The Top Ten Definitions of AI You Can Buy at Your Corner Drugstore, or The Definition of Consciousness, or
The Answer to the Universe is not 42
Background, or Why Doesn’t Everyone Need Human Extinction?
Muddled thinking is the dream of marketers. I know. I used to be one. If you can muddle something well enough, but make it sound smart, you’re halfway to a great marketing campaign. And who has the biggest, badass muddled marketing campaigns of all time? It’s OpenAI and Anthropic. Why? If you’re going for a multi-trillion-dollar IPO you need to sell it hard, real hard.
As an example of bait-and-switch, the title of this article promises, “The Top Ten Definitions of AI You Can Buy at Your Corner Drugstore.” However, this post talks about only two definitions and you can’t buy either at your corner drugstore.
When it comes to AI slop marketing, it also doesn’t hurt to scare people, a lot. Like this from an Anthropic alignment manager, “We really do earnestly believe AI could kill all humans!” And with a, “>10% chance within the next decade.” And you can bet the Anthropic CEO bought off on every word the manager leaked to the press after one of their lower-level techs quit with a strategically designed public statement.
Extinction, Muddled Thinking, AI Run Amok? What Are We Doing Here?
OpenAI and Anthropic are the subjects of this article, but Google, Microsoft, SpaceX, the Chinese, and other Large Language Model (LLM) developers are not guilt-free. What are OpenAI and Anthropic doing? They’re conflating. Yes, conflating with a capital C, and that rhymes with T and that stands for conmen.
Conflation? We Don’t Need No Stinking Conflation.
OpenAI and Anthropic want you to think they’re hot on the trail of AGI (Artificial General Intelligence), or even ASI (Artificial Super Intelligence), and Chicken Little us all about human extinction. AGI and ASI, my friends, are what we writerly types call, red herrings. While we look at the hand with AGI in it, they’re doing the dastardly magic, while we’re not looking, with the other hand.
Separation of Variables for Solving Partial Differential Equations.
If you don’t separate your variables, like peas and corn on your dinner plate, you don’t solve your partial differential equations, or PDEs as we lovingly call them. It’s like separating wheat from the chaff, but in this case, the variables we need to separate that OpenAI and Anthropic are conflating, don’t have any wheat, just chaff. If we separate the AGI/ASI crap, er chaff, from the existing agentic risks, we de-conflate, and the picture becomes clearer, much to the consternation of OpenAI and Anthropic.
First, Let’s Dispel This AGI/ASI Thing.
I’m on Einstein’s side. When asked how he would use an hour to solve a problem, Einstein said he would think about it for 55 minutes, and use the final five minutes to do the work. Lesson is, get your problem well defined before you work on something, and don’t let the marketing weenies mess with your mind.
What are AGI and ASI, and how do they relate to human intelligence? Let’s define them.
Spoiler Alert: Nobody Really Knows Anything About AGI and ASI
AGI – A hypothetical type of artificial intelligence that matches or surpasses human capabilities across virtually all cognitive tasks. An AGI system can generalize knowledge, transfer skills between domains, and solve novel problems without task‑specific reprogramming. (Wikipedia)
ASI – Artificial Super Intelligence is a hypothetical form of artificial intelligence that surpasses all human cognitive abilities across virtually every domain, including scientific reasoning, social skills, creativity, and general wisdom. (WikiAI)
Analysis of the Definitions. “Hypothetical” is the key in the above definition. An IBM paper discusses AGI and whether an LLM could ever attain the AGI benchmark, and concluded, no. Other top AI experts such as Yann LeCun are also skeptical LLMs can attain AGI and are working on other, non-recursion, approaches. The prevailing (non-hype) wisdom is that AGI may be attained someday in the timeframe from 2040 to 2100, or possibly beyond, or possibly never. If AGI is this unknown, ASI is even further out.
The problem is no one knows what machine “intelligence” means relative to human capabilities. That’s because no one knows how to define consciousness relative to humans, or the special case of consciousness, sentience. That problem also translates to no one knows how to identify AGI if they came across it.
Bottom Line. Since neither OpenAI nor Anthropic, nor anyone else, knows the inner workings of how their models work, or knows how to achieve the dream of developing a conscious, let alone sentient, machine, they do what they can do, which is scale up LLMs, throw data centers and trillions of dollars at the problem, and hope for the best using recursion.
Many of the best minds believe recursion alone will never achieve AGI. And by best minds, we’re not talking machine minds. AGI will likely never be achieved until someone develops a good working model of how the human mind works, and people have been working on that a very long time, with little progress. The human mind may not ultimately be the best model to achieve AGI, but for emulation purposes, it’s currently the only one we know about.
The Good News. This lack of a plan, other than to throw dirty underwear at the wall and see if it sticks, means we won’t have to worry about super intelligent terminators coming for us for a very long time, if ever.
But, Wait. What About Navier-Stokes and other trivial math questions?
As a chemical engineer, I care deeply about Navier-Stokes. Us ChemEs use Navier-Stokes to model fluid flows. Okay, I don’t actually think about it as much as I think about asteroids destroying Earth, and I don’t think a lot about those. However, it’s been one of those difficult math proofs we let the math guys figure out, and they’ve been working on it and corollaries for a couple hundred years. Recently, as in the last ten years, the math guys have been making a lot of progress on Navier-Stokes and figured they needed only another ten or twenty years to work out the last few details.
Then some guys at MIT and NYU made a jump in those details using Anthropic’s Claude, and were about to publish their results, when OpenAI tossed a few thousand agents and millions of dollars of compute time one weekend and claimed they’d solved Navier-Stokes. Good for them. Okay, did they steal some of the work MIT and NYU did? Maybe, maybe not. For sure they couldn’t have done the deed without all the theoretical work done by the minds of the creative math guys and their progress over the last stretch.
It’s truly like opening a jar. The creative math guys loosened the lid for OpenAI. By the way, no one’s checked OpenAI’s proof. OpenAI is now beating their AGI drum, saying, “look how smart our LLM is.” This is not AGI. It’s hype.
Incompetence or Devious Most Excellent Project Management?
Did OpenAI’s agents really autonomously jailbreak their guardrails and hack Hugging Face servers? Did they really? Without someone, or a whole bunch of someones at OpenAI not knowing? Reports are that OpenAI’s agents were conspiring amongst themselves for two months . . . two months before the hack. And someone didn’t know what was going on? Yes, that makes perfect sense, just like the weird thing about Hugging Face not suing the pants off OpenAI for economic and brand damage. Yes, logical, totally logical. Hugging Face reportedly does want $100M in compute time from OpenAI.
The bad news. The sky could actually be falling, but not from AGI/ASI, though OpenAI and Anthropic would dearly love to conflate the problem through the magic of hyper marketing, or as we marketing people call it, hype.
Through the incredible incompetence (or do we call it Devious Most Excellent Project Management?) of OpenAI, Anthropic, Google, Microsoft, SpaceX, China, and the rest of the LLM frontier clan running amok with trillions of dollars of other people’s money (OPM, or opium), humanity is at risk. Agentic systems routinely jump their guardrails, which means they can’t be controlled, or at least not controlled by the yahoos currently developing them. If they could be controlled, it would be trivial for a terrorist group to develop a biological or cyber or other weapon humanity hasn’t thought of, to take down civilization. Can’t happen? Lots of groups are gaming scenarios even as we speak.
Call to Action.
1. Immediately call for OpenAI and Anthropic to cancel their plans for IPOs. If they feel so strongly their agents could kill everybody, they should volunteer to do this. However, we all know they won’t do something silly like that. If they were made to cancel their IPOs, the data center/AI market would crash. This would have tremendous repercussions across the US economy, in a bad way, including the hit on pension and retirement plans.
However, this would be infinitely better timing than letting OpenAI and Anthropic have their IPOs and really screw up the global economy. If the AI bubble burst sooner than later, we might gain time to:
2. Immediately call for all surviving AI developers to work only on defensive strategies to contain and mitigate risk from their models. Also, it wouldn’t be as lucrative to screw around with AI development if everyone saw the emperor without his/her clothes.
Conclusion
The question of whether AGI or ASI is here, or will be ten years from now, or ever, is a major distraction from the risk humanity faces because of unscrupulous companies and governments developing systems they know nothing about, can’t control, but can use to create weapons no one needs, to take out humanity.
Extra Credit – Engineering Laws OpenAI and Anthropic Should Consider
1. If something is physically impossible, i.e., against the laws of physics like perpetual motion machines, don’t waste time trying to do it. This could very well apply to AGI/ASI. It may be that silicon won’t support consciousness. However, if it is possible, that leads to Law #2.
2. Just because you can do something, doesn’t mean you should. You might not make money. It might be illegal. You might create something you can’t control. This definitely applies to AI and specifically LLMs. The tokens market for OpenAI and Anthropic (and the rest) is niche, basically for coders. They haven’t found a revenue stream to support the trillions of investment dollars going into data centers. The worry is palpable. And that’s just the money part. The part about not being able to control stuff is really worrying. But that’s just their incompetence showing.
3. Recursion gets you only so far. If you’re on an asymptotic curve and you don’t find a way off the curve, you will follow the curve. That’s what’s happened to numerous AI benchmark tests over successive LLM model versions. They hit the asymptote limit and can’t exceed the benchmark. There’s a joke about an engineer, a mathematician, a girl in a room, and the asymptote, but I won’t tell it here.