Software has its fair share of naming conventions and traditions. Here are a few infamous ones: PHP means “PHP Hypertext Pre-compiler” but used to mean “Personal Home Page;” GNU means “GNU is Not Unix;” and Google is a misspelling of “Googol,” a large integer that can be written as a 1 followed by 100 zeros (10^100). Software engineers, like lawyers, know that words have intrinsic meaning and it's important to name your concepts aptly to explain it well. But what if you want to misinform, or covertly hide meaning?
What I'm getting at is the notion of when first mover's choice of naming is imbued with an ulterior motive, akin to delusion of the uninformed for personal gain. When you peek what's under the hood, the foundations of these software tools are transformer matrixes (i.e. some fancy linear algebra with feedback loops) as “Large Language Models,” or LLMs. Microsoft and OpenAI then brainstormed how to make this more approachable, giving rise to the name “Copilot,” like a trained, skilled, and educated human officer who can take command of an airliner. This kicked off the term for these automated interlocutors (i.e. text input fields) as “agents,” with human names like “Claude,” after Claude Shannon, a Bell Labs scientist who authored Information Theory.
The term Large Language Model is weak when you're critical of it. It's a hollow name for what's going on. “Large” is an arbitrary adjective for the advancement achieved. Language Models have existed for decades. They are matrixes of values with probability weights assigned to them to inhibit or to elicit an output from its input. These probabilities are refined over a period of training, to calibrate the input and output such that the words that emerge “make sense” to the user. The leap that happened in the 2020s is due to reaching critical mass. The datacenter cluster capability to train these language models are large enough for the training and inference to operate. If we have learned anything from the early ages of computation with its room sized computers, device volume will continue to expand and contract as we refine the operating motives to our desired expectations. LLM will be an outdated term, so we must adapt.
I propose SLiME. SLiME software tools are Sequenced Linguistic Modeling Engines, meaning an Engine that does work, using a Model of a knowledge domain, based in a human Language, Sequenced from artifacts. A Sequenced Linguistic Model Engine, like an internet search engine, is an operable software tool that extends the manual ability of the user to interact with a language. Intelligence is not advertised because the user will find none inside, instead it's an intricate knit of factoids and tabloids.
These are SLiME tools. Copilot is a Microsoft SLiME tool. Claude is an Anthrophic SLiME tool. And SpaceX's Grok is a CSAM SLiME tool. In my other post, regarding the capacity of software tools that use LLMs, I argue that LLM/SLiME tools will not approach AGI qualities because SLiMEs fail Searle's thought experiment The Chinese Room, despite what Functionalists clamor. The end result is seeing Copilot and Claude for what they are: not operators of the rooms, but phantom tools: fluent sequencers of text (signal sources) that seamlessly pattern match based on combing the entirety of the internet, petabytes of data from decades of existence.
By adjusting the fixture upon which our perspectives leverage, pattern descriptions can be interpreted differently. Like a magnifying scope boring into a protozoa. A classification of text output could be labeled “behaviors,” with peculiar species that exist outside of “normal” results be named “hallucinations,” but it's just a malformed Sequence from the Modeling Engine. That even more slippery classifications are those that are subliminal (from sycophancy, to racist dog whistles and suicidal ideation), can make more sense when knowing the engine is statistically based in the training dataset from the AI labs. SLiME tools, based in their weighting, traverse a hyper-dimensional maze, towards a satisfying result based on prior art. Starting from the “question” location, along probabilistic paths of text, until it arrives at a “statement” destination. Along the way, the sequenced training data informs the transforms to which symbol leads to the next, until a carriage return.
This essay is to practice language as any English speaker would since Shakespeare's time: to innovate on new terms for new concepts to model what is observed in the world. It's important to call out bad practices by anyone, no matter how large the company.