Was watching a show on Netflix last night about an AI app for your spouse to keep connected with you after death. It scraped data from your social media, email, browsing history etc to create the digitial AI version of me that would live on. I’m not sure which version my wife would like better.
Life has a way of teaching we wee humans that whenever the original premise contains this or that flaw, workflows dependent upon that fractured foundation will inevitably, ineluctably, and invariably find a way to crumble the infrastructure.
The available evidence of several decades pasts strongly suggest that so-called ‘tech giants’ - aka ‘FAANG’ - which have so-warmly embraced AI have yet to learn that ever-present lesson; for my own part, I would posit that they are unlikely to ever do so.
The prospects do not appear to be improving, as this eloquently-stated NSFE discussion - one of the many that have appeared in recent weeks re: the Offensive Products-policing Amabot - starkly illustrates (emphasis mine):
The few times I tried to use AI for product descriptions, it fluffed them up over the top. Maybe these Federal employees should use AI to generate their bullet points.
The order settles allegations that Workado promoted its AI Content Detector as “98 percent” accurate in detecting whether text was written by AI or human. But independent testing showed the accuracy rate on general-purpose content was just 53 percent, according to the FTC’s administrative complaint. The FTC alleges that Workado violated the FTC Act because the “98 percent” claim was false, misleading, or non-substantiated.
It’s because AI is mostly use as an agentic task execution bot. It has to reconcile what it knows, what you asked it to do and often without much context. The magic happens when you stop using AI to perform tasks but to iterate a vast array of data that is part of the larger set of data - a subset that revolves around: your thoughts, your views, your iteration, your interests, your aims and objectives - that is building a general ethos - then another subset could be tasks that you’ve completed together successfully to satisfaction with multiple iterative revisions. a second subset - if you continue to build these subsets - and you combine them as a total subset of the data it has - it will open up many doors. It’s not the LLM that is the issue (or rather that is only part of the issue - some ar better than others) - its because people are using it incorrectly and asking it to guess what you want, which it can’t do because it’s not telepathic (yet).