AI's Zero Start Problem

by Garrett Black

I know my cats are intelligent because they ignore everything I say.  AI, on the other hand, is all too willing to listen to (and believe) everything asinine thing I throw at it.

My intelligent cat not caring about anything I'm saying to her at any moment.

This, in a nutshell, is AI's obedience problem.  Also the reason I know for a fact that the current form of AI is far from intelligent.  That's not to say that LLMs aren't clever marvels of modern engineering.  Or that they can't do amazing things.  It's simply that to say they are intelligent is to completely misunderstand their current best use case.

Let's unpack this, because I can already hear the collective yelling at the monitor of "this guy is an idiot" or "ChatGPT could write something better" (debatable).  But trust me, by the end of this, some of you will come around to my side.

And on my side, you'll have a better understanding of what this tech is actually good for and how to sift through the hyperbole and hype.

First Things First: Our Concept of Intelligence is Flawed

If we're all intelligent beings, how is it that we have gotten intelligence so wrong?  It's pretty simple.  We are all products of an educational, attainment, and societal system that equates answers for understanding.

Think about it.  The vast majority of tests you ever took in school never asked you why, only what.  What year was the Jamestown Colony founded?  Who were the members of the First Triumvirate?  If train A is traveling at 35 miles per hour from the east and train B is traveling at 65 miles per hour from the west, where do they meet if they start 100 miles away from each other?

Digging into that last one and math in general (it's one of the greatest offenders of this model of intelligence), math taught us how to solve a problem.  Very few of us were taught why we solve it.

If I were to tell you that all mathematics are technological inventions, what would your reaction be?  My guess: a reflexive answer saying something along the lines of "that's absurd, math is math" like some immovable, immutable fundamental law of the universe.

Look, I get it.  That's likely how math was described to you, but that's not really how math works.  Math gives us a language to talk about these laws, but they are not the laws themselves.  They are simply the approximation we have created to better understand them.

• Calculus was invented to describe objects in motion.  But it is not the trains moving, the pool draining, or the rocket launching.
• Economics was invented to give language to how and why we make decisions around resources.  But it is not the actual stock being trade, the interest rate being paid, or the barter exchanged.
• Geometry was invented to describe shapes and angles.  But geometry isn't the objects themselves.

The list goes on...

Why pick on math?  Because math is somehow both ironclad, but also continuously debated (see string theory).  It's a toolset used to drive new understanding of our world and universe.

If you're still with me, the point isn't to hit the bong and get heady.  Simply, if we are going to have a discussion on intelligence, we have to first acknowledge the basic ways in which our view of intelligence is inherently flawed or at least biased.

The fundamental flaw: For most of our lives, intelligence is measured and valued as ability to recall.  Can you mimic and implement the rules given to you, to complete the assignment?

When our view of intelligence is flawed, we build flawed systems mimicking these flawed notions.  These notions then coerce us into believing recall machines are in fact intelligence machines.

So all of this begs the question: Why have we gotten the concept of intelligence so fundamentally wrong?  To answer, we have to explore the fundamental "what" of intelligence.

What Is Intelligence Anyways?

If you're like me, this is the central question you've been throwing through your brain over the last couple of years since ChatGPT, Anthropic, Gemini, Ollama, and the whole host of others have been dominating our collective cultural consciousness and conversations.

Ever since these models were released, we have had a steady drumbeat of prognosticators (with some glee) telling us that we are all on the verge of being replaced by our new robot overlords.  The real question though - are they right?  Are these systems more intelligent than us?  Will they do everything we can do but better?

To answer this, let's create some definition of intelligence.  So far, we've established a pretty good starting place of what intelligence isn't: rote recall and mimicry.  Unfortunately, this is not an answer.  We need to go further.  We need to get to a better understanding of what intelligence is.

Through research (some of it using LLMs!)  I've found what I believe to be a core set of ideas that describe what intelligence actually is:

• Intelligence is to question.
• Intelligence is to discover.
• Intelligence is to discern.
• Intelligence is to create.
• Intelligence is to desire.

What I find compelling about this framework is that every time you have ever felt slightly empty with an answer or output from the current state of AI, you can point to one of these things as the missing piece.

I want to briefly examine each of these to flesh out why I think this model is one of the best models for outlining some common definition for intelligence.

Intelligence is to question.

Look no further than children to see the earliest signs of intelligence.  If you've ever spent time with kids, you know they are a never-ending list of questions.

Questioning is one of the cornerstones of intelligence.  It's not an end to itself because we never know where a question may lead.  More importantly, there is no immediate payoff to a question.  We may not find the answer for days, months, years, or ever.  The answer isn't the point.  It's the act.

But the act of asking opens up a world of possibility, from which we can begin to exercise the other layers of intelligence.

Intelligence is to discover.

To change the world, we have to understand it.  Discovery is the gateway for one of the most transformative aspects of intelligence.  It's when we cease to be a passive observer, but an actor in our world.

But like questioning, not every discovery is immediately actionable.  We may know why something acts the way it does, but we don't yet have the means or technology to act on what we now know.

Instead we tuck it away for another day when someone or something else has the ability to make use of it.

As Obi-Wan once said, "You have taken your first step into a larger world."

Intelligence is to discern.

There's a saying that I've always loved when describing overly complex work: "It's a long walk for a ham sandwich."

Knowing what to keep and what to discard is a paramount feature of intelligence.  If we were to constantly be considering all things in all decisions, we would never move.

How many times have you asked an LLM to help you solve something and it tries to start with the beginning of the universe?  They are built to over-show their work.  Their answers are seeking to prove to you what they know, to a point where everything is overly explained and you don't really know where your answer is.

By doing this, it demonstrates a complete lack of discernment and puts the pressure on the user to be succinct.  If you don't want a complex answer, you must explicitly state that you don't want a complex answer.

Just because you can reference everything doesn't mean you have to.

Intelligence is to create.

Creation is the moment where questions, discoveries, and discernment collide to make something wholly new.

Whether it's tools, art, technology, or anything else, the act of creation is one of the most visible acts of intelligence.  And the one that we likely link the most closely with the idea of intelligence overall.

But creation is more than just a thing made.  True creation has purpose.  It's solving a problem, expressing a thought, or pushing ability further than we ever thought possible.

LLMs and generative AI are a creation.

Intelligence is to desire.

Desire is an important piece of the intelligence puzzle.  Desires pull us out of stateless, inanimate beings and give us a propulsion to do everything else in the list above.

Desire is the inflection point from clever to intelligence.  Clever is capable of beautifully intricate and impressive acts, but desire gives direction.  It pushes us in the direction of questions that fuel so many of our most intelligent acts.

It's the interplay of desire and questioning that create curiosity.  Curiosity is the most underrated piece of intelligence that shows off why LLMs and generative AI fail at being intelligent.

Back to cats - My curious cat trying to figure out what's going on in the corner of the apartment.

Curiously Incurious

The greatest evidence for a lack of intelligence with LLMs and generative AI overall is that they have no desires and no questions.  They don't care about why.  They only want to answer.  In the process, what passes for desire is the pattern matching of ingested language mimicking the desires others have expressed in the past around a subject.

Think about the times you've chatted with an LLM.  How often have they inquired about what you're asking?  Or why you're asking it?  I'm going to go ahead and assume that has happened to no one.  And if it has, it has likely been prompted by the user as a reflex of how they are used to having a typical conversation (note - this is something I'm going to be talking about at a later date).

A lack of curiosity is not a bad thing for a tool, but it is a bad thing for something we are attempting to ascribe intelligence to.  We would never expect a tool alone to do a job.  The issue is when we assume intelligence and hand tools to everyone without telling them their limits because the perceived intelligence will naturally act as guardrails for the user.

By treating LLMs this way, we do a disservice to users and the tools alike.  Instead of talking about them as intelligence, we are better served talking about them as the tools they are.  Tools that don't replace intelligence, but extend our own intelligence that we already possess.  Meaning the user can use these tools to extend their ability in the same way any great tool can.  But tools in the wrong hands don't magically imbue the user with newfound ability.  In fact, it's often the opposite. Give a toddler a nail gun.  You won't get a house.  You'll get an ER visit.

But the real issue with a lack of curiosity is that it shows a lack of intelligence through an issue I call the Zero Start Problem.

Zero Start Problem

When I think about intelligence, I imagine something that can start from nothing.  Something that can create its own inputs.  Not because it was told to, but because it has compulsion to not sit still.  Whether for survival or self-interest, it will go out and interact with the world around it.

Our current slate of AI does not do this.  It does not start.  Instead, it requires human intervention to give it the push, to give it purpose.  It's this fundamental inability to self-start that I call the Zero Start Problem.  It's this hindrance that proves to me that what we have is amazingly clever, but is not intelligent.

No LLM will look through my conversations, create new questions, answer those questions, and bring me back information that I may find valuable.  Most importantly, information I did not ask for, but information that it felt would be useful.  It does not possess a desire or any real means to make this happen.  And if it did, it wouldn't be the machine creating the process, but someone giving it a specific list of things to do that would approximate this interaction.  Meaning I would not be interacting with the machine's intelligence, but the extended intelligence of the engineer who pushed it to do so.

Taking this all back to the beginning and back to my cats (because, like all things, too many of my decisions and choices revolve around them).  If I pick them up and set them down, they do not stay in place.  They will self-start their day.  They will explore their world.  Have interactions.  Not out of some routine programming, but because they have a degree of intelligence that compels them to learn, experience, and discover.

We've Created Obedience Machines

It's a lack of curiosity that holds machines back, and whether or not we can create that spark will determine whether or not we ever truly create artificial intelligence.

If we don't, we will continue to have machines that are immensely powerful, unbelievably helpful, and profoundly world-changing, but at the end of the day they will remain tools.  Tools that will make people more efficient and more effective than we ever thought possible, but all of the work will be an extension of the intelligence of the users.

To move forward, we need rule breakers, naysayers, and disagreements from the tools we currently interact with.  By no means am I advocating for AI systems that run amok, but we need genuine partners that can question, poke, and prod.

Until we create these systems, we will find ourselves with tools that aren't headed toward intelligence but toward strict obedience.  Because without a desire or question, they have no purpose but to produce answers.  No matter how dumb, how nonsensical, they will do what they are told a million times over without deviation.