While I agree with your general assessment that true AGI is a ways off. I think your view of what is currently possible is off kilter though.

Let's look at DALL-E 2 at first glance it just a painting/art AI. But its ability to recognize different objects and more importantly combine multiple objects into a coherent design is pretty staggering. It's not the one trick recognition pony you are talking about. If allowed to expand into non art work I am sure it would quickly identify the important data/objects.

The Chinchilla AI on the other hand shows us the you can get better results with an order of magnitude fewer data points if the AI is full/correctly trained. This and GPT3 already outperform most high school students at writing tasks.

Okay, last thing. It doesn't taking AGI to devastate the population just 10x better/cheaper than a human would be in the economy and we are awfully close to that already.