A Definition of AGI

John: Alright, welcome to Advanced Topics in AI Cognition. Today's lecture is on 'A Definition of AGI' by Dan Hendrycks and a large team from places like the Center for AI Safety and MIT. John: We've seen a lot of recent work trying to pin down what AGI means, like 'Levels of AGI for Operationalizing Progress' or Chollet's 'On the Measure of Intelligence'. This paper takes a different tack by grounding its definition in human psychometrics. It argues that the current, nebulous definitions are hindering real progress. Yes, Noah? Noah: Hi Professor. When you say it's grounded in human psychometrics, does that mean they're proposing a kind of IQ test for AI? John: That's a good way to frame it, though they are careful to build a more comprehensive framework than a single IQ score. The authors argue the term 'AGI' has become a moving goalpost. Their central objective is to provide a concrete definition: AGI is an AI that matches or exceeds the cognitive versatility and proficiency of a well-educated adult. Noah: So the emphasis is on the breadth of skills, the 'versatility' part? John: Exactly. Versatility and proficiency. This is meant to combat what they call 'capability contortions'—instances where an AI uses a narrow strength, like a massive context window, to fake a more general ability, like long-term memory. The goal isn't just to see if an AI can pass a test, but to diagnose whether it possesses the underlying cognitive machinery. Noah: That makes sense. So it’s less about just solving datasets, and more about having the right internal architecture? John: Precisely. The framework is designed as a diagnostic tool to pinpoint architectural weaknesses that are currently being masked by clever engineering or immense scale. John: To do this, their entire methodology is built on the Cattell-Horn-Carroll theory, or CHC theory, which is the most empirically validated model of human intelligence. They adapt it to define ten core cognitive domains for AGI, each weighted equally to stress that versatility we just discussed. Noah: What are some of those domains? John: They include things you'd expect, like General Knowledge, Reading and Writing, and Mathematical Ability. But critically, they also include On-the-Spot Reasoning, Working Memory, and both Long-Term Memory Storage and Retrieval. They also specify Visual and Auditory Processing, and even cognitive Speed. Noah: Wait, Long-Term Memory is split into Storage and Retrieval? Why? John: An excellent question. It's a key insight of the paper. They separate them to highlight a specific failure mode in current models. An AI might be good at retrieving facts from its training data, which mimics retrieval, but terrible at storing new information from an ongoing interaction. That's the storage part, and as we'll see, current models score a zero there. It's also how they can measure issues like hallucination, which they classify as a failure of retrieval precision. Noah: So they use existing benchmarks for these, like ImageNet for visual or AP exams for math? John: Yes, they provide examples like those, but they frame them as 'task specifications' rather than fixed datasets. This is to prevent models from just memorizing the test set. The evaluation has to be continuous and robust, using the best available tests for each narrow ability within those ten domains. John: When they apply this framework, the results are quite revealing. They get what they call a 'jagged' cognitive profile for models like GPT-4. It scores well on knowledge and language but gets near-zero on things like Long-Term Memory Storage, Auditory Processing, and true reasoning. Noah: So it's not a smooth progression of capabilities. That zero on Long-Term Memory Storage is stark. It reminds me of the work on 'Personalized AGI via Neuroscience-Inspired Continuous Learning,' which explicitly tries to build separate memory systems to address that kind of amnesia. John: That's a great connection. This paper's findings strongly suggest that such architectural changes are necessary. Simply scaling current transformer models might not fill these gaps. For instance, they project GPT-5 will make significant gains, reaching 57% on their AGI score compared to GPT-4's 27%, but it still scores zero on memory storage. The model remains fundamentally amnesic, unable to learn from experience over time. Noah: So all the talk about Retrieval-Augmented Generation, or RAG, is essentially a patch for a bad memory retrieval system? John: They would argue it's a perfect example of a 'capability contortion.' It's an external tool bolted on to compensate for a core cognitive deficit—in this case, imprecise retrieval and a tendency to hallucinate. This framework makes those architectural weaknesses impossible to ignore. John: Ultimately, the significance of this work is that it moves the AGI discussion from a philosophical debate to a problem of scientific measurement. It provides a diagnostic tool and a roadmap, shifting the focus from simply achieving high scores on narrow benchmarks to building integrated, versatile intelligence. John: The key takeaway is this: to build true AGI, we need to stop focusing only on what AI can do and start rigorously measuring and addressing what it can't. This paper offers a clear, human-centric lens through which to do that. John: Thanks for listening. If you have any further questions, ask our AI assistant or drop a comment.