The author argues for concrete definitions of AGI, RSI, and ASI, then claims AGI has already been achieved because GPT-6 Astra scored 99.9% on the ARC-AGI-3 benchmark and demonstrated generalization to novel tasks.
You've heard this before, and now you're hearing it again. You roll your eyes. The problem with this statement is that AGI has no accepted definition, so it's meaningless. Fine. I'll give you a concrete definition. I'm tired of the pointless arguing. Let's discuss upon a single shared definition. In fact, I'll give a concrete definition for the whole trinity: AGI, RSI, and ASI. Definitions follow: AGI AGI stands for Artificial General Intelligence. It is any AI that can successfully perform tasks that it was not trained to do. That is, the AI can generalize its specific intelligence (learned from training) to new, novel tasks. RSI RSI stands for Recursive Self-Improvement. It is the ability of an AI model to improve itself and create more intelligent versions without any human intelligence input. ASI ASI stands for Artificial Super-Intelligence. It is any AI that can outperform every single human on every single task. --- Okay, beuno? Now I'll explain my definitions, starting with AGI. AGI is the definition that is most disagreed upon. My definition of AGI is relatively narrow and strictly defined. In fact, I can point to a single test that determines if an AI model is AGI: the ARC-AGI-3 benchmark. (If you don't know what this is, look it up. TL;DR: it tests if AI can learn in novel environments it has never seen before.) GPT-6 Astra is the first AI model to score 99.9% on this benchmark, matching or outperforming humans. Because of this, I do agree with Sam Altman's announcement that claimed GPT-6 Astra was the beginning of the "AGI" era. We also see this generalized intelligence in real-world scenarios: Astra can play Minecraft, beat Portal, operate robotics, and do things that it was not specifically trained to do. Note that this definition does NOT necessitate any level of skill. It does not mean that AGI can replace all human jobs. Astra is not extremely good at playing Minecraft or operating robots—the point is that it can. RSI is a pretty agreed-upon definition. Once human researchers aren't needed to create new AI models, this ability is reached. If I could add one thing: I don't think RSI is necessarily a single point in time. Some basic level of RSI may be achieved, but it might suck. For example, the self-iteration cycle might take too long or cost too much. Likely the "foom" explosion point will only happen once a highly effective form of RSI is achieved. ASI is the one that everyone is dreaming about. An AI overlord is ASI. A cure to cancer will be achieved by ASI. Infinite energy will be discovered by ASI. This is the sci-fi goal we are trying to achieve. Note that just because current AI is super-intelligent in some areas (e.g. math) doesn't mean it is ASI—under my definition of ASI, it must be super-intelligent in all areas. --- Therefore, under these definitions: YES, we have already achieved AGI, and we are on the way to ASI. Probably RSI will be achieved soon. Written by human.
The article argues that AGI has already been achieved based on historical definitions, as LLMs outperform most humans in intelligence, and suggests that ASI is the next frontier.
The article analyzes GPT-6 Astra's 99.9% score on the ARC-AGI-3 benchmark, revealing that the score varies based on testing harnesses and questioning OpenAI's AGI interpretation amid clarifications from the benchmark creators.
The article explores whether a future AI model, GPT-6 Astra, could meet Demis Hassabis' proposed benchmark for AGI by independently developing general relativity from a 1911 knowledge cutoff.
The author argues that current AI models can solve narrowly defined tasks and speculates that iterative improvements could address LLM bottlenecks, potentially leading to AGI through recursive self-improvement.