Is AGI Here? A Sober Look at What AI Can't Do

The answer is no. We do not have Artificial General Intelligence.

Now that's out of the way, we can talk about the more interesting question: why does it sometimes feel like we do? GPT-4 and its peers can write poetry, pass the bar exam, and generate working code in languages they've barely seen. These aren't just parlor tricks; they're genuinely useful capabilities that feel like a new kind of intelligence.

But they are not general intelligence. They are incredibly sophisticated prediction engines. Conflating the two is a great way to misunderstand what these tools are for and a terrible way to build a product. The limitations are not just edge cases; they are fundamental gaps in understanding.

The 'Sparks of AGI' Mirage

Microsoft researchers published a famous paper claiming GPT-4 showed “sparks of AGI.” They gave it novel problems, like drawing a unicorn in TiKZ (a LaTeX graphics language), and it produced a shockingly good result. This looks like reasoning. It feels like the model understands the concept of a unicorn and the syntax of a new language and merges them.

What’s really happening is pattern matching on a scale humans can’t comprehend. The model has ingested a vast corpus of text that includes unicorn descriptions, code in countless languages, and tutorials on programming graphics. The most statistically probable sequence of tokens to satisfy “draw a unicorn in TiKZ” is, in fact, the code to draw one.

It’s a high-wire act of statistical prediction. When it works, it’s magic. When it fails, it reveals the lack of a world model. Ask it a simple logic puzzle that isn't well-represented in its training data, and it will confidently give you a nonsensical answer. It doesn't know why an answer is correct, it only knows that it's a plausible-sounding sequence of words.

Where It All Falls Apart: The Real World

AGI implies an ability to operate with and understand the messy, physical, unpredictable world. This is where current models completely break down. They are text-in, text-out systems with zero grounding in reality.

Consider a simple physical task: unscrewing a stripped screw. An LLM can give you a perfect list of common techniques:

• Use a rubber band for extra grip.
• Try a different sized screwdriver.
• Use pliers to grab the head.
• Cut a new slot with a Dremel tool.

But it has no concept of the feel of the screwdriver slipping, the subtle shift in torque, or the frustration. It can't improvise when the rubber band breaks or the pliers can't get a grip. It has never held a tool. This physical intuition, this embodied knowledge, is a core part of general intelligence.

This gap extends to common sense. The Winograd Schema Challenge uses sentences designed to test understanding of ambiguity. For example: “The city councilmen refused the demonstrators a permit because they feared violence.” Who feared violence? A human knows it’s the councilmen. An LLM might get it right, but it's a probabilistic guess based on similar sentences, not a true understanding of the social dynamics.

Finally, there's no continuous learning. You can’t teach a model something today and expect it to remember it tomorrow. Every API call is a fresh start, a brilliant amnesiac. A toddler learns object permanence once and has it for life. GPT-4 has to re-learn it with every single prompt.

Autonomy vs. Automation

The most useful distinction for builders is between autonomy and automation. We are very, very good at automation. We are nowhere near autonomy.

• Automation is using a tool to perform a well-defined task faster. This is the LLM sweet spot. It can write boilerplate code, summarize a meeting transcript, or draft five versions of an email. The human is the strategist; the model is the high-speed intern.
• Autonomy is the ability to set and pursue goals in a novel environment. This is the AGI promise. An autonomous agent would notice falling metrics, diagnose the cause in the codebase, write a fix, test it, and deploy it without human intervention.

Projects like Auto-GPT tried to fake autonomy by chaining LLM calls together. The results were telling. They’d get stuck in repetitive loops, misinterpret their own goals, and hallucinate progress. They were brittle because they lacked a real model of the world to act upon.

So, What Should We Build?

Don't wait for AGI. The automation tools we have right now are ridiculously powerful. The trick is to stop thinking about building a synthetic brain and start thinking about building a better tool.

Focus on workflows where the LLM can be a super-powered copilot. The model generates, suggests, and summarizes, but a human makes the critical decisions. It's not about replacing the developer, the writer, or the analyst. It’s about giving them a 10x boost.

Build a customer service tool that drafts replies based on internal documentation, but requires a human agent to click 'send'. Build a code refactoring tool that suggests changes, but needs a developer to run the tests and approve the merge. Keep the human in the loop.

The goal isn't to build a machine that thinks. It's to build a machine that helps us think better.