Understanding What AI Does with a Prompt and What is a Better Response

Teaching AI is Supervised Fine-Tuning.

Prompts are not going away. In Large Language Model fine-tuning, SFT helps craft more effective AI by providing tuning data in the form of Prompt-Baseline Response-Improved Response. These annotation data are exploited to power backpropagation updating of internal weights leading to more refined, more attuned, attention and feed-forward outputs within the LLM processing. A form of Reinforcement Learning (with) Human Feedback, SFT is “grad school” for rudimentary foundation models seeking improved performance, especially in quality of information and communicability in user dialog. Gain of concision and clarity is owed to the more, not the first or the few. It is consensus. Capability does not, however, confer ability. That’s a human thing.

A decade ago, SFT was data labeling, e.g. cat vs everything else image tagging, a la Fei-Fei Li crowdsourcing fame. Its evolved form today is data annotation, generating those associations. Annotators may be freelancers, working a gig likely to hang around in some form. The ranks of annotators teaching the most adept frontier models of AI will thin out. The survivors in Big AI annotation will likely be Subject Matter Experts in various aspects of the business. Fortunately, or not, “lightweight” models, including edge models, will find increasing use in smartphones, IoT devices, cars, robots, trains and boats and planes, and places we have yet to see; space? Of course. In quotidian realms, “tiny AI” will need continual grooming by ordinary users, particularly early adopters, to keep up with the evolving detail of their dissimilar applications. Tiny AI SFT may become a global cottage industry. Or else a few behemoths may corral it all. Only time will tell.

SFT is not Recursive Self Improvement.

Sadly, AI “doom trolling” has taken the idea of vanilla RSI for a wild ride. RSI is self-supervised machine learning. This is unsupervised by humans, rather by its own recursion to adjust its parameters, as guided by a cost-of-accuracy function. However powerful this schema, and it is, it may not necessarily lead to particular emergent capabilities. Emergence is, in general, the phenomenon of new behavior from an evolved, or even merely scaled-up, system. An optimistic industry vision is that “intelligence,” not to say “consciousness,” will one day emerge at high enough scaling up of a promising architecture. By adding dimensions and/or increasing the widths (of a dimension), i.e. by simply expanding machine complexity, it will just “happen.” Good luck with that. One possible antidote to the hysteria about an emerging Superintelligence as capable of “evil” as the good, is Cal Newport’s blog. Not saying Cal is right about everything. Just saying here’s a calm assessment amidst the chaos. You could do worse.