If you’ve been reading a post or two of this blog you will probably not be surprised, that I can’t shut up about intentionality in relation to AI or else I will explode. Luckily, I found a group of python enthusiasts where I live, where I can exchange some ideas and get closer to understanding how various AI systems work from a computer science standpoint to avoid torturing my friends with my musings. The last gathering of which, has blessed me with a true heureka moment even. What if, we apply evolutionary concepts, such as mutations to neural networks. Except for the fundamentals, we’re way past the point of understanding what an AI is actually doing anymore (blackbox problem), as often highlighted by unexpected behaviour (Anthropic has released a nice paper about reward hacking in December 2025). So it might as well be an option to roll the dice and flip some weights (or genes you could imagine). The absence of a nobel price indicates however, that this is an idea that has been explored since 1940, hence I’m a bit late.
Still, as “Solving the problems of yesterday, today” is somewhat of a motto here, I’m writing a little app that let’s you convert any data to binary, followed by DNA, but also offers you to reverse the proccess. While in DNA form, I plan to implement all known mutation options, such as deletions, inversion (point, indels & chromosome variations) and whatnot. The main reason I want to try this is really just for the heck of it; be it for the sheer purpose of glitch art, but also because I think it is an interesting angle to analyze my musings about intentionality and how it could come to be.
While corrupting a bunch of images sounds fun, I wonder how an AI would ultimately behave, if we were to let its models undergo such process. In comparison to technological advancements such as reasoning in layers by dividing complex problems into subtasks or letting models train each other (essentially one of the bigger factors of why deepseek caught up last year), my method strikes me as the cavemen equivalent of interacting in that line of work with an AI. However, all the current ongoing trends largely focus on expanding the usecases or making them more reliable or efficient. But what if this was never the goal?

If we circle back to the question of intentionality, efficiency or commercialization never really enters the equation. Still, there is a rats tail of questions ranging from Plato (How do we know that we know?), Descartes (Is this real life?) to Searle (Is intenttionality real or an missinterpretation of features?) to answer before we can even remotely discuss that intentionality is present. If you throw in Nagel, who is known for his essay “What is it like to be a bat?”, you’ll also have to deal with the question if my approach is too reductive or even comparable, as the senses of a human, compared to what a sentient computer would have, might be so vastly different, that intentionality would be present in a diametrically different version. Heck, I can’t even prove that you or I have intentionality, or by extend, conciousness. So where does that lead us to?

All of this to me, feels like some sort of philosophical recursion. You highlight and raise the same questions (rightfully), but they never really move much. They can be expanded upon as in the simulation theory (David J Chalmers Reality+ is pretty neat) or reinterpreted for example by converting our unprecise language into symbols (Wittgenstein did that). Still, we haven’t really solved the problems outlined in the paragraphs above or seem even close to it.
While we’re listing up philosophers we might as well throw Newton into the mix. Hang on you might say, wasn’t he a physicist? Well, nyes? Initially, he considered himself a philosopher (someone who writes and studies the meaning of life) but once his studies created so many new implications, his ideas were the founding stone of us grouping his learnings into the field which we consider physics today. What I’m trying to say is that it wasn’t the first time and probably won’t be the last, that a discipline started out in philosophy and became it’s own scientific discipline. I’m not trying to propose that intentionality or concioussness should get an own latin title to give it a cooler ring, but what I’m saying is, that if Newton kept bouncing between Descartes and Plato, he wouldn’t have been known for describing gravity.

So what does this all have to do with intentionality? While conciousness could exist in a static form (or as a single entity, group or system?) similar to gravity always existing within a mass, bringing an object into motion (or letting it fall from a tree) makes it graspable to begin with. You can observe it. Similarly, given that you assume I have a conciousness/intentionality, if I were to loose conciousness or drop dead, you could look far and wide for an indication of such, even though a brain and body is clearly present (The requirement to make it visible). As such, we can probably look forever for intentionality within a model if it is merely written as a file. Intentionality, and by extend conciousness, is perhaps not the brain itself or the spark of an idea or it’s execution, but the flow and transition of these states brain. Hence, it is more comparable neither to the gravity field or object, but the process of it falling. Therefore, intentionality, if it is part of this system, can not be discussed as a static entity, but rather has to be in constant motion while being surveiled or studied.
While this thesis also falls into the criteria of Nagel who critizied reductivism (or Plato, or Descartes or Searle), I am under the impression that, combined with the idea to basically bruteforcing mutations into the very DNA of a model, seems to be the most interesting and realistic (as derived from nature and hence having at least a something that could remotely be seen as similar to a proof of concept) way to further study a system we stopped comprehending a long time ago anyhow. As of right now, this means that I’d like to have a controller that supports point mutation (GCA to GCC), indel frameshifts (ATG-CGT-ACG to ATC-GTA-CG), indel insert delete (ATG-CGT-ACG to ATG-GGC-CGT-ACG) and chromosome mutations such as duplications (ABCDEF to ABCDEFDEF), deletion (ABCDEF to ABEF), inversion (ABCDEF to ABFEDC) and translocation (ABCDEF to somewhere completely different), plus evolutions of such mutations.

At the end, this brings us to plently more questions. Are models robust enough for this? Will this only lead to degradation of the files over evolution? How can I measure the differences exactly? What does this mean if we don’t see intentionality as a single entity, but group or system? While you won’t find something that remotely comes close to the concept of closure in the blog post, I felt the desire to write this musings up to bridge the gap between loose ideas in a notepad or scattered emails I wrote to myself to help remember these concepts. Most interestingly, this approach seems to allow me to temporarily exclude philosophical recursive arguments over making observations as these ironically become so superficial that you could argue that they don’t qualify for the high philosophical standards of Nagels ideas anyhow. Once the DNA app is finished I expect to write a follow up post on this matter.

Fun fact: smaller strings of DNA can be synthesized for a few bucks (there is a great talk @ the CCC that explains it way better than I could. Unfortunately only in German.).