The o3 and o4-mini models can think with images, code better than you, and locate your bistro with a photo of a sandwich. Is this artificial general intelligence - AGI?
With the o3 and o4-mini, OpenAI has raised the bar for artificial intelligence to a level where we’re talking about the ability to reason, visual understand, and solve complex problems on our own. Some say we’ve achieved AGI. Others say – almost. But we’re all thinking the same thing: “Holy s***, this is crazy.”
Welcome to the age where AI understands your handwriting. Even if you write like a doctor.
On Wednesday, April 16, 2025, OpenAI launched the model o3 and o4-mini. No fanfare, no hype – just a cold blog post. And yet… something was out of the ordinary. For the first time, people – and those within OpenAI – were starting to publicly hint at AGI.
AGI. That “ultimate boss” of artificial intelligence. A system that can think, understand, learn – and, most insanely, does it all better than you. Even if you have a PhD.
o3 can repeatedly zoom and crop into images in order to read small, handwritten text
it's CRAZY pic.twitter.com/bNve0OsJLJ
— Dan Shipper 📧 (@danshipper) April 16, 2025
Thinking with images – it's not just the eyes anymore. It's the visual mind.
One of the most mind-blowing features of o3 is its ability to “think with images.” Imagine an AI getting a picture of your high school graduation notes, written with your left hand on the bus. And – effortlessly – understanding them, correcting them, explaining them, and even solving your equations.
The model can zoom in, rotate, sharpen… and then draw conclusions based on what it sees. This is no longer passive analysis – this is active visual thinking. At the human level. Or even a step further.
Math? Only 99.5% pass rate. Almost boring.
On the elite AIME 2025 test, considered the “Dark Souls” of math competitions, the o4-mini scored 99.5 %. This is no longer “AI can do math.” This is “AI is best in class.”
OpenAI's Noam Brown has warned that the models are still not masters of mathematical proofs—so they're not ready for the Fields Medal. But for us mere mortals? They are. And they have been for some time.
And now the big “if”: is this AGI?
Model trainers at OpenAI say that when testing o3, they were “very tempted to call it AGI.” Sam Altman retweets rave reviews, Tyler Cowen says “I think it’s AGI.” And Twitter… Twitter goes into ecstasy.
Is this artificial general intelligence yet? If you define AGI as something that beats 99 % humans in 99 % intellectual challenges… then, hey – maybe we’re there.
Security, hallucinations, and the fact that AI can find your favorite bistro from a picture of a plate
The o3 can not only read your notes – it can also identify a location from a picture of a window. Or lunch. And yes, quite a few people have inadvertently “self-doxed” themselves with a photo of their pasta. Great for impressive technology. A little less great for privacy.
Furthermore – and this is perhaps the biggest paradox – o3 hallucinates more than its older brother o1. More reason, more delusions? According to OpenAI’s own data, this may be due to the “optimistic guess” that comes with advanced logic.
But hey – who isn't sane when they can think in pictures?
Conclusion: If this isn't AGI, then it's its annoyingly smart little brother
o3 and o4-mini are more than just an upgrade. They are a breakthrough. Not just in benchmarks, but in everyday use. They are AI that understands, sees, reasoned, and even acts agentically – like a digital assistant that knows more about you than you do. And if that doesn't seem like it, at least something similar To AGI... then you probably don't know what you want. But AI certainly does.
