AGI is Quicksand

In the 1960's, almost 3% of all films showed characters sinking in clay, mud, or sand (according to Wikipedia). Quicksand, in other words.

Quicksand requires only 3 ingredients: ordinary beach sand, water, and a place where the water can fully saturate the sand and not drain away. The planet we live on has at least 390,000 miles of coastline, and that doesn't include the shorelines of lakes and rivers, which can also provide quicksand's 3 ingredients. So quicksand and quicksand-related harm must be extremely common, right?

Nope.

Quicksand exists, but it's not extremely common. And quicksand-related harm is very uncommon. I'd link to some authoritative stats on this if such stats existed. They don't. Probably because it's so uncommon.

The reason we have 390,000 miles of coastline and we do not have 390,000 miles of deadly quicksand (and loads of quicksand-related deaths) is because the conditions that create quicksand are special and unique. The first 2 ingredients -- sand and water -- are extremely common, especially on coastlines. But the 3rd ingredient -- the context in which water and sand form quicksand -- is not very common. The 3 do come together! But only in a small fraction of the situations where water and sand are both present.

Quicksand is the special case of water and sand mixing, not the general case.

The Discourse About AI seems to have decided that where AI is inveitably headed -- and what we need to be Very Worried About -- is either a place called Artificial General Intelligence, or Artificial SuperIntelligence. The Discourse About AI is made up of people smarter, more successful, harder working, and better looking than me and yet, they're wrong.

The smarter/more successful/harder working/better looking folks are seeing a few special cases where mixing AI and jobs that need doing --> automating the job and they are making a big leap to believe that in 1 to 5 years AI will replace half of all entry-level white-collar jobs. That's nuts, and if you've ever looked closely at what it actually takes to produce economic value in the real world, I think you'll agree with me.

Some work can be fully automated with current state of the art AI. But God help you if you hire the healthcare.gov team to design and build the system to automate it! Some tasks can be executed by AI with variable accuracy and consistency. But how many humans have to be in the loop to ensure enough accuracy and consistency?

Maybe somewhere around 800 million people now use AI. That indeed represents a shocking and unprecedented adoption curve. (It took the 15 years from 1989 to 2004 for the admittedly much more real-world-infrastructure-intensive internet to achieve the same zero-to-800-million diffusion) And so we're all right to look at that and say... holy cow, there's definitely something going on here!!

And it's understandable to look at that zero-to-800-million user explosion over the course of 2.5 years and make some similarly... explosive inferences about what AI will be able to do in the future. Here's the thing, though. Not all users are power users, and special ability is not general ability.

Paying $20 or $200/month to get access to a research or coding assistant that has to be closely supervised and frequently corrected is not automation that can do everything a human can in the workplace in both normal and exceptional situations.

It's so much more helpful to call things what they actually are. It helps curb the magical thinking. AI is non-deterministic automation. That's both its superpower and its achilles heel. Because AI is "fuzzy" in its logic, it is able to encounter novel situations and respond to them in useful ways. This is genuine, 200-proof, un-cut dealer-grade magic. And also, there are so many places in life where we will not touch a non-deterministic system with a 10-foot pole.

I find it exhausting to try to argue why special cases don't quickly or easily become the general case. It just seems so obvious to me that they don't, and I'm baffled by people who don't seem to understand this aspect of reality. It seems obvious that applying deterministic automation to messy or complex systems is hard enough that the default choice is to not automate them and instead use humans. Applying non-deterministic automation to those same messy or complex systems?

Where AI is actually headed is Artificial Specialized Intelligence. And maybe a few people should be worried about that, but not society as a whole.

Artificial Specialized Intelligence is building AI into systems where non-deterministic automation produces acceptable variance and quality and where the cost of building and maintaining the system either has nearly immediate ROI or there is a wealthy sponsor who can mortgage the cost over a longer timeframe and many idential-enough use cases (think Amazon and AWS or the SaaS infrastructure ecosystem writ large under ZIRP). ASI (not the highly unlikely fantasty one, the real one I'm talking about here) will have an incremental positive effect. Gradual transformation; not tragedy.

But why then is the marketing world in particular so freaked out about AI?

Because marketing as a whole has become systemetized and simplified enough for LLMs to do a lot of the work that constitutes modern marketing. A lot of that work is using GUIs to drive APIs. In other words, modern marketing is full of systems where non-deterministic automation produces acceptable variance and quality and where the cost of building and maintaining the system has nearly immediate ROI.

The unkind way to say this is: marketing is full of people with expensive college loans doing work that is beneath their potential.