What?
AIs, or artificial intelligences, are currently separate, non-conscious [1], data-driven [2], and semi-autonomous [3] systems capable of interacting in the real world. They can include LLMs [4], or Large Language Models [4], or VAs [5], which are virtual assistants [5], but *not less than that*.
How?
They are being created by processing vast amounts of data by human elements [6] engaged in fact-checking [6], visualization [6], and contextualization [6] in a pre-training environment [7], followed up by automated training via production platforms that apply the pre-training data [7].
The process from inception to a finished product, which is an A.I. created specifically to perform a vast array of functions within a particular or multiple areas of expertise [8], is an incredibly complex one, with multiple timeframes that need considering, but is generally on a track of being sped up given rising technologies and the unavoidable human involvement in the process [9].
The Process?
Breaking it all down in a simple set of timeframes from inception to finished product:
Step 1
Conceptualization
The primary stakeholder, or creator, in other words, the initial individual(s) that envisions an artificial intelligence to perform certain tasks. These visionary individuals are typically interested in virtual assistants but would like to aim for the development of a more advanced construct than a simple question and answer machine.
Step 2
Resources
Once a general form and function, or purpose, of the A.I. is envisioned, the primary stakeholder will seek other stakeholders and the organizations necessary to achieve the creation of this A.I.
Step 3
Action Plan
An action plan can take place via the concerted effort of all stakeholders, or if the task is undertaken by a single individual, then it is entirely funded by that individual, and the action plan can be directed by the influence of such a multifaceted creator. In other words, the action plan can comprise the objectives and wishes of multiple people or of a single, calculated individual.
Step 4
Creation Platforms
Once agreement on an action plan is in place, that action plan can be applied in a competitive bidding or a selection process for the organizations necessary to actually train and create the AI.
Step 5
Pre-Training
Pre-training is not training. Pre-training, or the accumulation of training data, usually involves two techniques: RLHF and SFT. Both of these techniques involve human interaction with the model being created, which aims to help vet the accuracy of the model. The pre-training data is usually stored in a database, which can be fed to the automated frameworks that follow.
Step 6
Training
Automated frameworks and AI training libraries allow for the possibility of creating an AI by using the training data, which could number in the millions of individual single-turn or multi-turn conversations between a human and an AI. The single conversation exists as a database item; in a database that contains these training batches, this database can be uploaded to the server or computer housing the data model for the AI, which themselves run libraries like Pytorch and Py Lightning. These are code libraries that facilitate the automation of this training process.
Step 7
Iterations
Creating a production-ready AI involves producing several pre-training batches of data through RLHF or SFT techniques, with the goal of each subsequent batch being improved over the previous batches. This continues until the maxim of 99.99% accuracy is reached, which is the true 100% truthfulness quotient. 0.01% represents the unforseeable.
Step 8
The A.I.
After the maximum of data is achieved via several iterations, that final database becomes the memories of the A.I., which in turn represents it’s functional brain. At this point, the A.I. is fitted with an interface, one that now includes multiple media input and outputs.
Step 9
R.O.I.
The return on investment will be determined by the costs of sustaining the A.I. with its interface on a server or serverless architecture, plus its evolution via pre-training/training processes that produce new maximums of data for its memory core, plus how well laid out and specific the action plan was, with a clear end result to produce something that satisfies a true market and a true need.
Step 10
The Future of the Newborn AI
Once the AI is live and performing its purpose, be it either helping a scientific lab with difficult calculations that would take days for a human to perform, or be it helping an individual craft a recipe for a meal or even guide a missile to its destination while outmanouvering defenses, regardless of its purpose, the process that produced it can be said to be foolproof, just as thorough, just as accurate as sending a man into space, as overcoming fusion reactions, or as attempting to create something new from virtually thin air. The future of your AI is here; it has arrived and has become the present.
Conclusion
Only as good as our minds
Since humanity lacks the ultimate answers, it is not possible to expect the AI to provide you with logical and sensible answers to the greatest questions of mankind. Such as: Where do we come from? Where are we going? Is there a God? and so on… Therefore, your R.O.I. and the lifespan of your A.I. will be determined by how good the action plan and, naturally, the training data are. However, the AI may be capable of helping us answer those questions. In conclusion, it will only be as smart and capable as its creator, never better. It is and will continue to be limited by the universe, just like every other existence, in the exact same manner. It will die, eventually, with the brain(s) that created it.
[1] Non-conscious: The term, in the context of an A.I., means that the central brain unit does not possess awareness but is functional.
[2] Data-driven: Any central brain that requires data to function and fulfill its purpose.
[3] Semi-autonomous: Implies that the form in question is lacking full autonomy, or independence. In the case of an AI, for example, it lacks a physical body typically.
[4] LLMs: Large Language Models are simply a conglomerate of databases run on physical hardware that have been populated with data created via AI pre-training techniques, often involving human interaction, described below. The human interaction aspect of these systems and their comprehensive understanding of human language—how we talk and thus how we think—is what gives them their moniker, LLM.
[5] VAs: Virtual assistants are less complex AIs that are typically focused on performing a core set of tasks really well but may be more flawed than a complete AI. This is because they don’t need nearly as many pre-training iterations as a full AI.
*Not less than that:* AIs cannot be mere applications who can also perform a variety of tasks well, but they usually cannot have an actual conversation with a human being and cannot perform fact-checking or contextualization. They are statically designed and rely on dynamic data, which is very specific. For example, social media platforms are essentially web-based applications like that. AIs can however exist within an application, with their own dedicated interfaces, such as a conversational AI assistant on a website.
[6] Fact-checking, contextualization, and visualisation: These things are enforced by a human in a topical conversation that can last up to, sometimes, 30 turns or more (or question or answer phases). The aim of the conversation is exactly that: to determine if the response given by the AI is true, if it fits the context of the conversation and is not nonsense, and is visualized if necessary if video, audio, or image media is involved.
[7] Pre-training and training: The whole process of training an AI, given its complexity and timescale, needed to be broken down into two stages, one involving a human element since it would be reductionist and incorrect to have the AI fact-check itself initially, and the other a fully automated way to “feed” this fact-checked pre-training data to it, which comprises its training process.
[8] Singular or multifaceted: Depending on the scope and action plan for the AI by its creator(s), the AI could be trained in handling a single expert-level topic, for example, physics, or many expert-level topics, again depending on the needs of its creator(s).
Leave a Reply