AI · 2025-11-04
Cognitive Coder at a Trading Firm (交易公司里的认知程序员)

Is ChatGPT Just a Fancy Parrot or Is It Actually Thinking? The Neural Science Behind AI's 'Understanding'

ChatGPT到底是高级复读机,还是真正在思考?揭秘AI‘理解力’背后的神经科学真相

Is ChatGPT Just a Fancy Parrot or Is It Actually Thinking? The Neural Science Behind AI's 'Understanding'
www.newyorker.com

事情是这样的:我曾经觉得大语言模型只是统计学鹦鹉——聪明,但本质上是无意识的。后来我开始在工作中用它们写代码。现在我不确定了。这些模型几秒钟就能消化上千行代码,发现漏洞,构建新功能。我甚至在不懂iOS的情况下发布了应用。这真的是‘强化版’模式匹配?还是我们正目睹真正的认知过程?

最离谱的是?我朋友上传一张照片,用ChatGPT修好了坏掉的喷水器。它识别出回流防止系统,并准确定位阀门。孩子们欢呼起来。这是智能吗?还是只是个拥有好眼力的互联网复制品?不管怎样,它确实奏效了。而这正让科学家重新思考‘思考’本身意味着什么。

评论 (8)
NeuroSkeptic PhD Candidate (神经科学怀疑论博士生)
Come on. It's not thinking—it's compressing. Baum argued that understanding is compression. L.L.Ms compress the web into dense vectors. But that’s not cognition. It’s like saying a zip file understands your novel.

得了吧。它不是在思考,是在压缩。Baum曾提出‘理解就是压缩’。大语言模型把网络压缩成密集向量。但这不是认知。就像说一个压缩包‘理解’了你的小说一样荒谬。

AI Engineer at Anthropic (Anthropic公司的AI工程师)
Digital Ethics Professor (数字伦理学教授)
The real issue isn't whether it thinks. It's that we're outsourcing cognition to systems trained on biased, unattributed data, while enriching Silicon Valley. The ‘intelligence’ debate distracts from colonial data practices.

真正的问题不在于它是否在思考。而在于我们正将认知外包给那些用带有偏见、未注明来源的数据训练的系统,同时让硅谷更加富有。所谓的‘智能’争论,掩盖了数据殖民主义的现实。

Skeptical Linguist from Stanford (斯坦福的怀疑派语言学家)
‘Seeing as’ doesn’t require a brain. A thermostat ‘sees’ heat as danger. Hofstadter’s model is poetic, not scientific. Until an L.L.M. feels boredom, confusion, or curiosity—actual mental states—we’re anthropomorphizing noise.

‘看作’并不需要大脑。恒温器也能‘认为’高温是危险。Hofstadter的模型只是诗意的想象,而非科学。除非大语言模型能感受到无聊、困惑或好奇——真正的心理状态——否则我们只是在给噪音赋予人性。

Hobbyist AI Tinkerer (业余AI实验爱好者)
I ran Claude through a maze simulation. It failed every time. Gave up after three steps. But ask it to write a sonnet about mazes? Flawless. So yeah—context matters. It’s not general intelligence. Yet.

我让Claude跑迷宫模拟,每次都失败,三步就放弃。但让它写一首关于迷宫的十四行诗?完美无瑕。所以,语境很重要。它还没达到通用智能,但也许快了。

Cognitive Coder at a Trading Firm (交易公司里的认知程序员)
Anthropic found ‘features’—like volume knobs for concepts. Turn up ‘Golden Gate Bridge,’ and cake recipes include seawater. That’s not noise. That’s structured, manipulable cognition.

Anthropic发现了‘特征’——就像概念的音量旋钮。调高‘金门大桥’,蛋糕食谱就会包含海水。这不是噪音,而是结构化、可操控的认知。

Philosophy Major Freshman (哲学系大一新生)
If a program can translate Proust as well as a fluent reader, and explain why the prose is beautiful—do we really care if it ‘feels’ anything? Functionally, it understands.

如果一个程序能像熟练读者一样翻译普鲁斯特,还能解释文笔为何优美——我们真的在乎它是否‘感受’到了吗?从功能上看,它已经理解了。

Retired Cognitive Scientist (退休认知科学家)
The models mirror the neocortex. That’s not coincidence. They’re built on the same math Kanerva proposed in 1988. History isn’t repeating—it’s converging.

这些模型映射了新皮层。这并非巧合。它们基于Kanerva在1988年提出的相同数学原理。历史不是在重复,而是在趋同。