chaosmith-site/content/posts/transformers-and-NNNs.md

Transformers and Neural Networks

2026-06-25

When Wetware Fails, and Hardware Tries Its Best

Every time I read another article about AI replacing programmers, I roll my eyes. The code is usually mediocre. The architecture is just wrong. The output needs UITL supervision, steering, and sometimes - purification by fire.

And yet I use large language models every day.

They don't make me a better programmer, or a better engineer, or a better architect. Hell, I'd argue they negatively influence my cognitive depth. I use them to conserve durability.

I have a neurological condition. On a bad day my limbs, my fingers - are in pain. They stiffen. They stop bending properly. They lose sensitivity. They can't reliably be where I need them to be. The core, the concepts, the thought process - it's all still there, but the bandwidth between thought and keyboard has packet loss worse than ADSL over FM radio.

Most people evaluate LLMs by asking whether they save time, whether they know things, whether they make the "right" decisions. I evaluate them by asking whether they convert instruction to intended result.

Time is abundant, but the energy isn't.

Given enough hours, I can write every line myself. The question is whether I can physically or psychologically afford to. An LLM lets me spend words instead of keystrokes. It lets me spend intent and get the uncompressed version of it with remarkable 60-70% accuracy with frontier models, and 40-50% with whatever MoE I can run in 8GB of VRAM.

The result is rarely perfect. But perfection has to earn its place in queue. Point is - continuing to create, think and test things, even when it's the equivalent of wiping with sandpaper: the thing you did was good, but the price you pay is just this painful.

Transformers did What?

Some time after something gets invented, the meaning of the word describing it changes forever. This is something I have experienced when interacting with Americans: CrockPot, Dremel... All of those are brand names that were fortunate enough to become the names of the category of things they describe. Nobody thinks "Multi-cooker" when they look at one. Same thing happens with Transformers and Large Language Models. After so much advancement in manual and automatic language mangling, people tend to forget what a phrase or word like that would mean literally:

• Transformer - Something that transforms itself or something else from one configuration to another
• Large Language Model - A model of human or non-human language that is large

The kicker: most advancements in LLM's emergent behaviors are just re-discoveries of the field called "Linguistics" - the science of language. But that is a topic for another time. If you can compress data, then it needs less bandwidth. Most HTTP requests we do every day are gzip-compressed and we only notice it when they are not, and suddenly the load times are crazy. Language is a compression of information stored in one's brain, their latent space, if you will. So when it comes to language models - that is the layer they operate on: language. Language is representative of our intents. Portraits are also representative of the original human. Both are heavy in losses, no matter how much you try. Just because you can fit the entire Shrek movie in a jpeg, doesn't mean that its representation will keep everything present in the original. So: you operate at that level.

Rant about a future topic

LLMs are not AI. They are artificial, but they possess no intelligence. They only possess a model of language granular enough to have a representation of intelligence.

So: how about I compress my intent into one language (English), and the LLM simply translates that into a different language - a programming language? This is exactly where the LLMs shine. And all the "groundbreaking new research" in prompt engineering is re-discovering this completely unthinkable phenomenon.

And this is where I live. This is where I can compress intent into language, and a language model translates that language into a programming language. It sucks at coding, but it excels at translation.

It's the prosthetic I never had, yet the one I need until we figure out optic fiber, gold or copper axon replacements.

From the moment I understood the weakness of my flesh, it disgusted me. But in our universe, this leads to making either a machine or a human employee try to decompress your intent from lossy compressed protocol called "language" into actions. And yet, it works. Worked for millennia. Except for machines that try to do anything other than translation.

Reminds of a joke: If a deaf person loses a finger, is that an accent or a speech impediment?

---

---

DECLARE RULE UNDERSTANDING_MAKES_RELIABILITY:
  CONDITION ALL {
    NODE System AS S
    EDGE Uses AS U
    PROP S.reliability < 1.0
  }
  EXPECTS CHANGED EDGE Uses
  FOR INFER U.reliability_total = S.reliability * (1 + U.understanding)
  UNLESS U.understanding < 0.1
  AS LOGICAL

BEGIN_TICK 2

FIND NODE h OF TYPE Human BY PROPERTY alias = "G[REDACTED] G[REDACTED]"

LET llm = CREATE NODE System
SET NODE llm.name = "Large Language Model"
SET NODE llm.reliability = 0.59

LET u = CREATE EDGE Uses WITH h llm
SET EDGE u.purpose = "prosthetic"
SET EDGE u.understanding = 0.74

END_TICK