3 ms·
Who? It has been known that LLMs cannot reason transparently nor can these black-boxes explain themselves without regurgitating and rewording their sentences t
by rvz 3y ago
Who?
It has been known that LLMs cannot reason transparently nor can these black-boxes explain themselves without regurgitating and rewording their sentences to sound intelligent, but instead are confident sophists no matter what random anyone tells you otherwise.
EDIT: This is the context before it was deleted by the grandparent comment:
>> i have yet to see any ai system properly implement individual level-adjusting tutoring. i suspect because the LLM needs a proper theory of mind (https://twitter.com/swyx/status/1697121327143150004 https://twitter.com/swyx/status/1697121327143150004) before you can put this to practice.
My point still stands.
- gojomo 3y agoEmphatic assertions "it has been known" are anti-convincing.
- BoorishBears 3y agoYou're showing why I'm so annoyed by this perfectly! It's malicious to rope theory of mind into justifying that point because it's just wrong enough. If the reader doesn't think deeply about why on earth you would ever to rope theory of mind into this, your brain will happily go down the stochastic parrot route: "How can it have theory of mind, theory of mind is understanding emotions outside of your own, the LLM has no emotions" But that's a complete nerdsnipe. — If instead you distrust this person's underlying motivations to not be genuine intellectual curiosity, but rather to present a statement that is easily agreed to even at the cost of being wrong... you examine that comment at a higher level: What is theory of mind adding here besides triggering your typical engineer's well established "LLMs are over-anthropomorphized" response? Even in psychology it's a hairy non-universally accepted or agreed upon concept! Theory of mind gives two things at the highest level: inward regulation: which is nonsensical for the LLM, you can tell it what emotion it's outputting as, it does not need theory of mind to act angry outward recognition: we've let computers do this with linear algebra for over 2 decades. It's what 5 of the largest companies in technology are built on... — Commentary like that accounts is built on being just wrong enough: You calmly state wild opinions. There are people who want to agree with any calm voice because they're seeking guidance in the storm of <insert hype cycle>. They invent a foothold in your wild statement, some sliver of truth they can squint and maybe almost make out. Then you gain a following, which then starts to add a social aspect: If I don't get it but this is a figure head, I must be looking it wrong. Now people are squinting harder. This repeats itself until everyone has their eyes closed following someone who has never actually said anything with any intention other than advancing their own influence. They don't care how many useful ideas die along the way, there's no intellectual curiosity to entice them to even stumble upon something more meaningful, it's just draining the energy out of what should be a truly rewarding time for self-thinking.