My date with an octopus - FT中文网
登录×
电子邮件/用户名
密码
记住我
请输入邮箱和密码进行绑定操作:
请输入手机号码,通过短信验证(目前仅支持中国大陆地区的手机号)。如果该手机号尚未注册,将自动创建 FT中文网账号。短信可能需要几分钟送达,验证码15分钟内有效,请耐心等待:
请您阅读我们的用户注册协议和隐私权保护政策,点击下方按钮即视为您接受。
卧底经济学家

My date with an octopus

Whatever the request, ChatGPT is happy to help, often with unfortunate results
© Alberto Miranda
00:00

{"text":[[{"start":4.6,"text":"A funny thing happened to me this week. After trusting a dating app to arrange dinner with a suitably vivacious and intelligent lady, I arrived at the restaurant at the appointed time to find that in fact my date was with an octopus."}],[{"start":20.689999999999998,"text":"For the avoidance of doubt, everything in the paragraph above is untrue. I am not on the dating market, there was no octopus and nothing funny ever happens to me. Nevertheless, I typed this scenario into the latest offering from ChatGPT, asked why it had sent me on a blind date with an octopus, and demanded an apology."}],[{"start":42.19,"text":"“I owe you both an apology and an explanation — and possibly a towel,” ChatGPT began, despite the fact that I had never asked it for any dating advice in the first place. “You dressed up, you made the effort, and you deserved a romantic dinner — not a cephalopod-related debacle.”"}],[{"start":61.47,"text":"ChatGPT went on to explain why it had made the mistake — a weak grasp of “human courtship norms” — and in its defence pointed out that the octopus was intelligent and vivacious, and “left saying it was the best date she’d had in years”. Which, in fairness, is not a bad line. ChatGPT finished by offering to draft a “lessons learned” report and a formal apology to the restaurant. (The apology isn’t bad either: “While my guest, ‘Octavia’, displayed considerable intellect and curiosity, I now appreciate that these qualities do not mitigate the disruption caused to your other patrons, your wait staff, or your fish tank . . . ”)"}],[{"start":103.66,"text":"Janelle Shane is the author of You Look Like A Thing And I Love You, a book about how neural networks succeed and fail. She has recently demanded that ChatGPT apologise to her for advising her to trade her mother’s cow in exchange for some magic beans, and for releasing an army of cloned T-Rexes into Central Park. The responses are deft pieces of improv comedy."}],[{"start":128.95,"text":"This, like so many things Generative AI can do, is both impressive and a bit weird. It is also instructive. Improv is all about accepting the premise: taking whatever is thrown at you and building on it."}],[{"start":142.19,"text":"A computer which responded “I have never arranged a date for you, octopus or otherwise” would be a terrible improv partner. However, in every other situation I can imagine, that would be a more appropriate response to a demand for an octopus-date apology."}],[{"start":160.97,"text":"What role does the AI think it’s playing? Confusion over that question can cause serious headaches surprisingly quickly. I recently asked ChatGPT-o3 for help with a research question. I dimly remembered a story told by the moral philosopher Jonathan Glover — probably, I thought, in Glover’s book Humanity — about a Nazi bureaucrat haggling over the fee for slave labour, punctiliously fussing over petty financials and ignoring the grotesque human cost. I wanted to find the details."}],[{"start":195.87,"text":"The computer was happy to help: the story in question concerned the Buna-Monowitz works, the argument concerned pay rates for prisoners who were sick or who died half way through a shift, and the details could be found on pp288-292 of the first edition or pp300-304 of the second edition."}],[{"start":217.11,"text":"This seems to be incredibly impressive work, except that ChatGPT was still in improv mode. When I checked Glover’s book, I realised ChatGPT had invented it all. I found the story in question but I had misremembered the details and ChatGPT had fabricated them with exactly the same commitment and mental agility that it had fabricated an apology for a date with an invertebrate. Suddenly, the improv is less than hilarious."}],[{"start":246.65,"text":"AI researchers have long worried about what they call the “alignment problem”, the question of whether AI systems (and algorithms more broadly) will do what we want them to do, or somehow misunderstand our true goals."}],[{"start":261.75,"text":"There is a long tradition of this in our stories and legends, from the unhappy King Midas, who wished for the golden touch but turned his food and drink and even his own daughter into gold, to the malevolent monkey’s paw. In the famous WW Jacobs short story, a man who wishes for £200 on the monkey’s paw receives the money shortly afterwards as compensation when his son dies in a workplace accident."}],[{"start":289.23,"text":"Jack Vance’s masterful fantasy trilogy Lyonesse offers the supernatural servitor Rylf, instructed by the wizard Murgen to follow an enemy who had shape-shifted into a moth. Rylf did so, but the moth-shaped enemy soon found a flaming torch “where it joined a thousand other moths, all careering around the flame, to Rylf’s confusion.” Rylf had superhuman powers, but alas, no common sense. His instructions were to pursue the shape-shifted enemy, and yet, “As he waited . . . one of the moths dropped to the ground and altered its form to that of a human man . . . By the laws of probability, as Rylf reckoned them, the moth of his interest remained in the throng.”"}],[{"start":335.07000000000005,"text":"There are so many ways to offer catastrophic compliance, whether maliciously, like the monkey’s paw, or through a lack of judgment, like Rylf, or because the instruction itself is confusing. You and I might think it is obvious that the request for an octopus apology cannot be taken seriously, while the request for help tracking down a story about the holocaust cannot be taken lightly. The machine, like Rylf, may see things differently."}],[{"start":366.26000000000005,"text":"It may be that such problems will soon be fixed. When I copied my Jonathan Glover request into the latest model, ChatGPT-5, it began with a vague fabrication before pivoting hard towards the truth: “Unfortunately, I couldn’t find the exact phrasing online . . . I recommend checking in your own copy of Humanity.” Much better. Not actually helpful — but far less harmful than the previous invention."}],[{"start":392.46000000000004,"text":"As for the confident bullshitting of GPT-o3, what to do? I decided to play to its strengths. I asked for an apology and an explanation."}],[{"start":404.29,"text":"Find out about our latest stories first — follow FT Weekend Magazine on X and FT Weekend on Instagram"}],[{"start":417.61,"text":""}]],"url":"https://audio.ftmailbox.cn/album/a_1756161896_6121.mp3"}

版权声明:本文版权归FT中文网所有,未经允许任何单位或个人不得转载,复制或以任何其他方式使用本文全部或部分,侵权必究。

太空探索技术公司转向太空之外

星际基地,我们遇到麻烦了。

瞄准OpenAI和Anthropic的廉价新AI模型

初创公司TypeSafe AI的“Jev”模型承诺为开发者提供更快速、更高效的AI。

没有产品也没关系:投资者押注AI新型实验室

的确,这些数字可能掩盖了一些障眼法。

唐纳德•特朗普押注AI,Maga阵营退缩

美国总统对大型科技企业的追捧,正在疏远曾经忠诚的支持者;距离中期选举仅剩数周。

从1998年走到现在,他还在回家的路上

从智利南部出发近30年后,卡尔•布什比即将回到英国。他是日益壮大的极限旅行者群体中的一员。

我们该如何应对AI的金融风险?

公众视线之外,监管官员正试图用AI来管AI。
设置字号×
最小
较小
默认
较大
最大
分享×