGI. Then it became ASI. And if you’ve been hanging out on AI Twitter (or X) ever since GPT-6 Astra, Fable, and Mythos 5.1 came out, you’ll probably have seen both acronyms thrown around as though they’re synonyms. But they are not. One is a point on a timeline. The other is an endpoint nobody has really seen yet, and anybody who says they have a clear idea of what awaits beyond it is making a guess, regardless of how fancy their hardware is.
Also read: Inside online AI selfie boom: Hidden cost of viral AI image trends explained
AGI, in other words, is an artificial intelligence which has been able to learn all intellectual skills performed by humans, regardless of the fields involved. Your smartphone’s artificial intelligence could be asked to write emails and summarize documents for you. However, if it is asked to perform an operation for which it was not trained, it fails to complete the task successfully. That would not be the case for AGI, which would reason about law, programming, biology, or small talk in the same manner a person reasons on Tuesday.
However, the problem is that everyone disagrees on the finish line. According to OpenAI, for example, AGI has been defined at various moments as the ability to perform “most economically valuable work” better than humans. Researchers want to have proof that the machine has actually been able to reason and not simply memorize data. Therefore, when an institute claims that their new model is “a taste of AGI,” it means that their marketing department is overstating the results of their benchmark.
ASuperIntelligence is the next step up, and a much larger leap. The term refers to an AI system which not only equals but surpasses human intelligence across all domains – scientific research, planning and strategizing, creativity, and so forth. This isn’t an “intelligent colleague” but rather an entity which “could redesign the Internet, cure a disease and negotiate better than a bunch of diplomats before lunch time.”
Also read: Consumers today are not price sensitive; they are more value-centric: Cellecor CMO Jubin Rawal
It’s here that things start to get murky. AGI, at least in theory, is something we can actually test for. There are metrics, benchmarks, real-world applications and implementation. ASI is, by contrast, a bit more hypothetical in nature, since we don’t have a working model to compare to. All projections about ASI are speculative.
Due to the fact that the past couple of months have been busy. GPT-6 Astra published benchmarks that got the OpenAI CEO to mention the word AGI. Anthropic’s Fable and Mythos 5.1 increased capability for reasoning and agentic tasks. Each model release sets the bar higher for what “impressive” means, and each set sends the debate about AGI a little bit further into ASI realms of discussion, despite the fact that the models are far from general in nature. A model that performs well on a coding challenge does not necessarily understand its motivation.
This is the issue I have with the discussion. When talking about AGI versus ASI, it becomes some sort of a clock ticking away to some kind of a defining moment. Nothing is going to happen like that. Progress is sloppy, inconsistent and involves many systems that can do superhuman things in certain areas (folding proteins, chess games, code reviewing) but are absolutely incompetent with regards to things that could be done easily by a child of five.
Whenever you read an article saying that the model is close to achieving AGI, ask yourself what it is good for, rather than believing what the slides say. If someone mentions ASI in your daily conversation as if it is some kind of an appointed event, then keep in mind that there isn’t a solid consensus about its predecessor. The difference between AGI and ASI lies not in science, but rhetoric. AGI is something that we are getting closer to. ASI is our story about inevitability of that closeness.
Also read: The AI Music land grab: Why every major label is suddenly signing AI deals