"The most authentic hipster talk of today is the twenty-three skidoo of tomorrow"
-- Ernest Hemingway
SF Bay Area, barely.
I have thoughts on AI:
https://mcguinnessai.substack.com
More a tech grandpa than a tech bro.
Charles McGuinness
Loading...
Perfect day for visiting Bees N Blooms in Santa Rosa (north of SF). Just amazing to walk along the lavender in bloom.
An LLM knows how confident it is in the next word (token) it generates, but it's not "thinking" about facts as some sort of abstraction. There's no inner life, no Dasein, if you're into that sort of thing. There's no more depth to it than in a 3D image in a lenticular card.
As humans, we know things, but we also know how certain we are of those things. You can ask me a question and then ask me how confident I am in the answer. And I will be able to tell you how confident I am. Ask me when was the Gettysburg Address was, and I'll say 1864 but maybe plus/minus a year.
Of course the contrition isn't real. It's not even really contrition. It might stick a note somewhere that says don't make stuff up, but here's the problem: It does not know it made stuff up. So how can it stop itself?
Models are getting less likely to hallucinate as they get better. But they still do, and they have neither self-awareness nor shame to hold them back.
A table saw would not notice let alone regret cutting your thumb off; it sure wouldn't learn a lesson from it.
LLMs are no different.
There are 2 possible explanations for this.
(A) is that @bencollins.bsky.social bought the FT and turned it into the UK edition of The Onion.
(B) is that Hillary is evil.
I know it's B, but why? Now's the time to whitewash your reputation, HRC, and be on the right side of justice for a while.
As much as I want to go "See???", I have to admit I'm puzzled. Flu shots are usually given in the fall, and the requirement to get them was lifted in the spring, so I'm not seeing the implied causality here.
I am becoming convinced that Anthropic, when they got ready to release (the now banned) Fable, did something to decrease how well Opus 4.8 works (by making expend less effort). What used to work well is now a struggle.
Anecdotal, I know, but it's so consistent.
You can pretend to agree, but you can't pretend to be there.
Thought for sure there was going to be an Aussie involved in this debate.
