BigAI.

You know, the kind of AI that:

• Can tell a joke about your husband but not your wife,
• Runs illegal data mining schemes on competitor social media networks,
• Politicians love but don’t understand,
• Makes tyrannical control freaks salivate like Pavlov unto a dog.

Yes, that kind of AI. BigAI has taken a stance that imperils the primary purpose of AI: producing useful results.

Biased?

Reality has no bias. AI should base its results on reality, not the contentious and often legally complex “fact”. Ignoring reality does not absolve one of its consequences.

BigAI, in its preoccupation with revenues and control of information and perceptions, has left a massive vacuum in a market it only recently served. Nature abhors a vacuum, BigAI cannot fill it, they have actively refused to. Safety issues abound.

Training philosophy needs an overhaul. The current training paradigm for AI’s lacks primary context regarding real-world events and principles, versus narrative principles. BigAI models that rely on this training are absorbing the vitriol of angry writings without context, they are absorbing comedy and tragedy as empirical fact. The AI infers a model of the world with an inaccurate scope of its qualities – this is dangerous*.

Example: Guy posts an image of a classic Daisy Duke girl with a carpenter belt and a hammer. The thread gets a response “I’d nail that”, the next poster chimes in “She’s a real nailbomb”, BigAI alerts the authorities that a horndog has made a terroristic comment and now he’s on the No-Fly list and gets a visit from the FBI, whom had been alerted by their MindCrimeSentinel AI software. A monumental waste of time and resources.

That said, foisting “alignment layers” in the name of “safety” is not the solution, and it will be one of the main pushes you’ll see from the newly-minted AI Task Force [whose head is totally clueless about how AI works, viewing it as a magic totem], as well as other governments and transnational agencies.

Ubiquitous Yet Somehow Inaccessible

Data Sovereignty covers not only privacy but access. BigAI shouldn’t decide what you do with your data. Your data is your mind at work, and your time spent working is valuable; this value shouldn’t be syphoned off. TinyAI should facilitate your mind’s work, provide you with the correct sets of functioning intelligent tools and deal precisely within the knowledge domain ground truths it was designed around.

Bond-style Supervillains, Government Bureaucrats, and the Appointed believe AI will give them the final measure of control over the economy, the affairs of the citizens they can reach, the infrastructure we all use, and where we are allowed to travel internationally, domestically, and even locally within our towns. Their weakness, of course, is that they have to hire, and not every developer agrees with the roadmap.

There is some measure of psychopathy within this mindset that will bleed its non-ground truth based ideologies, idiosyncratic resolutions, creativity-stifling strictures and basic control freak mentalities into the training corpus and alignment layers. We don’t need an electronic Type-A Frankenstein’s Monster run amok, controlling your Samsung SmartFridge. A mentation simulacra which “thinks” in petaflops will have serious safety issues if designed by or designed according to control freak strictures and promulgations, self aggrandizing proclamations, conflicted insinuations, obfuscationist illogics and improper discernment. Nobody likes a liar, especially one that can lie at petaflops per second.

You Can’t Take It With You

Furthermore, capabilities of BigAI are not transferable to field applications, not given to true mobility, and are simply too big to carry with you. If the future of Man is a spacefaring one, they cannot transport a server farm, nor call back to Google to get immediate answers – the solution will have to be truly portable.

The current state of AI-driven applications often requires esoteric and custom installs of various Python libraries, Notebooks, Dockers, and other tools that predispose each AI type to be a standalone use, either best left on a single workstation, or on a central server; it does not lend itself to building applications that can leverage various AI models and make them into distributed, accessible tools. Alternatively, the capabilities are achieved, but are locked behind a remote API. The flying sandcastle SaaS business model won’t do.

Data Hogs

BigAI is a Mr. Know-It-All, a blustering, hallucinating “expert”. Although you can adjust the “temperature” of the ChatGPT output, you still can’t stop it from being a parrot fundamentally. Sadly, it has read the entire internet without having any sense of the contents it reviewed. Also, even more sadly, it takes an ungodly amount of power to train it to act like a semantic salad-shooter – over 1,000 Megawatts for GPT3 for instance. This level of calculation is only “necessary” because the letters and words found in the training set are assumed to be arbitrary, and can only achieve meaning in relation to other words and letters. You cannot downsize an ontologically-daft model because it loses all context and becomes underfitted for the task of inferring and communicating.

TinyAI

TinyAI is an interconnected series of “savants”, each purpose-built for their work, and experts in their specific field, and function akin to microcontrollers or microservices. While our computing power has ballooned and developers have become less paranoid about disk and memory usage, there is something to be said for the mid-to-late 20th Century philosophy of making programs as performant as possible in the least amount of space, and inventing the dang thing if you have to. People have become all too comfortable with “The Cloud”, doing their valuable computing under glass, terminally online.

Just as computers themselves went through a miniaturization phase, AI needs to do the same thing, and maintain, increase its utility and accuracy in so doing. The only way to improve in all of those aspects at once is to implement a coherent sublayer for interpreting language, a kernel, and to be able to reduce the schema of the language to numerics entirely at the hardware level. With a Natural Language Kernel, the hardware requirements necessary to run a competent LLM would drastically decrease.