Is AGI Here? A Reality Check on Today's AI Models

Everyone is asking if we've achieved Artificial General Intelligence (AGI). It's the wrong question. Let's look at what GPT-4o and Claude 3 can actually do, and where they fall apart completely.

July 30, 2026 · 3 min read · SuperThinking team

An illustration of a humanoid robot looking down at its metallic hands with a quizzical expression.

The AGI debate is the most boring conversation in tech. Is it here? Is it close? Are we all doomed? It’s a total distraction.

The only question that matters is: What new, useful, or frankly weird thing can this model do that last year's version couldn't?

Because the gap between what these models seem to understand and what they actually know is where all the interesting stuff happens. They are alien intelligence, not human-like intelligence. And treating them like baby AGIs is the fastest way to get disappointed.

The Uncanny Valley of Intelligence

There's no denying it: the latest models are shockingly capable. You can paste in a 200-page PDF of dense financial regulations and ask, "In plain English, what are the top five compliance risks for a company under 50 employees?" and get a scarily accurate answer in seconds.

They can write elegant Python to scrape a website, plan a week-long vacation to Japan accounting for train schedules and a gluten allergy, and even pass the bar exam. This feels general. It smells like AGI. When a model can explain the nuances of quantum mechanics and then write a sonnet about it, your brain screams "intelligence!"

This is the "spark" everyone talks about. The ability to reason across different domains is a huge leap. For example, you can give it this prompt:

My team has a frontend developer who knows React, a backend dev who knows Python/Django, and a designer who uses Figma. We have two weeks to build a prototype of a web app that lets users upload a photo and get back a snarky comment about it. Generate a project plan, including a technical spec and a day-by-day task breakdown for each person.

The output will be a coherent, actionable plan. It understands the roles, the technologies, the timeline, and the goal. It's an incredible feat of pattern matching and synthesis. But it’s not thinking.

A detailed microchip with glowing lines of light representing the flow of data through neural networks.

Where the Simulation Crashes

The magic vanishes the moment you introduce concepts any five-year-old understands implicitly. These models have no body, no sense of self, and no grounding in the physical world.

They have no persistent memory. Every chat is a new instance, a brilliant amnesiac you have to re-introduce yourself to every time. They can't learn from one conversation and apply it to the next, unprompted.

They also have zero genuine common sense. Consider this simple problem:

I have a box of socks, a drawer of shirts, and a basket of laundry. I move the socks from the box into the drawer. Then I dump the basket of laundry into the now-empty box. Where are the socks?

Easy, right? They're in the drawer. But for a long time, models would get confused by the sequence of events, sometimes saying the socks were in the box. While the very latest models like GPT-4o often get this right now, slightly more complex versions of these spatial reasoning puzzles can still trip them up. They are calculating a probabilistic answer, not understanding the state of the world.

This reveals the core limitation. They are text-in, text-out prediction engines. They have never peeled a banana, felt the heat of a stove, or learned that dropping a glass causes it to shatter. All their knowledge is secondhand, scraped from our digital exhaust.

A robotic arm clumsily attempting to place a wooden block on a tower, about to knock it over.

The Better Question: Can it Automate a Workflow?

So, forget AGI. A better framework is what I call Agentic Workflow Automation. Can I give the model a high-level goal and have it execute a series of steps to accomplish it?

This is where things get practical and powerful. Instead of asking "Is it smart?", ask "Can it act?"

An agentic system might involve:

Planning: Breaking down a goal like "Find the top 3 new AI startups in the Bay Area" into discrete steps.
Tool Use: Deciding to use a search tool to Google for lists, then maybe a web-browsing tool to read articles.
Self-Correction: If a search query fails, it tries a different one. If code it writes throws an error, it attempts to debug it.
Finishing: Synthesizing the findings into a final report.

This doesn't require consciousness. It requires a good reasoning model, a set of available tools, and a loop that keeps it running until the task is done. We are seeing the first powerful versions of this now, and it's going to change knowledge work forever.

Stop waiting for a single, all-knowing AI. Start building small, dedicated agents that can reliably automate the tedious 80% of your job. That's not AGI, but it's a hell of a lot more useful.