Are We Closer to AGI Than We Think?

The AI models we are using today are the dumbest models we will ever use.

And it shows.

Current LLMs often stumble over common sense problems that most humans solve intuitively - like recognizing logical fallacies or maintaining consistent reasoning across long chains of thought.

This leads to a lot of things, one of them being AIs tendency to "hallucinate" answers not based in reality. Effectively, it just "makes stuff up" if it doesn't know the answer.

But, despite these limitations, they are already transforming industries and sparking ethical debates.

So, what happens when AI overcomes these hurdles?

What happens when a model can truly think and reason?

Well, the rumor mill is churning at OpenAI where another leak has revealed Strawberry: the codename for the Q* project that I made a video on late last year.

If you zoom out and take all the leaks and public statements from the like of Sam and friends, it seems clear that reasoning is the focus of OpenAI. It's what they have been driving towards for the past couple of years because they believe this is one of the keys to AGI and beyond.

Intelligence Explosion

AI doesn't have to get "good at everything" to change the world. Once we are able to combine reasoning and logic capabilities with the vast speed and scalability of existing models, we approach the point where AI will automate AI research itself.

Getting good at that one thing will kick us up that exponential curve.

There's still going to be friction there due to the stubborn concrete-like nature of reality: the fact that it takes time to pour concrete to build all the datacenters needed to power the compute the world will all of a sudden be clamoring for.