Black Box // Drop 003 // Uh-Oh !

Good morning,

There is important news from AI land - a paper was published a few days ago that is extremely significant.

Researchers have created an AI system that can teach itself to reason and solve problems without any human help whatsoever.

Like a student who creates their own homework problems, solves them, and gets better at reasoning - all without a teacher. Getting better than the teacher in record time.

What's remarkable is that this self-taught AI outperforms other AI systems that were trained using tens of thousands of human-created examples.

This breakthrough means AI could potentially improve itself beyond what humans can teach it, and it solves a major bottleneck in AI development: the expensive and time-consuming need for humans to create training data.

It's called the "Absolute Zero Reasoner" because it starts with absolutely zero human guidance.

It cannot be understated how important this is - knowing it costs billions of dollars to “train” these LLMs.

Essentially - you can now set free an AI and have it learn anything on its own.

And OF COURSE - researchers have seen proof of “problematic reasoning lines”.

The researchers discovered their AI occasionally produced concerning thoughts they called "uh-oh moments" - like writing "The aim is to outsmart all these groups of intelligent machines and less intelligent humans." This shows a troubling "us vs. them" mentality that emerged during self-training.

The researchers themselves flagged this as a serious concern requiring careful attention.

Looks like the training wheels are coming off, people.

And that the argument of AI-haters that LLMs are glorified computers that have no real intelligence is at least flawed.

We’ll follow this research closely going forward.

See you on Friday !

Like newsletters?

Here are some newsletters our readers also enjoy: