TE
科技回声
首页24小时热榜最新最佳问答展示工作
GitHubTwitter
首页

科技回声

基于 Next.js 构建的科技新闻平台,提供全球科技新闻和讨论内容。

GitHubTwitter

首页

首页最新最佳问答展示工作

资源链接

HackerNews API原版 HackerNewsNext.js

© 2025 科技回声. 版权所有。

Ask HN: Why is GPT-4 or Claude-2 so bad at tic-tac-toe?

2 点作者 zurfer超过 1 年前
I was surprised to learn that GPT-4 can&#x27;t play tic-tac-toe, but thought people who tried just didn&#x27;t prompt it correctly.<p>But after trying for 2h to get it work (even with GPT-4V) it seems like a fundamental limitation.<p>I&#x27;ve found a HN submission [1] where someone used a brute-force prompt to get it to play correctly, but as the top&#x2F;only comment points out, it&#x27;s a limited action space and enumerating most of it seems moot.<p>I was hoping for a more reasonable prompt. After all humans are able learn tic-tac-toe rapidly.<p>current hypotheses:<p>1) tic-tac-toe requires &quot;spatial reasoning&quot; and LLMs train on sequences (somehow GPT-4V didn&#x27;t elevate that constraint)<p>2) tic-tac-toe requires &quot;search&quot; of future scenarios<p>Would love to hear what you think&#x2F;know!<p>---<p>Previous discussion about T3 and GPT-4: https:&#x2F;&#x2F;news.ycombinator.com&#x2F;item?id=35216614 (7 months ago)<p>[1] https:&#x2F;&#x2F;news.ycombinator.com&#x2F;item?id=37626918

5 条评论

outdig超过 1 年前
I agree with jqpabc123 except for one point: that it cannot &#x27;reason&#x27; and it is not smart. I disagree with this.<p>The reason it cannot play well is because it has very little &#x27;experience&#x27; (training data) with it. It&#x27;s been trained on &#x27;what the game is&#x27;, it has not been trained how to win.<p>You can think of it a bit like driving. You can know what driving is, but it doesn&#x27;t make you a driver if you&#x27;ve never driven before.<p>You can ask a genius who&#x27;s never played before to play tic tac toe with you, tell it the rules, they will likely not win on the first attempt or play optimally. This doesn&#x27;t mean that person isn&#x27;t a genius.<p>You said humans are able to &#x27;learn&#x27; to play it rapidly. So is GPT, in training mode, it can process a million games in seconds, where a human can&#x27;t.<p>The problem here, is it simply has no experience.<p>If I told you every time you played tic tac toe against me, you would forget all your experience the next time we played, would you play optimally?
评论 #38134534 未加载
sharemywin超过 1 年前
I think this paper has an interesting way of training problem solving:<p><a href="https:&#x2F;&#x2F;venturebeat.com&#x2F;ai&#x2F;microsoft-unveils-lema-a-revolutionary-ai-learning-method-mirroring-human-problem-solving&#x2F;" rel="nofollow noreferrer">https:&#x2F;&#x2F;venturebeat.com&#x2F;ai&#x2F;microsoft-unveils-lema-a-revoluti...</a><p>I submitted to HN but nobody seemed to care:<p><a href="https:&#x2F;&#x2F;news.ycombinator.com&#x2F;item?id=38128012">https:&#x2F;&#x2F;news.ycombinator.com&#x2F;item?id=38128012</a><p>I looks like it basically uses GPT-4 to train a smaller model on problem solving.
jqpabc123超过 1 年前
<i>But after trying for 2h to get it work (even with GPT-4V) it seems like a fundamental limitation.</i><p>It obviously hasn&#x27;t been &quot;trained&quot; for tic-tac-toe. The way to train it is using statistics --- present every possible position and the correct response so it can build a database.<p>There is no logic or reasoning involved --- it&#x27;s all statistics. It&#x27;s not what we call &quot;smart&quot;. Any ability to &quot;reason&quot; is just a statistical illusion.
bell-cot超过 1 年前
The Turing Test sounds cool, and &quot;is he a clever conversationalist?&quot; is a fairly good test - <i>of social intelligence and class, for casual use in human society</i>.<p>But current &quot;AI&#x27;s&quot; are intelligent kinda like pocket calculators are intelligent.
david927超过 1 年前
The New Yorker wrote a nice article on LLMs and how they work, &quot;What Kind of Mind Does ChatGPT Have?&quot;