The difference, of course, is dictionaries existed for centuries as a physical tome before it was evolved into a digital application. Wikipedia is an (imperfect) extension on the idea of the encyclopaedia (its technical bridge the likes of the old Encarta) — which, as well, existed for generations before the internet came about.
(And yes, it is handy, mass-wise, to have a dictionary, especially the
OED, on one’s device as an application, without a need to carry a book along, and also handy to not have to lug an entire Encyclopaedia Brittanica or World Book along for the ride.)
LLMs, however, enjoy no analogue precedent — unless by
precedent, we mean to say human beings.
For things like recreating the
Earth’s dynamic features, digitally, in real time, an LLM-based application is indispensable and necessary. For mining long-shuttered troves of data for pattern recognitions which human minds, even the sharpest, are rarely able to pick out, a LLM assigned to this is also indispensable. A good example is identifying
all possible proteins RNA could template into synthesis. Also, predicting and tweaking city-wide traffic patterns, to better foresee ways to move people about more efficiently and quickly, no matter the hour, is a place where LLMs can help transportation planners and civil engineers.
But having a LLM “app” on one’s desktop supplants or replaces no tool which already exists, and as such merits being an optional/opt-in component of an OS — not something from which one is unable to opt out.
That bridge is one which LLMs have yet to cross.
There is still an inherent
bias problem and a
drift problem with LLMs which makes neutral fact-checking a goal, but one which hasn’t arrived to 2024. Until then, we may expect more of
these stories.
Welp.
Unless there’s a way for an LLM to be programmed without even a faintest whiff of the developer or development team’s own human bias, then this is a Utopian idealization with little hope of realization.
Which means taking LLMs away from a proprietary/closed/black box realm to a completely transparent, examinable, and juried realm in compliance with both published standards (e.g., IEEE) and published protocols/deployments (e.g., ISO).
OK. This is a technical challenge, unrelated to the three social-oriented points preceding it.
Of course, a public —
public — LLM is just that: public, like a public utility (like hydro/electricity infrastructure, water/sewage/reservoir infrastructure, transportation infrastructure, and the physical/logical internet).
What I don’t see right now are any of the leading LLM/AI principals suggesting that’s where LLM development should, ultimately, be headed — not so long as Bay/Wall/Bond Street are permitted any leverage in that discussion.
So again, a compulsory LLM client, one dependent upon proprietary and/or for-profit LLMs, integrated into an OS — when that LLM is still chock-full of the above issues and revenue-adjacent liability — is really inventing and putting the cart before the horse (in ecosystems where the horse hasn’t yet been introduced).
As such, it has no place being integrated into an OS — unless of course the objective of that integration is to amplify the aforementioned problems, only at a mass user scale and to yield more negligent misinformation and conscious disinformation by bad (human) actors.