DarkHorse Podcast asks, Did AI just Evolve?

I don't think so

Image generated by LTX from the prompt: Show old fashioned computers evolving to modern day computers that can think.

On October 23, Bret Weinstein and Heather Heying, hosts of the DarkHorse podcast, reported the discovery, which is described in a pre-print paper (not peer-reviewed). Weinstein and Heying were spooked by the claim, supposing that this instance might be just the beginning. Soon LLMs might advance science too rapidly into unprecedented frontiers before proper cautionary measures can be taken. After listening to the DarkHorse episode, I decided to look into this.

It is a fairly common practice in many areas of science these days to use LLMs to try to find unnoticed patterns across various databases. I would expect an LLM to be an excellent tool for sifting through vast amounts of data to highlight what’s already known. The LLM in this case, called Cell2Sentence, is trained on datasets mapping RNA in the cell (which gives you an indication of the state of the cell) to cellular responses. Using this tool, researchers can predict and retrodict cellular activity based on past data and even generalize about novel combinations of states. Nothing miraculous here.

The researchers were looking for possible uses for 4,000 known cancer drugs. This is typical activity for cancer research: find a use for a synthetic compound for which a company has a patent. It’s a solution looking for a problem.

(Why not look for environmental toxins that cause cancer instead? You know the answer.)

The paper is very wordy, so I have simplified the prompt they gave the LLM: “Find a drug … that would boost the immune signal [antigen presentation] only…where low [and inadequate] levels of interferon… [are] already present.”

Let me unpack this as best I can.¹ In a healthy body, a cancer cell is marked by an antigen that functions as a signal to the immune system to destroy that cell. Interferon is a protein produced by the cell to signal when it is in trouble. Interferon increases the expression of major histocompatibility complex (MHC) antigens: those are part of the “kill me” tags. In other words, interferon leads to the unhealthy cell being tagged for destruction by the immune system.

The researchers were looking for a drug that would amplify interferon to increase the expression of MHC antigens (the little tags) that would get the immune system to kill the cancer cell.

The researchers claim that the LLM made a novel hypothesis when it suggested that silmitasertib, one of the 4,000 drugs explored, would increase MHC antigens only in cells that had some low levels of interferon already and silmitasertib would have little or no affect in cells with no interferon. They claim that this is a novel discovery because in the literature that Cell2Sentence was trained on this had not been noted. According to the Google announcement about the paper, “it has not been reported in the literature to explicitly enhance MHC-I expression or antigen presentation [my italics].”

Silmitasertib received a patent for its ability to inhibit an enzyme called Casein kinase 2 (CK2) that is essential for numerous cellular processes and can be overactive in cancer cells. Since CK2 plays a role in regulating interferon signaling pathways, one might expect silmitasertib to have some kind of an effect on interferon, which is probably why this drug was included in the list of 4,000 to be tested with the prompt.

I suspect that the novel hypothesis was implied in the literature, since inhibiting CK2 might stress the cell, triggering interferon production; interferon leads to increased MHC. The fact that it did not really effect cells with no interferon might be the surprising thing. It’s not entirely clear to me.

In any case, as Weinstein and Heying note, at best it’s a “small leap away from what is known.” Weinstein assures us that this is not an instance of AGI (genuine general intelligence). However, says Weinstein, if this is not a total fluke (which he doesn’t rule out), “we are on the foothill of AI creating hypotheses,” and that “scares the crap out of” the DarkHorse hosts.

My conclusion is that Cell2Sentence’s discovery is not that remarkable. It’s the sort of pattern that one might expect a computer to be able to find when you specifically ask it to look for that type of pattern. My question is, How many LLM generated “hypotheses” had to be tested in a wet lab before one was found to be correct?

More importantly, what are the potential off-target effects of silmitasertib? An LLM might have a much harder time with that prompt.

I may be wrong. I am not a biologist. I was initially trained on a philosophy dataset that included many obtuse post-structuralist and postmodern texts. I learned how to filter through bullshit to get at the rather simplistic claim that was being made in order to analyze it. I was able to apply my skillset on bullshit papers sponsored by the pharmaceutical industry, and here I apply the same method to this Google-sponsored paper about a new product they want to sell.

The Cell2Sentence LLM has 27 billion parameters. This says to me that, like all computer processing, it’s working with brute force, trying out all possible combinations and patterns. Biological systems tend to be much more efficient, and I think I know why.

The novelty of LLM performances, I wager, will always be in the prompt, insofar as the researchers come up with good questions. They wanted to find a drug that requires some interferon to be able to amplify interferon. Somewhere in the data training set this was implied. They found what they told the LLM to find.

LLMs are Good Tools for Sifting Data; They are not Creative Minds

LLMs will never be sentient or truly creative. (I could have softened that with “In my opinion,” but I decided not to in order to hook you.) I contend that if AI cannot invent novel hypotheses, it is because it uses only one type of sign, a symbol or code, and does not use the kinds of other signs that biological systems use, through which they make cogent novel hypotheses by gleaning qualitative information from physical contexts.

Computers only manipulate abstract symbols in abstract contexts.

Biological semiosis takes place in a wet messy soup. Instead of wires connecting nodes, as in an electronic circuit, there are chemical reactions, series of transductions/transformations of molecules. These local reactions give rise to coordinated electrochemical fields that differentially affect cells.

There are molecules (synthetic, foreign, or natural) floating around with shapes that might mimic the shape of other molecules, leading to unpredictable side-effects. In biology, the physicality of the sign matters. Spatial position matters. Interference matters. Simultaneously activated signal pathways—which can interfere and interact with each other if they share molecules—can become associated; in this way one thing sensed by the organism can come to stand for another.

Living systems can create new signs. They don’t have to be programmed.

For those interested in more details, see my paper, “Self-Reinforcing Cycles and Mistakes: The emergence of subjective meaning,” which was included in a book in a series dedicated to astrobiology, a field where philosophers and biologists explore the basic conditions necessary for life and intelligence to emerge. Astrobiology is an appropriate ground for interrogating theories of artificial intelligence.

Weinstein, Evolution, and AI

DarkHorse is one of my favorite podcasts. The secular intellectual interests of Weinstein and Heying match my own: they are evolutionary biologists and complex systems scientists. Heying originally majored in Literature. I like their banter. I feel at home with them. I suspect they would be shocked by nothing I am saying here.

In an interview with Joe Rogan a few years ago, Weinstein acknowledged that Creationist critics of the theory of evolution by natural selection are correct in arguing that the time scales required by the theory are too great. (I do similar but secular work with butterfly mimicry as my test case.) Weinstein guesses that some other kind of “computer language,” that we don’t know about, is running on top of the digital hardware of genomes that enables adaptations to occur at a more rapid rate.

The selection for reproductive fitness “trains” a population’s genome slowly, changing to the settings that multiply fastest. An individual’s acts of rare genius are forgotten if they don’t lead to lots of offspring. This may be compared to the way the LLMs learn. Unique individual decisions matter very little because it’s the most common patterns that prevail.

But let’s drop the computer metaphors because trying to understand vastly complex systems (organisms) by comparing them to simpler systems (computers) might be more confusing than not. I do think Weinstein is on the right track, but what he is looking for is a theory of biosemiosis, not of coding per se. Organisms use different kinds of signs, not just codes, to rapidly evolve.

As I say again what I have said before, this time I highlight the fact that those who tend to dismiss me come from two opposite directions, spiritualism and materialism. I travel a third way: I argue that intelligence and creativity are biological concepts.

Spiritualists and Materialists Versus Biologists

Spiritualists are content to say that humans have souls and God-given freewill and therefore humans are different from machines. This answer satisfies only a certain segment of the population. I also want to speak to that segment of the population whose metaphysics does not posit supernatural spiritualism. Besides, it was that madman Descartes who popularized this line of thinking. Ultimately, he is no friend of spiritualists.

Rene Descartes on a podcast by ChatGTP

Descartes claimed the world is divided into the material, which is open to empirical investigation, and the immaterial, which is not open to empirical investigation. Descartes set in motion the idea that what we call the “spirit” or the “elan vital” can neither be proven to exist nor analyzed. This is in part why what we call “modern science” has mostly limited itself to mechanistic approaches, even in biology, and any scientist who is a little too interested in non-linearity, irreducibility, and complex system science is often viewed as suspect by materialists.

Materialists insist that biological intelligence can eventually be fully modeled with enough information about neural activity. They believe that ultimately a neuron is like a passive node in a computer network insofar as it is simply biased by electro-chemical effects. They concede that, yes, neurons aren’t just on or off; they have multiple states. But that, they say, is just a matter of complication, not complexity, and computers are well-suited to processing the multiple dimensions that contribute to a neuron firing or not. At best, they might concede that a single neuron needs to be represented by an entire network of on/off switches, but ultimately, the materialists insist, intelligence can emerge with a sufficient number of weighted connections, discrete groups divided into definite hierarchies. All information can be digitized and expressed as objective quantities not subjective qualities, they say.

Biosemiotics

A sign is some thing/quality that stands for something else, to an agent like a cell or an organism. Because Biosemiotics is a science of how cells form meaningful associations and respond to signs, it is a science of the immaterial, in a way. It overturns the destructive dualistic notions of Rene Descartes—whom, it appears from a latter day CT scan of his skull, had a brain tumor, may have had Exploding Head Syndrome and had symptoms of right hemisphere damage insofar as he often mistook living organisms for machines.

There’s a common assumption that all thinking involves symbol manipulation. The etymology of the word symbol, is syn (together) + bol (throw), throw together. It refers to an arbitrary connection. A symbol is something that arbitrarily stands for another thing. The term “code” has the same meaning as “symbol” in semiotics.

Examples of symbols/codes:

• The word “tree” stands for a big leafy plant with bark

• In Morse code, dots and dashes stand for letters

• Ones and zeros in computers stand for numbers or letters or functions

• Python programming language stands for ones and zeros.

How are Codes Created?

• A relationship between two arbitrary things can be created by an external intelligence: programmed or encrypted according to a rulebook

• An arbitrary relationship can be created by a physical tie. For example, vaccinologists can use a conjugate to link a protein to a toxin; or, for a natural example, a nucleotide triplet is yoked to one amino acid using a protein tRNA.

• An arbitrary relationship can be created by a lengthy selection process, which picks up common patterns through repetition, as with LLMs or neural networks. This is also a kind of “fire together wire together” idea, such as Donald Hebb proposed for neurons.

• You can also think of a symbol as a habit, a well-worn path, a convention, connecting one thing arbitrarily to another.

The key thing is you have to learn what a symbol refers to. Its associated meaning is not inherent; it’s forced, in a way.

Biological systems do evolve codes/symbols over time (for example, DNA), but they also use other kinds of signs that don’t require training or programming from an external source. Biological systems also use signs whose associations are derived from qualities in context.

Let me present an example of these different kinds of what we can think of as grounded signs. Imagine that you are traveling to the airport in a country where you do not know the language, say China. On the way, you see a sign with the image of an airplane pointing in the direction that you are headed. That’s an icon sign. You know to stay your course instantly without having to decrypt any arbitrary code, such as 機場. Soon you encounter an exit from the highway and you see another sign with an image of an airplane. This time the airplane is pointing to the right in the direction of the exit. This kind of sign is called an index. Icons and indexes take their meanings from qualities (similar shapes) and physical contexts (pointing).

You don’t need a brain to make associations via icons and indexes. All biological cells can do this. This very primitive form of biosemiosis likely preceded the more complex semiosis that eventually emerged with human language. Likewise, words probably began as icons and indexes, then were abstracted from their contexts and became symbols that one has to memorize.

LLMs only use symbols, that is, conventional signs, that are learned through repetition. Rote memorization is for dummies.

Living systems can generate new hypotheses spontaneously, based on context. They don’t need an external programmer or millions of repetitions to learn something new.

To illustrate rapid biological sign creation/learning, let me first mention the “mind palace technique” for memorization. If you want to be able to recall the lines of a very long poem, you can walk through a house with several rooms, looking at some items in the rooms and associating one object with each of the lines of the poem. When you want to recall the poem later, you can imagine walking through the rooms. As you look at each object in your mind’s eye, it will trigger the recollection of the right line.

Second, when people develop Post-Traumatic Stress Disorder (PTSD), a random object (like a green shoe) that happened to be in the context of that traumatic event can later trigger the memory of that event and cause the person to have anxiety.

PTSD is similar to what is called a “flashbulb memory.” Some people can recall the details of the place where they were when they heard surprising news. For example, recall where you were when you learned that the Twin Towers had exploded on 9/11.

From my own experience, I note another example. I listen to a lot of audiobooks in the car. I often lose my place and I have to search through the recording to find where I left off the day or week before. When I hear lines that I have already listened to, I picture in my mind’s eye exactly where I was on the road when I heard those words.

These are all examples of a kind of memorization based on the proximity of some arbitrary object to something else. Importantly, this seems to happen instantly in a living system. AI would have to be extensively trained to make such an association.

Importantly, this kind of recall, using an index sign, is possible even in simple organisms like slime mold or pea plants. In my work, I argue that such examples of instant Pavlovian conditioning are probably made possible when two different signal pathways for recognizing different objects interfere with each other.

In contrast, computer nodes in a “neural” network are passive. They are modified as information flows through them and changes their “weighted” biases. They do not interpret the information that flows through them (they do not associate the information with the objects in the room that they are in or notice coincidental similarities in physical shapes of symbols) and they pass the information on to another node just as they received it (minus friction).

Living cells make subjective decisions in response to signs in their environments. They are not passively molded. Subjectivity means that a cell can respond to, say, a molecule in its environment that is merely similar in shape to one that it is evolved to respond to. Cells can make physical puns, in a way.

Using index and icons signs might be how cells can have novel responses to things. Importantly, they don’t explore new tools randomly. They can only interact with signs of similar shape or that were previously associated with a specific context. Cellular creativity is helpfully constrained.

The Future of LLMs

I asked DuckDuckGo’s AI search assistant to summarize my above mentioned paper:

Overview of “Self-Reinforcing Cycles and Mistakes”

The paper explores how subjective meaning arises in living systems.

It emphasizes the role of memory in shaping responses to environmental signals.

Two levels of subjectivity are identified: one from memory and another from interpretative responses to signals.

Key Themes

Memory and Anticipation: Memory modifies internal pathways, allowing organisms to anticipate future conditions.

Interpretation: Responses to signals can be seen as misinterpretations, contributing to the emergence of subjective meaning.

Implications

Understanding subjective meaning can enhance insights into cognitive processes in both biological and artificial systems.

The findings may have applications in fields like robotics and artificial intelligence, particularly in developing adaptive learning devices.

In my paper, I do not suggest that my work could be used to advance robotics or AI. Probably most of the papers on intelligence can be applied in the tech sector, because that’s the kind of research that gets funding these days. So the LLM just says that my work might be applied in robotics or AI too.

It is not accurate to say that I emphasize “the role of memory in shaping responses to environmental signals.” The traditional literature that the LLM has been trained on undoubtedly states this obvious fact. Differently, I emphasize the creative distortion of memory in learning new responses to environmental signals.

Also, I say that only certain “responses to signals can be seen as misinterpretations,” and this is rare but potentially significant.

It’s funny how AI converts every novel idea back into the consensus opinion that the author is trying to overturn or move beyond. We might not ever advance our knowledge if we rely on AI to think for us.

1. I have taken three graduate-level courses on the immune system with James Lyons-Weiler at IPAK-EDU.org, which gave me the confidence to critique the Google-Yale paper.