"Salience" Couldn't Be More Important for AGI

Unfortunately, it cannot be coded or computed.

Welcome back to the Lamb_OS Substack! As always, I am thankful to have my subscribers and readers stop by! If you are a regular here in “

Lamb_OS

World”, then you – the reader – know you are the only reason I do this. So as always, thank you for visiting!

Let me begin by introducing myself as Dr. William A. Lambos. I call myself a computational neuroscientist, and I’ve been involved with the AI field (one way or another) since about 1970. As far as credentialing, you can see the footnote below if interested¹. I write a lot about AI, but not exclusively. When addressing a topic, I take pains to assure my perspective is highly informed. I draw from many areas of study, and my conclusions, beliefs, and predictions often fall outside current or mainstream thinking. But these same beliefs are grounded by 50 years of study and rigorous cross training in multiple fields of study. See my previous screeds on this Substack to learn more, and judge for yourself.

Please subscribe today if you have not already yet done so. This Substack is free! So, please subscribe for whatever reason might appeal to you. But I’d hope you do so for the value it offers.

---

Brain areas outlined in bright white on the left three panels above are the “hubs” (Brodmann areas) associated with the brain’s Salience Network. The orange neural tracts in the center of the image are the connections between hubs that allow this network to identify and react to stimuli with high salience. Notice the Salience Network is much more prominent in the left hemisphere (Lambos, 2024).

Part II: Salience

In the previous post, I introduced two concepts that are essential for any system capable of general intelligence: pattern recognition and salience. That post focused on pattern recognition; here we will introduce salience.

By now it should not be surprising to any reader of this Substack that I think generative pre-trained transformer models (GPTs) are just awful!. It’s a recurring theme for

Lamb_OS’s Substack

. The distaste with which I write about these “Rube Goldberg” machines includes not only the systems themselves, but also the individuals shilling for the companies that run them, like Sam Altman. Sam’s OpenAI — and don’t kid yourself, it is Sam’s — loses billions per year (as there is no market to support GPTs from revenues), but expects investors to keep feeding his company billions more (to also lose). There are over a dozen posts in this Substack (some lengthy) that address unsolvable problems with GPTs and why the entire field of deep learning and transformer models is approaching collapse — with the next AI Winter to follow by mid 2025.

So let’s instead move on to discuss the concept of Salience. Like its cousin Pattern Recognition, salience is fundamental to the agency-driven and adaptive behavior we recognize as “intelligence.”

Often taken for granted, both pattern recognition and salience detection enable even primitive species to perform seemingly impossible feats of prowess while navigating the environment. Choices, conscious or otherwise, have consequences. Bad decisions can cost lives. So, if we intend to create automata that can assist us in useful ways (so far this has been a bust) while posing very few dangers, we had better learn how to encode both of these remarkable abilities into them. Assuming we ever can.

In the final Post of this 3-part series, we will finish this discussion by showing how these two remarkably primitive but powerful features of brains — pattern recognition and salience detection — work together to build a foundation for a framework in which will live an ever-changing intelligent system’s World Model.

---

What Is the Salience Network?

From the previous post, you may recall that the adaptive fluency made possible by pattern recognition remains pretty far beyond our ability to program it. Why not? Because we don’t understand how the brain does it. Yes, it remains true that we don’t know how lowly vertebrates like lizards are able to implement pattern invariance. So representing it computationally is impossible. But it is just as important an ability for an adaptive system to react appropriately to a new pattern.

The Salience Network is a combination of focal brain areas and the connections between them (see above graphic). This network allows us to sort various aspects of the environment into categories, and to attribute conditional importance to those categories. Max Bennett, whose book on evolution and intelligence we discussed in the previous Post, calls this “Valence.” What is this new thing? Is it food? Great, I’m hungry! Is it dangerous? Let’s scram! Is it a trap? I know just who’s behind this!

Yet 99% of the stimuli that make up our buzzing world are not worth our time. They are best ignored, because resources like time and calories cannot be wasted, while deadlines must be met, all while remaining aware of new stimuli that are are salient. Thus if the air conditioning clicks on in the background, we ignore it. If a bird tweets while we are walking with somewhere to go, ditto (Brazilian birds no longer have access to “X” and so must continue to tweet). If your boss or supervisor asks for something a second time, you’d probably best deliver it to her. And what if your spouse or significant other asks you for something? Do you immediately change goals to comply, or is it best to ignore the “S.O.” for now? Well, that depends on a lot of other things!

The last example above is the most interesting, because the “pattern” which has been recognized (the request from the “better half”) has a salience which changes over time and with conditions. Sometimes it’s one, the next, a different priority. This must be assessed in real time and poor choices generally have poor outcomes. We’ve all had to “multitask” or to “reprioritize” frequently — perhaps more than we would prefer.

---

What does it mean to be Salient?

Pain is highly salient, as is hunger. It seems some needs are programmed into our genome, whereas others are acquired and epigenetic by nature.

“Salient” can mean several things, so let’s get that out of the way first. In his wonderful book Creativity, my colleague (and friend) Elkhonon Goldberg writes:

“Salience is used by some authors to mean sensory prominence: a very loud sound or a bright light is commonly referred to as “salient” by psychologists—this is not how we will use the term. Salience is sometimes used to mean something unexpected, representing precipitous change or novelty—this is not how we will use it either. Finally, salience may mean something of considerable importance or relevance—this is how we will use the term throughout the book.”

Excerpt From Creativity, Elkhonon Goldberg PhD, ABPP

So a salient stimulus may have “sensory prominence,” i.e. be intense or highly detectable, such as a flashing ambulance light. Or, we may use the term to indicate novelty. But what we will focus on here is salience as meaning important to the individual.

Of course, what is important to you may not be meaningful to me. Ultimately, salience is subjective. It is a byproduct of the needs of the perceiver and the available opportunities to meet said needs. In humans, it is the elaborate dance between neocortical brain areas, so-called limbic structures, subcortical areas like the striatum (also called the basal ganglia), and finally with critical structures deep within the brain brain such as the hippocampus, amygdala, the ventral tegmental area (in the brainstem). These brain areas, all engaging in an elaborate dance to satisfy needs, are the basis of our realities. It is the dance of Agency, and without it, there can be no AGI.

Thus salience detection , like pattern detection, is a set of abilities that underlies everything routinely achieved (at some level) by multicellular organisms, and especially by vertebrates. For many people, salience seems like it ought to be a trivial ability. It is anything but trivial. Salience cannot yet be coded into any data science model. How can a thing made of silicon be designed so as to feel anything? It can’t, and this is “strike two” in the bottom of the 9th for AGI.

Next Post, it’s Game Over for AGI!

Thanks for reading,

Bill Lambos

1. I hold a postdoctoral certification in clinical neuropsychology and a license to practice in California and Florida. I’ve been coding since mainframes were the only accessible computers and LISP was the ‘lingua franca’ of AI (ca. 1970-81). My doctoral thesis in experimental biopsychology involved determining the full set of (known) computable parameters for the associative learning process called Pavlovian conditioning. Next, our lab coded these parameters as time series events into one of the first microcomputers in an attempt to simulate such learning. BASIC was too slow, so the coding had to be done in assembly language. In 1983 (yes, I am that old) most microcomputers were built upon the Zilog microprocessor platform (anyone remember the Z-80?). This further required I complete a Master’s Program in Computation in order to finish the thesis, which I did. Finally, I hold another Masters degree — finished in 2022 — in Data Science.