AI Is a Terrible TherapistInvestors are betting that AI chatbots can treat people who are struggling with mental-health issues and unable to access professional care. But the way large language models are currently trained makes them sycophantic, and thus fundamentally incapable of providing effective treatment.
So yet another job LLMs will royally screw up.
Why was the design decision made to make LLMs sycophantic? I'm guessing, because most people enjoy having some smoke blown up their ass. So they will come back for more. And pay more $$$ to the AI company.
Yet again, capitalism ruins everything.
Currently in my ongoing Bullshit project, I am reading & annotating "On Truth", Harry Frankfurt's follow-up to "On Bullshit". Mercifully short, but still a slog.
I realized while walking the other day that one of the root causes of bullshit fatigue is that all of this philosophical reading is just words, words, and more words. Very few pictures. No graphs or charts. No tables. Very few diagrams. I even wouldn't mind a code sample.
And I think that bullshit fatigue leads to bullshit contagion. Bullshit wears you out, you lapse into writing bullshit.
And I think that the fact that LLMs are mostly creatures of nothing but words is why they are a complete dead end on the road to AGI. I don't think you can ever get CommonSense™ from words alone.
Here's the home page/directory for my posts on Bullshit.
No comments:
Post a Comment