Stop Asking 'Is It AGI?' — The Answer Is Boring

The 'Spark of AGI' is a Funhouse Mirror

Everyone wants to know if we've achieved Artificial General intelligence. GPT-4 triggered a thousand takes. Claude 3 Opus just fanned the flames. Researchers talk about 'sparks of AGI' and tech Twitter debates sentience.

It’s the wrong conversation.

AGI isn't a light switch that flips from 'off' to 'on'. It's a messy, uneven spectrum of capabilities. Today's frontier models are like a genius savant who can write a fugue but can't figure out how to tie their own shoelaces. They can do things that feel like genuine reasoning, then make a mistake a five-year-old wouldn't.

The real question isn’t whether it’s “general intelligence.” The real question is: where can I trust it to work without a human babysitter?

Moments of Frightening Competence

Let's be clear: these models can do things that were science fiction three years ago. You can give a model like Claude 3 Opus a high-level goal and watch it produce genuinely surprising results.

For instance, I recently fed it a messy 20-page PDF of user research notes and asked it to act as a product manager.

My prompt was simple:

Read this user research. Identify the top 5 user frustrations. For each frustration, propose a concrete feature idea to solve it. Prioritize these 5 features using the RICE framework (Reach, Impact, Confidence, Effort) and explain your scoring. Present this as a markdown table.

What it produced wasn't just a summary. It synthesized conflicting user quotes, identified subtle patterns, invented five plausible feature ideas, and then correctly applied a prioritization framework with defensible, albeit estimated, scores. It did the work of a junior PM in about 90 seconds. That feels intelligent.

This is where the AGI talk comes from. When a model can connect disparate concepts, apply a learned framework, and generate a structured, novel output, it's performing a task that requires multi-step reasoning. It feels like more than just pattern matching.

The Stupidly Simple Failure Modes

Then, in the next chat window, the same model will fail spectacularly at something simple. This is the other side of the coin, the part that proves we are not dealing with a human-like general intelligence.

These models have no real-world grounding. No physics model. No common sense baked in from lived experience. They know the word 'heavy' is associated with 'anvil' and 'light' with 'feather,' but they don't understand gravity.

Ask one to plan a trip to a grocery store:

• It can: create a perfect shopping list and find the most efficient route on Google Maps.
• It can't: tell you not to put the eggs at the bottom of the bag under the canned goods.

This isn't a one-off quirk. It's a fundamental limitation. They are systems of statistical relationships, not understanding. Other classic failures are still common:

• Basic Arithmetic: They can solve calculus problems by mimicking textbook examples, but still make simple addition errors on large numbers because they're 'predicting' the answer, not calculating it.
• Logical Contradictions: You can often trick a model into holding two contradictory beliefs in the same response. It has no internal consistency checker.
• Spatial Reasoning: Ask it to describe the layout of a room based on a text description and watch it get tangled in knots.