In September 2026, OpenAI launched its new model GPT-6 Astra. OpenAI president Greg Brockman said, “Welcome to the AGI era.” AGI is Artificial General Intelligence, which I believe to the be the threshold for achieving true artificial intelligence. It is a holy grail for AI companies, many of whom are trying to convince us that AGI has already been achieved. After all, those venture capital investment dollars aren’t going to raise themselves.
Brockman did acknowledge that many people don’t really understand what AGI is.
Everyone has a different definition of AGI. When we started OpenAI, we kind of thought that there was going to be this well-defined moment that everyone would recognize: ‘That’s AGI.’ It hasn’t played out like that. It’s a much more gray, fuzzy thing.
It’s certainly advantageous to promote the fuzziness of the term AGI when you’re trying to convince everyone that you’ve achieved AGI without actually having achieved it. Brockman is correct, though, when he says that the term AGI isn’t well-defined. So let me well-define it.
AGI is Lt. Cmdr. Data from Star Trek: The Next Generation. If you’re not familiar with the show, Data is an android, a fully artificial being who serves as second officer aboard the starship USS Enterprise-D. He has a fully human form, but is better than humans in some respects. He can process information much faster than humans and has perfect recall.

Data has several attributes that separate him from today’s Large Language Models (LLMs) such as GPT, Claude and Gemini.
Continuous vs static learning
LLMs are pre-trained. It’s what the P in GPT stands for. Pre-trained models are frozen after training. They require costly fine tuning or retraining to incorporate new knowledge. Data learns continuously, updating his models in real-time to accommodate new knowledge and experiences.
True reasoning vs. pattern-matching
LLMs are probabilistic content predictors optimized to guess the next piece of content (e.g., word or token) given a prompt and context. They mimic logic, but they operate completely in the syntactic domain. As a result, they struggle with abstract reasoning and have problems with hallucination. Data has a true reasoning capability. He operates in the semantic domain in addition to the syntactic domain. He doesn’t suffer from hallucinations like LLMs do.
Common sense and world models
LLMs do not inherently understand physical reality or cause-and-effect. They have functional linguistic competence, meaning they can speak fluently about a world they don’t actually experience. Data understands physical reality. He can map words and text onto the physical objects, feelings and emotions they represent. He can understand cause-and-effect, and thus can understand the consequences of his actions. LLMs cannot understand the consequences of their actions1.
Learning and agency
LLMs generate immediate outputs rather than spending time pondering. If you’re using Claude Code, it says it’s “thinking” or “pondering” or any other present imperfect verb, but that’s just filler while it’s doing its computations. Data has system 2 thinking, the ability to slow down, plan multi-step actions and adapt strategies mid-execution. We exhibit this level of thinking every time we stop execution of a plan because it’s gotten too complicated, and we back up and rethink the plan. An LLM will do this if we tell it to, but not on its own.
Initiative
Data wants to be more human. Towards that end, he has taken up painting, acting, poetry, playing the violin and trying to learn how to tell jokes. He even keeps a cat as a pet. No one told him to do these things, he did them on his own because he wanted to. LLMs don’t have anything approaching this level of initiative.
LLMs have a long way to go before they can reach the level of AGI. Today’s AIs may look intelligent based on their external behaviors, but unlike Data they are not intelligent in any real sense. We are intelligent, so it’s imperative that we remain in the loop with AI. To responsibly use AI, we need to keep it human.
1 Work in the area of world models has the potential to allow AIs to understand cause-and-effect relationships by training them to model the real world and predict the next state of a world model that results from an action applied to the current state of the model. Any true AGI will have this capability.
Leave a Reply