11 Comments
User's avatar
Jethro Hedstrom's avatar

Big fan of your writing on AI Ben, please keep it up.

Thomas DeWitt's avatar

I guess I am skeptical about definitions that are relative to human ability, whether expert or randomly selected, because human ability varies by orders of magnitude based on which tools you allow and how much time is given, and this seems arbitrary. A human expert + pencil and paper is far more capable in math than a human expert without. Give them mathematica and their ability grows again by a substantial amount.

Marcus Seldon's avatar

This is a good point. Really there are many ways we can compare AI to human ability in a domain:

* Better than a human with no tools

* Better than a human with non-LLM tools

* Better than a human with tools AND access to an LLM ("centaurs")

Eric Ngoiya's avatar

After years of applied research and engineering, we present Artificial General Intelligence with Earth.

"A system describing the interoperation of increasingly heterogenous data points without compression"

Demonstration. https://youtu.be/_uAyox15xNQ?si=RdTJw4iwJW5XStoI

Visit our website at https://www.queue.cx to become our early customers and supporters. Connect with our community and follow us on https://www.x.com/tryqcx

Immanuel Giulea's avatar

Your note was written on March 22. GPT-5.4 was already released. GPT-5.3 was never released to ChatGPT.

Benjamin Todd's avatar

I’m not sure I follow. Are you saying GPT-5.4 would already hit the definition of AGI? I address that in the post.

Gergely Máté's avatar

Here's my take about the same process of becoming economically obsolete, from a human's perspective: https://www.lesswrong.com/posts/SwpYdEu4rEXwhsxGA/let-s-reason-about-your-job-security

But given the likely somewhat step-function nature of what you describe in AI development (the milestones), my method probably needs further refinement.

Kris Lyser's avatar

You're right, we haven't reached AGI. But the reason is fundamental.. qualitative rather quantitative. We are measuring software milestones while ignoring the biological hardware reality. Speaking of more precise terminology: LLMs are an "externalized cerebellum", rather than a modeled, complete brain. If that's understood, then it becomes obvious why current AGI attempts have sticky failure properties. We're trying to replace brain function by scaling the function of but one of its minor structures.

Tim's avatar

>Only 70% of AI researchers know what AGI stands for

The tweet you cite is just Leo Gao walking around and asking random people at NeurIPS.

Probably more accurate to say 30% of attendees of NeurIPS are either a) not AI researchers, or b) just messing with Leo.

Benjamin Todd's avatar

I’ve edited to “people at the biggest AI conference said”.