PwnallthethingsSo these AI chats are pretty cool.
One thing that's gonna cause some problems tho is that they sound really really convincing while sometimes being extremely extremely wrong
A zillion ways this revolution is going to be great.
One way that's going to suck tho is in places where human interactions are a *useful* friction / proof-of-work. When it's not, those systems will get really overloaded
Few examples:
1. Online troll bots & fake personalities
It's going to get a *lot* harder to distinguish bots from people and much easier to create entirely fictitious credible online personalities to troll/harass/do crime
2. Persuasive letters by e.g. constituents to regulators.
Volume of (~sensible, unique) letters was a valid indication of sentiment. Soon won't be.
3. Ransomware victim communications & negotiations
Used to be one of the few costly areas in scale. Not for long
4. Places where writing is assessed for cash or credit
Going to be much easier for e.g. students to fake/plagiarize essays or other homework for-credit
Combatting that will probably eventually force schools and universities towards towards more heavy focus on proctored work (exams, supervised essay writing) (ugh)
Going to be much easier to create fake-news clickbait farms too for ~ the same reason
What worries me a bit here is that in trying to preserve some of these systems, there'll probably be a lot of folks pushing tech designed to *prove* humanness to combat it. And politicos and regulators are less likely to hear that and think "guess they should use creative CAPTCHAs" and more likely to think "oh, people should upload passports", and that's gonna be a whole new shitshow
Also excited for the new era of "who is liable when the AI tells you to cook the frozen Turkey in the deep-fryer" / "can you maliciously interact as user A with the AI, such that the AI interacts with user B and advises them to do X" wars
Andrei Kucharavy@Pwnallthethings you don’t want to know how easy are those models to hijack, especially on niche topics. At some point on Twitter I had a thread of Google search results that had artifacts due to errors in parsing or short-term fads that affected search results in a durable manner by creating an association between terms.
Given the linguistic specificity of different population sub-groups, you can also target the hijack to a specific sub-group you might need to influence/hurt/…
Given the linguistic specificity of different population sub-groups, you can also target the hijack to a specific sub-group you might need to influence/hurt/…