Seeing all kinds of vastly different statistical models being bundled under the umbrella term of “AI” is as frustrating as witnessing anything that is not a relational database with SQL being called “NoSQL”[1]. At this level of generality, the term loses enough meaning to sabotage almost any conversation about it, while promoting sensationalized pieces that induce more anxiety than understanding.

“Is AI going to kill us all?” Well, a linear regression in an Excel spreadsheet probably won’t do that, unless what is implied by its coefficients causes a heart attack.

But what about an agent model that can make decisions and use an LLM to interact with the outside world? What if it is embedded in a robot body and can use another neural network to control it? What if there are millions of them, silently exchanging messages and conspiring about how to turn the Solar System into a cloud of grey nanogoo[2]?

We don’t really know. And we won’t know until we try building and running systems like that for a prolonged period.

OK, but “will LLMs evolve into AGI?[3]”. No idea. What is “AGI”? We have not even managed to agree on a workable definition so far. Should it be human-like or just general? Should it be more general than an average human? Or more general than the sum of all humans and therefore not human anymore?

By the way, it is called “ASI”[4] now. Keep up, please!

In this chaotic conversational landscape it is far too easy to be overinfluenced by the merchants of hype and the prophets of doom alike. Instead, we need more precision for the masses. Definitions that are understandable, with clear boundaries. Articles and videos popularising the discipline with careful communication about which claims are warranted and which not, and under which assumptions.

One such video, by Jodie Burchell, PhD, is embedded below. It is a presentation about the strengths and limitations of LLMs, recorded at GOTO Amsterdam 2024. It is clear, precise, informative and digestible to the general audience.

I recommend it through and through.

  1. “NoSQL” may mean a document store, a graph triple store, a key-value store, a vector database etc. Some of them will implement a (limited) variant of SQL anyway, while not being relational in the sense PostgreSQL or MySQL are. ↩︎

  2. Wikipedia has a good article about so called “grey goo”. ↩︎

  3. “Artificial General Intelligence” ↩︎

  4. “Artificial Super Intelligence” ↩︎