OK, I’ll admit – I tend to click on some AI articles offered by my MSN.com feed – enough to ensure that about 30% of all the MSN-offered articles are AI-related.
And my Themesis email inbox? (The one that I use as a grab-all for everybody’s EVERYTHING?) That’s full to the point of toppling over, every single week, with AI blogposts.
And typically, these are GOOD blogposts.
And all of this is really at the level of “surface debris.”
You know, the kinds of lightweight sticks and leaves that float on the surface of the water, as it churns its way merrily towards the nearest drain.
The stuff that is here today, gone tomorrow.
After all – we might want to know how to write a good prompt, but – will prompts even exist 2 1/2 years from now? (Let’s take our “initiation point” as the first ChatGPT release, circa November, 2022, a little more than 2 1/2 years ago from writing THIS blogpost.)
Will the trade-offs between one gen-AI system vs. another be an issue even SIX months from now?
This means that the REAL question – for you as an AI practitioner, and maybe even as the leader of an AI company, or an AI scientist / developer (“AI Magician”) seeking to create the NEXT LEVEL in AI – IS: where to put your attention?
Even more precisely, where is your true north?
Meaning, how do you orient yourself? You’re establishing an “attention framework” so that you automatically choose the BEST avenue for that most limited and valuable resource that you own – your time.
Transformers – The Problems Are Not Going Away
Likely, you already KNOW the transformer-related issues.
But THIS blogpost (by Logan Thornloe) does a good job:
And on the same theme we have, from Apple researchers:
- THIS is good: “Apple Just Pulled the Plug …” and
- Here’s the Apple article that they reference: https://machinelearning.apple.com/research/illusion-of-thinking
So let’s move on — to agents.
Agent-Based Systems Are Not the Answer, Either
And agents … how about agents?
Despite the hype and hoopla, agents are fairly simple, primitive creatures.
They go out, perform a task, come back. Maybe they trigger the next agent, and so forth.
Agents are typically either rules-based or reinforcement learning-based.
Simple Rules-Based Systems Are Not the Answer
So … rules-based systems?
The last time that rules-based systems were really, REALLY popular was circa the 1980’s, when they were in the form of expert systems. And it was precisely the PROBLEMS with expert systems that made everyone so excited when neural networks first emerged, because expert systems were fragile and breakable, very NON-adaptive, and couldn’t handle much complexity. (Not without getting MORE fragile and breakable.)
So – the emergence of something that did not have to be pre-specified as an “if/then” was exciting. Novel. Earth-shaking.
And that lasted about ten years – from 1987 to about 1997, when the problems with (then SIMPLE) neural networks were fully revealed … and does this sound like an old story?
Similarly, reinforcement learning. Very goal-directed. NOT adaptive to mapping out a new environment, or to doing ANYTHING other than achieving their “Prime Directive.”
So What Do We Do?
The first thing to do – in terms of determining our focus – is to identify the LEVEL of the game that we’re playing.
There are three basic levels:
- Skilled Tradesperson,
- Engineer, and
- Magician.
For EACH of these, our focus is different.
If you’re a skilled tradesperson, your focus is on using tools (the techniques and tools of AI & ML) to solve an immediate, existing problem that CAN be solved by intelligent, knowledgeable application of your tools and your expertise.
The focus is: solve an existing problem, RIGHT NOW, with what you’ve got on hand.
If you’re an engineer, you are more likely to DESIGN a system – but again, using the tools and techniques at hand.
And if you’re a magician, you’ll create “something out of nothing.”
Your Role Dictates Your Focus
One of the most useful ways to sharpen your focus is to be very clear and honest with yourself about your role, and about your aspirations WITHIN your role.
Skilled Tradesperson
If you’re a skilled tradesperson, then knowing WHICH transformer-based system to use is an IMPERATIVE. Whether you’re doing text-based processing, image-based, creating ANYTHING using any of the commercially-available transformer systems, all the current industry buzz is actually CENTRAL to your work.
Similarly, the current set of “tradecraft” skills – crafting good prompts, crafting a useful set of agents – all this is extremely pertinent.
The only downfall is – all these skills have a short (sometimes VERY short) half-life.
It may be that 2 1/2 years from now, “prompts” will have gone the way of the Model T Ford. (And so on, for the rest of the currently most-popular trades skills.)
The upshot is: being a skilled tradesperson means that you’ll LIKELY always have a job – in the sense that there is ALWAYS a need for plumbers, electricians, and the like.
BUT, you will ALWAYS have to upgrade your skill set; what you’d “bring to the job” even a few months from now could be RADICALLY DIFFERENT from what you’ve got in your toolbelt right now.
Engineer
I love engineers.
Meaning, I totally ADORE engineers.
Engineers keep the Universe running. They solve problems.
Engineers are practical, down-to-earth, solutions-oriented folks.
If you’re an engineer, then you need a somewhat more elevated toolset, and a wider (and more in-depth) scope of reading.
It will help to understand a BIT MORE OF THE FUNDAMENTALS as you design the solutions to your problems (or your client’s problems).
As an example: you might be designing and building a RAG (retrieval-augmented generation) system on top of a client-specific, domain-specific knowledge base or corpus.
It will help you, then, to understand that a RAG uses cosine similarity. That means, the vectors that you create for your RAG will function a LOT LIKE the vectors that would define different clusters if you were using a k-means clustering algorithm on a corpus.
Knowing that, you’d be very thoughtful about everything that contributes to effective k-means (cosine similarity-based) separation and also how those vectors REPRESENT your corpus. Lots of tech details would follow, and we won’t go into them here … this is just an example of when a little algorithmic depth would be highly appropriate.
Magician
There are far more skilled tradespeople than there are skilled engineers. And similarly, there are far more skilled engineers than there are true “magicians” – those who innovate radically new concepts and methods.
Magicians are relatively rare.
Some of my favorite “magicians” are:
- Geoffrey Hinton (invented generative AI),
- Paul Werbos (invented backpropagation), and
- Karl Friston (invented active inference).
I’m a magician – much more so than anything else.
A magician’s study will necessarily be far more conceptual (often more mathematical, or use more physics, etc.), and far more complex than what tradespeople and engineers will need.
But to get a sense of how to IDENTIFY when someone has “created magic,” look at THIS YOUTUBE:
For now, Socrates’ adage applies: “Know thyself.”
Whatever you are is what you are.
All are good.
Be well! (And “live long and prosper”!) – AJM
Resources & References
The following arXiv paper is the one I mentioned in the vid – it’s my 70++ “tutorial” on Karl Friston’s core equations in his 2013/2015 papers:
- Maren, Alianna J. 2024. “Derivation of Variational Bayes.” arXiv:1906.08804v6 [cs.NE]. (Rev 6: 18 Aug 2024). (Accessed July 23, 2025; available at arXiv link.)