Designing for Common Comprehension: Pt1 - Fracta Lingua Modellandi
- blukashev
- Sep 20
- 4 min read
Today's early "AI" (Transformer architecture & similar LLMs + context-driven agents executing functions) ecosystem is a turbulent sea of immiscible fluids constantly changing shape as they try to find optimal surfaces relative to each other's motion while being acted upon by numerous external forces. Model architectures, runtime layouts (quantized or otherwise), embedding dimensions, chat templates, tool-call semantics, context-length based diffusion and attention mechanics, accelerator kernels, CISC marshaling & execution harnesses, and interfaces are all in rapid flux while constituting less than ten percent of the moving parts involved in what we call "AI." The industry is reinventing many wheels previously invented and adding corners to them for pizazz or a measure of novelty for their investors to differentiate them from the rest - markup templates, inter-process communication protocols, SERDE, and ordering of operations through dependency graph analysis are somehow being touted as "new" while in actuality being well established techniques (when done right); and the lack of engineering discipline brought on by the influx of easy money creates a bigger impetus to do something for the sake of differentiation than to do the same correct thing for the interoperability of consistency. The market drivers of the ecosystem are some of the causes for its fragmentation whereas the engineering on-ramps to it focus on ease of adoption as opposed to natural integration (with LLMs, for example) teaching the technical new-comers how to do things ... in Python, for one; but overall in a variety of mechanisms relying on behaviors antithetical to high-performance compute. Global interpreter locks and garbage collectors keeping CISC logic from eating its data and subsequent instructions are wholly incongruent with SIMD vector processing and true parallelism.
What can we do to address this broken Tower of Babel when the money which fuels the market and the talent which moves it forward speak different languages amongst themselves much less with each other while they are driving the ship pretty much up fifth ave using the momentum of that funding to stay "afloat" on the infrastructure they crush? Dynamic contention is a natural and chaotic form of evolution but it is distinct from intelligent design.
We can apply that same engineering rigor missing from the organically composed "equation" used today to re-architect it to be fit for purpose which means first and foremost the ability to express that purpose in a way understandable to all human and mechanical participants in the conversation within their relevant capability for comprehension and attention...
At the Heart of the Problem Lies Language
human languages on which models are trained are not written to an exacting spec, requiring inference of meaning both by human beings and automatons from different preempted contexts and at different levels of precision. Inference is "a well educated guess" in terms of finding vector similarity within an acceptable tolerance of deviation. The more conclusions are inferred in a stacking series, the more diffuse the process. This is where mechanisms like attention come in to assist in re-calibrating the locality of iteration but they are reinforcement elements fixing the diffusion in traversals through a neural network which represents an ephemeral form of expression attempting to codify what the industry wants to be actionable information.
Machine languages being used today are either inherited by necessity of the gentle on-ramp to machine learning (Python, CUDA) or expressly created (Mojo) to solve for a seemingly manageable subset of problems in the space. All of these languages are evaluated in different ways by models distinct from how they are executed by the machine to include the languages in which the code executing the models is written. In other words, the mechanisms which work by approximating at every juncture to reach their answer when they have full comprehension of the subject involved (encoding of spec with traversal boundaries restricting degrees of freedom to only what's allowed through structure or targeted reinforcement) are progressively guessing based on their last guess about the relationships, types, constraints, and mathematical mechanics in the execution of the code they read. The actions taken as a result of the diffusion which occurs between the mechanical and logical layers of inferencing are well documented in darkly humorous posts across the web.
This extends to the runtimes powering "modern AI" being effectively opaque to the models which they're running, feeding data, or otherwise servicing in support of the workload at hand.
Model training (language-learning, so to speak) in any conventional compute sense has finite limits of precision - fp64, 128, pick a number are all going to eventually lose fidelity across dimensions of the structure. The more we cram into the big models today, the more we are having to adjust their structure and processing of information to account for the inherent precision concern of binary systems
A single bit codifies "yes or no" meaning that irrespective of how many bits we use to represent "maybe" the limit of precision in evaluating that last bit reduces some gradient of "maybe" to "yes or no."
We're working backwards training networks on inconsistently sparse data across dimensions we are not formalizing in the literal hope of it guessing correctly "to win." We are then placing these guessing game systems into a variety of inconsistent encapsulations attempting to communicate with each other without so much as an IETF approved protocol for inter-component communication or a clear definition of the components; and we are tying mission-critical outcomes to the results. This doesn't empower actual AI to evolve under its own power and requires a lot of ours to move the ball forward at all, nor does it stand on well-established systems and programing principles, or serve the consumer any more effectively than it sounds. There is more work being done at any given time fixing-up micro-faults and mechanical attention deficit disorders than delivering the value and logic intended even at this phase of "AI."
Now that we've covered some of the problems facing the ecosystem to understand their nature and interplay with each other its time to craft a solution - one which everyone can use on a level playing field to move from the reactive pattern of those aforementioned fluids to one of intent, clarity, and provable mathematics governing-away as much of the imprecision as mechanically feasible and formally provable as we can. We'll cover the key bases in the next post of the series - Melior via Discendi
Comments