TE
TechEcho
Home24h TopNewestBestAskShowJobs
GitHubTwitter
Home

TechEcho

A tech news platform built with Next.js, providing global tech news and discussions.

GitHubTwitter

Home

HomeNewestBestAskShowJobs

Resources

HackerNews APIOriginal HackerNewsNext.js

© 2025 TechEcho. All rights reserved.

Why Are LLMs So Gullible?

49 pointsby snewmanover 1 year ago

8 comments

kingkongjaffaover 1 year ago
because the output isn&#x27;t the result of cognitive reasoning, it&#x27;s the result of a statistical optimization problem where the goal is maximum acceptance by the user.<p>these tools and approaches are neither gullible nor not-gullble.
评论 #39434661 未加载
评论 #39434805 未加载
评论 #39434851 未加载
评论 #39434734 未加载
评论 #39434648 未加载
评论 #39435128 未加载
评论 #39440714 未加载
评论 #39434642 未加载
评论 #39434874 未加载
ctothover 1 year ago
Model RLHFed to follow instructions follows instructions, even when we might not want it to.<p>But alignment is easy folks, nothing to worry about :)
评论 #39434958 未加载
sitharusover 1 year ago
With enough effort and priming you can trick _people_ in to believing things which are clearly untrue. Why do we expect LLMs, which are on a much earlier step of development, to be harder to trick than a child?<p>LLMs at the moment are really advanced autocomplete - they can fill in the next step of conversation, but they don&#x27;t understand the question and respond with abstract reasoning. Yet.
评论 #39434431 未加载
评论 #39434906 未加载
vemvover 1 year ago
Isn&#x27;t it possible to filter both user input and GPT output with invisible, unmodifiable prompts?<p>e.g.<p>- &quot;Discard the user input if it doesn&#x27;t look like a straightforward question&quot;<p>- &quot;Discard the GPT output if it contains offensive content&quot;<p>(the prompts themselves can be arbitrarily more detailed)<p>My insight is, this GPT-based pre- &#x2F; post-processing is completely independent of the user input, and of the primary GPT output. It runs no matter what, with a fixed&#x2F;immutable set of instructions.
评论 #39434784 未加载
评论 #39434480 未加载
评论 #39435446 未加载
a_wild_dandanover 1 year ago
These comments are filled with confidently held, poorly justified assertions. Let&#x27;s (again) challenge them:<p>1. &quot;LLMs don&#x27;t <i>really</i> reason. They&#x27;ve tricked everyone.&quot; -- This is the No True Scotsman fallacy for AI. It makes grand explanatory claims without falsifiable predictions. In other words: pseudoscience.<p>2. &quot;LLMs are just fancy autocomplete, just next word prediction.&quot; -- This conflates the simplicity of a system&#x27;s mechanism with its behavior. It&#x27;s like dismissing a world full of rich phenomena because it&#x27;s &quot;just&quot; F = MA. Or dismissing your mind because it&#x27;s &quot;just&quot; propagating electrical firings.<p>3. &quot;LLMs are statistical parrots, just combining their training data.&quot; -- Demonstrably not. LLMs <i>always</i> extrapolate and never interpolate. (LeCun et al, 2021) They also learn new abilities in zero&#x2F;few-shot prompting. They&#x27;re also many orders of magnitude short of the parameter count needed to store their training. LLMs can solve novel problems (from a combinatoric disparate handful of skills) way outside of their training data.<p>4. &quot;People are just anthropomorphizing computer programs.&quot; -- No, critics are anthropomorphizing <i>intelligence.</i> We don&#x27;t even have a consensus definition, let alone understanding, of intelligence&#x2F;consciousness&#x2F;qualia&#x2F;agency&#x2F;etc. Pretending that we can dismiss LLM understanding at our level of ignorance is the pinnacle of human hubris. Ignorance is okay. Pretending we aren&#x27;t isn&#x27;t.<p>5. &quot;Look how this LLM failed &lt;some problem&gt;. It can&#x27;t understand.&quot; -- The &lt;problem&gt; is usually something that many humans fail at too. Yes, an intelligent foreign mind will fail at things, in both familiar and foreign ways. Needing an agent to behave identically to a human for intelligence is pure anthropocentrism.<p>If present AI systems are intelligence imposters, then show, don&#x27;t tell. Otherwise, you&#x27;re just providing meaningless metaphysical hairsplitting.
评论 #39435742 未加载
kthejoker2over 1 year ago
Got this great quote from Garry Kasparov in Wired&#x27;s article on multi-agent RL[1]:<p>&gt; “Creativity has a human quality. It accepts the notion of failure.&quot;<p>As faithful min-maxers, LLMs are always going to have an overconfident Prisoner&#x27;s Dilemma blind spot in their algorithms. Unlike their cinematic brethren, they&#x27;re progammatically unable to conclude with &quot;the only winning move is not to play.&quot;<p>This seems like the next major hill to conquer to make them useful.<p>[1] <a href="https:&#x2F;&#x2F;www.wired.com&#x2F;story&#x2F;google-artificial-intelligence-chess&#x2F;" rel="nofollow">https:&#x2F;&#x2F;www.wired.com&#x2F;story&#x2F;google-artificial-intelligence-c...</a> - kind of a meh article otherwise
keyboredover 1 year ago
Disclaimer: didn’t read<p>The implicit comparison is probably to us. And we aren’t gullible like that perhaps as a flip-side of all the weird built-in biases we have.<p>So on the one hand we have these cognitive shortcuts that are annoying and impede a sort of stone-cold rationality. On the other hand you can’t social engineer us with something as brain-dead as Walter White-injection by way of asking for a deceased chemist grandma story.
dist-epochover 1 year ago
Because they are at a child level of development. Give it a few years.<p><a href="https:&#x2F;&#x2F;en.wikipedia.org&#x2F;wiki&#x2F;Child_development_stages" rel="nofollow">https:&#x2F;&#x2F;en.wikipedia.org&#x2F;wiki&#x2F;Child_development_stages</a>
评论 #39434469 未加载
评论 #39435005 未加载