OpenAI 的重大调整 + 医生诊室中的 AI + Talkie，一款前 1930 年代的大型语言模型

也告诉我们你是否同意我们的选择。

And let us know if you agree with our picks.

我猜你不会同意。

I bet you won't.

凯文，屋里有医生吗？

Kevin, is there a doctor in the house?

有啊，凯西。

There sure is, Casey.

今天我们要

Today, we are going

和一位医生聊聊人工智能和医学，因为最近这个领域发生了太多事，我们得请个专业人士来给我们做个梳理。

to have a conversation with a doctor about AI and medicine because this is an area where there has just been a lot happening recently, and we needed someone qualified to come in and debrief us.

对。

Yeah.

过去几个月，我们纵观整个领域，发现一家接一家公司都推出了AI与医学结合的产品。

You know, as we've sort of looked across the landscape just over the past few months, we've seen company after company introduce their own product at the intersection of AI and medicine.

还有ChatGPT Health，专门给医生用的ChatGPT。

There's ChatGPT Health, ChatGPT for clinicians.

亚马逊搞了个叫Health AI的东西。

Amazon has something called Health AI.

微软推出了Copilot Health。

Microsoft has Copilot Health.

与此同时，医生们也在自己琢磨着用这项技术。

And, of course, all the while, doctors are experimenting with this technology.

据我们所知，他们还真越用越兴奋，觉得这玩意儿挺有料。

And as best as we can tell, actually getting really excited about what they're seeing.

是啊。

Yeah.

最近我去看病，感觉变化特别大——现在每次去看医生之前，我都会先跟AI聊一通，弄清楚自己到底哪儿不对劲。

And this has been a huge change in my recent visits to doctors, which is that I now am having this series of conversations leading up to the visit with AI systems about what is going on.

所以我去的时候脑子里已经有了一堆靠谱的信息，跟医生聊起来也更对等、更深入了。

And so I am coming armed with what I believe to be good information about what is going on, and that allows me to sort of have a different, more elevated conversation with the doctor.

这可不是我一个人这样。

And this is not just me.

现在越来越多的人会去找聊天机器人问医疗问题。

Like, people are increasingly turning to chatbots for medical information.

根据一些最新的数据，大约有三分之一的美国人会用AI来查健康信息。

According to some recent data, approximately a third of Americans report turning to AI for healthcare information.

各大公司也在拼命推出更专业的工具，专门针对医疗场景做优化。

And companies are racing to respond to that demand by making better tools that are specifically designed for use in health care.

为了帮大家理清AI在医疗领域的现状，我们请回了我们特别喜欢的一位医生——罗德曼。

So to help us make sense of the landscape for AI and medicine and health care, we invited back to the show one of our favorite doctors, Rodman.

他是贝斯以色列女执事医疗中心的内科医生，也是哈佛医学院的助理教授。

He's an internal medicine physician at Beth Israel Deaconess Medical Center and an assistant professor at Harvard Medical School.

是的。

Yeah.

我们上次采访他是在2024年11月。

We last talked to him in November 2024.

从那以后，他一直研究人们和人工智能在医疗领域中的互动方式。

And since then, he's continued to study the way that people and AI interact in the health care space.

我们对他有好多问题想问。

And we have a lot of questions for him.

比如，凯文，你这皮疹到底该怎么处理？

Like, what should we do about your rash, Kevin?

没错。

Yes.

那咱们赶紧付个挂号费，把亚当·罗德曼医生请上来吧。

So let's fork over our co pays and bring in doctor Adam Rodman.

亚当·罗德曼医生，欢迎再次做客《Heartfork》。

Doctor Adam Rodman, welcome back to Heartfork.

能来这儿真是太开心了。

Oh, it is a pleasure to be here.

我现在算不算这节目的老朋友了？

Am I a friend of the show at this point?

那咱们看看这次访谈会怎么样。

Well, let's see how this interview goes.

至少你已经是节目的医生了。

You're at least a doctor of the show.

你可是我们的家庭医生。

You are our primary care physician.

上次我们采访你是在2024年底，那时候医学界好像开始有人嘀咕：等等，有点不对劲。

So when we last talked to you in late twenty twenty four, I think this was a moment where the medical community was starting to say, wait a minute.

这些AI模型在诊断方面确实越来越厉害了，但当时整个行业还是抱着观望态度。

These AI models are getting pretty good at things like diagnostics, but I think a lot of the field was still kind of in wait and see mode.

现在快两年过去了，我们有了不少新工具，也出了很多关于AI在医学中应用的新研究。

Now, almost two years later, we have a lot of new tools and a lot of new studies about the use of AI in medicine.

所以你给我们捋一捋，过去这一年半，AI在医学领域到底发生了啥？

So just catch us up on, like, what has been going on with AI in medicine for the last call it year and a half?

是的。

Yeah.

简直太疯狂了。

It's been crazy.

在医学领域，AI的普及速度之快，按某种标准来说，可能是有史以来最快被采纳的医疗技术了。

AI in medicine has gone from well, depending on how you measure it, it's probably the fastest adopted medical technology of all time.

以前这东西还特别新潮，几乎没人用，现在却成了大多数医生每周工作的常规部分。

We went from this being super novel, almost no one used AI tools, to this being a routine part of most doctors' weekly practice.

你能给我们讲讲医生们用的AI工具都有哪些吗？

And give us a sense of, like, the AI stack for a doctor.

他们现在都在用什么工具？怎么用的？

What are the tools that they are using right now and how?

尤其是那些普通的医生，

And particularly, like, what are the mainstream doctors?

就是还没站在技术最前沿的那些人。

Like, the people that, you know, aren't yet on the bleeding edge.

普通医生？

The normies?

是的

Yeah.

你可以这么理解

If you will.

对

Yeah.

大多数普通医生用的、你的听众很可能已经接触过的主流技术，叫AI文书助手。

So the biggest sort of normal doctor technology, which most of your listeners or a a good portion of your listeners have, have encountered are what are called AI scribes.

这是一种语音转文字的算法，能听着你和病人对话，然后自动生成病历的初稿。

That's a sort of voice to text algorithm that listens to you, talk to your patients and then writes a first draft of your note.

这种工具短短两年内就从新鲜实验品变成了标配。

And these have gone from, like, kind of a novel experimental technology to commodity in probably less than two years.

到处都能见到。

They're everywhere.

医生们特别喜欢，病人也喜欢，因为医生能花更多时间跟他们聊天。

Doctors really like them, and then patients really like them because they spend more time, like, talking.

第二个比较常见的医生应用场景是辅助决策。

And then the the second sort of, I'd say, normal doctor, use case is for decision support.

有个叫Open Evidence的公司，做了一个免费工具，从零开始，短短时间就爆火了。

So there's this one company called Open Evidence that has created a free tool that has gone from, again, zero to crazy numbers of adoption.

我跟你说，像我带的住院医这些年轻医生，天天都在用。

I will tell you, younger doctors, like my residents, use it all the time.

具体数字我不清楚，但估计现在差不多有一半的美国医生都在用它。

I don't know the actual numbers, but it's probably close to half of US doctors are using this right now.

天哪。

Wow.

我看到的数据是，现在超过四成的医生都在用这个工具，才两三年时间，普及得真够快的。

So, yeah, the the statistics that I've seen are that more than 40% of doctors now are using this, which is pretty crazy uptake for something that was just started a couple of years ago back in 2022.

今年三月，Open Evidence公布过，仅24小时内，医生们就向这个AI系统咨询了一百万次。

In March, Open Evidence reported that in a single twenty four hour period, doctors consulted the AI system a million times.

我对Open Evidence特别感兴趣。

Been fascinated by Open Evidence.

我本人没用过，但我有朋友是医生或护士，他们都说得跟你一样——尤其是年轻医生，几乎人人都在用这个工具。

I've never used it myself, but I have friends who are doctors or nurses, and they have said what you've said that basically that just everyone, especially on the younger end of medicine, is just using this thing constantly.

那你能不能给我们讲讲，这个Open Evidence到底是怎么运作的？

So like, give us a sense of like how this open evidence tool works.

它一般用在哪些场景？有什么优势和不足？

Like, what situations is it used for and what are its strengths and weaknesses?

这问题问得真好。

Oh, that's a great question.

Open Evidence的运作原理其实算商业机密，但大致就是用了某种检索增强生成技术，加上一个证据检索系统。

So so how open evidence works, like, all of these tools is is a trade secret, but it uses some sort of, like, retrieval augmented generation and an evidence retrieval tool.

他们跟很多顶级医学期刊都有合作。

And they have all these deals with the big medical journals.

比如《新英格兰医学杂志》、《JAMA》这些。

So, New England Journal of Medicine, JAMA.

当你提出一个临床问题时，它会去检索相关文献，筛选出高质量的证据，然后所有回答都严格基于已发表的研究。

And when you ask a clinical query, it searches the evidence and then tries to identify high quality sources, and then it always grounds what's coming back in the literature.

像我这样头发都花白了的人，用Open Evidence就跟用谷歌搜索或者以前那些老工具差不多。

So you have gray hairs like me who kind of use open evidence the way that I would use a Google search or one of the old tools.

我拿它来快速翻文献，效率高多了，经常直接跳到原始论文，或者就图个方便，快速找个参考文献。

So I use it as a souped up way to search the literature rapidly and often go to the primary sources, or I use it as a faster way to get a reference.

比如我很久没用过的一种药，Open Evidence能直接从FDA调出药品说明书。

So a a drug that I haven't dosed in a long time, open evidence pulls the drug monographs from the FDA.

我一两秒就能调出来。

I can very quickly pull that up.

对。

Mhmm.

我注意到年轻医生更爱问这类问题：这到底可能是啥情况？

Younger doctors, I have noticed, and I don't know this empirically, but younger doctors are more likely to ask questions like, what could be going on?

你能给我个第二意见吗？

Can you give me a second opinion?

接下来我该怎么做？

What is the next thing that I should do?

所以呢，我平时不用这类决策支持或参考工具，但现在有了新的用法。

So, ways that I don't traditionally use decision support or reference tools, but sort of a new way.

当然啦，年轻医生们也会用它来查资料，就跟我的用法一样。

And, of course, younger doctors also use it in the reference ways that I do.

那他们会不会把病人的实际数据上传上去？

Now are they actually uploading patient data to this?

还是说，他们只是用一些通用的、匿名的描述来获取建议？

Like or are they are they just sort of describing patients in generic and anonymized ways to get back some decision support?

据我了解，大多数人是第二种情况。

My understanding is largely number two.

我相信公司心里有数，希望没人会把受保护的健康信息直接复制粘贴进去。

I'm sure the company has a good sense on how many people I hope no one is copying, protected health information and putting into it.

就我观察到的，不管是同事还是学生，大家基本都把它当搜索引擎用——有疑问就查。

Certainly, what I've observed from, like, my colleagues and my students, most people use it the way you would use, like, a search tool when you have a question, which is, hey.

比如说，我正给这个病人用头孢曲松。

I like, I'm giving this person ceftriaxone.

腹腔感染的话，合适的剂量是多少？

What what's the right dose for an intra abdominal infection?

所以这些都是一些比较笼统的问题，由医生来解读。

So sort of generic questions that are being interpreted through the physician.

有没有哪些AI工具是和患者病历系统直接整合的？

And are there any AI tools that are integrated with patient health records?

这方面我一直挺有顾虑的，我不希望自己的个人健康数据、受保护的医疗信息被传到那些云端AI系统里。

This has been an area where I think there's just been a lot of pushback about, like, I don't want my personal health data, my protected health data going into one of these cloud based AI systems.

但有没有医院或医疗系统已经把这类工具直接接入患者数据了？

But are there hospital systems or medical systems that are bringing this stuff directly into contact with patient data?

绝对有。

Oh, a 100%.

没错。

Yes.

目前，真正接触患者数据的AI应用，更多是用在账单处理这类事情上，而不是直接辅助医生做临床决策。

Right now, most of the sort of in contact with patient data are less about physician facing decisions and more about like billing.

有些公司正在跟电子病历系统做整合。

There are companies that are like integrating with the electronic health record.

不过这些还都不是行业标准。

Those are not standard yet.

还有那些电子病历系统本身的厂商。

And then the EHR companies themselves.

比如Epic，它是美国最大的电子病历供应商。

So like Epic is obviously the biggest EHR vendor in The US.

他们正在大力开发内置的功能。

They're doing a lot of work on building in native things.

就拿我所在的医疗系统来说，我要给病人发消息时，顶部的AI助手会建议：‘你或许该这么说。’

So for example, at my health system, if I wanna send a message to the patient, the helpful AI at the top already has like, maybe you should say this.

但这建议通常没啥用，我这辈子都没用过一次。

It's usually not that helpful, and I don't think I've used it once in my life.

不过确实有不少这类功能，正在被试验并直接集成到患者的健康数据里。

But there are a lot of those things that are are being experimented on actually built into the to patients' health data.

我很好奇医生们对这些事是什么看法。

I'm curious how doctors are feeling about all of this.

我们看到美国医学会的一项调查，发现超过八成的医生已经在工作中使用AI了。

We saw a survey from the American Medical Association that found that more than eighty percent of physicians now report using AI professionally.

是医生们争着抢着把这些工具带进诊室，觉得真有用吗？

Is that physicians racing out and grabbing these tools and bringing them to the office because they're so helpful?

还是说，这不过是老板们在说：‘你必须用AI，不然就走人’那种情况？

Or is this, the classic case of a CEO saying, hey, you gotta use AI or, you're out of here?

医生们都是自己带AI来用。

So doctors are BYO AI.

很多AI应用其实是语音转录助手和临床决策支持软件。

A lot of that AI use is AI scribes and decision support software.

说实话，有些人直接拿ChatGPT、Gemini或者Claude当决策助手用。

And I'll tell you, some people are just using straight up, like, ChatGPT or Gemini or Claude for their decision support software.

我觉得医生们到现在对AI态度还比较正面，原因之一就是这些工具大多是他们自己主动选的，觉得真能帮上忙，而不是被硬塞过来的那些东西。

So I think one of the reasons doctors thus far have been more positive about it than perhaps the overall population is they're largely tools that doctors are bringing themselves that they think make their lives better and at least not yet many things that are being imposed upon us.

是的

Yeah.

我现在发现，我和朋友去看医生时，常常先跟聊天机器人说一遍情况，然后带着机器人给的分析结果去见医生。

I've noticed that when when I and my friends go to the doctor now, we often are presenting our information to a chatbot first and then coming into the doctor with sort of a readout of what the chatbot has told us.

这当然不是什么新鲜事了。

This is, of course, not a new phenomenon.

好多年了，大家就一直拿WebMD的搜索结果去问医生。

People have been doing this, like, with, like, WebMD, results for many years.

但现在你有没有发现，越来越多的病人是先跟聊天机器人聊完自己的症状，才来见你的？

But is this something that you're seeing now is that many more patients are coming to you having already discussed whatever's going on for them with a chatbot?

对。

Yes.

另一个大变化是，诊室里现在多了一个‘人’，经常就是ChatGPT。

This is the other big changes is that there's, you know, there's someone else in the exam room with me, and often it's it's ChatGPT.

有时候我跟住院病人聊天，他们一边跟我说话，一边在跟ChatGPT讨论病情。

They're talking sometimes with my hospitalized inpatients, they're talking to ChatGPT while I'm in the room with them.

我觉得这其实挺有意思的，因为这对医生来说是一种全新的能力。

And I think this is it's interesting because this is a kind of a new competency for doctors.

我们得跟病人聊聊人工智能。

We have to talk to our patients about AI.

我已经开始跟病人讨论，哪些用法是安全的、值得参考的，哪些是绝对不能听的。

And I have started to talk to my patients about what I think are, like, safe uses, what are, like, safe uses while telling me, and then things that they definitely shouldn't do.

可能因为我既是医生又是AI研究者，病人更愿意跟我聊这个，但其实很多病人早就习惯日常使用AI了。

My patients may talk to me more about it because I am, like, a doctor and an AI researcher, but, like, a lot of my patients are are using AI routinely.

那你能不能举个例子，你平时是怎么跟他们说的？毕竟我每次去看医生前，也老喜欢上网查各种信息。

Well, give us a flavor of what you're telling them because, you know, I am I am definitely somebody who has looked up by subgroups before I've gone to the doctor.

而且说实话，我觉得这对我帮助特别大。

And I would say I've I've found it enormously helpful.

但我也能想象，有些医生可能会觉得烦，毕竟病人拿着聊天机器人的话来当真，跟你较真。

But I can also imagine, you know, more skeptical doctors, being annoyed, you know, at at a a patient telling them, you know, what chatty bitty says to do.

所以

So

是的。

Yeah.

我来跟你说说我的做法。

So here's my I'll give you my spiel.

我会给他们讲一个……怎么说呢？

This is I I give them a what is it?

绿灯、黄灯、红灯三种情况。

A green light, yellow light, red light.

绿灯就是一些普通的健康问题。

So the the green light uses are general health questions.

我最近被诊断出糖尿病。

So I I'm recently diagnosed with diabetes.

我特别爱吃海鲜。

I really love seafood.

你能帮我设计一份适合糖尿病患者的饮食吗？

Can you help come up with a diabetic diet for me?

绿色信号还用来为看医生做准备。

The green light uses are also preparing for clinic visits.

我马上要去见罗德曼医生了。

So I'm about to go see doctor Rodman.

我想确保自己问对问题。

I want to make sure that I ask the right questions.

这是他上次写的最后一句话。

Here is the last note or the last thing he wrote.

当然，把所有能识别身份的信息都删掉。

Obviously, strip out anything identifying.

别把你的个人健康信息放进去。

Don't put your personal health information.

帮我想想该问他些什么好问题。

And, like, help come up with a good questions to ask him.

其他绿色信号的活动可能还包括可穿戴设备的数据。

And then other other green light activities might be, like, wearable data.

我不确定他们在可穿戴设备数据方面有多在行，但我要告诉你，如果一个病人能给我五年来的苹果手表数据，那他从ChatGPT那儿得到的建议，可能比从我这儿强多了——毕竟我只有二十分钟，哪看得完五年数据啊。

I don't know how good they are at wearable data, but I will tell you if a patient is going to give me, like, five years of their Apple Watch data, they're probably gonna get a better from, ChatGPT than from me pretending to look at five years of Apple Watch data because it's a twenty minute visit.

凯西，我觉得你说的那些，基本都属于黄灯范围。

The yellow light, Casey, I think is a lot of the things that that you're saying.

所以我跟病人说，新出现的症状，你尽管去探索。

So I tell my patients it's okay to explore new symptoms.

甚至找聊天机器人咨询一下，做个备忘、提前准备，也没问题，只要清楚一点：它不能替代医生。这一步，才是你跟真人医生沟通的起点。

It's even okay to seek out second opinions when talking to a chatbot that can really help prepare you as long as you understand that it's not a replacement for a doctor, and that is the first step to talking to a human being.

所以大语言模型确实很强大。

So LLMs are really powerful.

当然，也有证据表明，人用这些工具的时候，效果并不总是能达到实验室级别的精准。

And, I mean, there is some evidence, of course, that like, when any human uses them, you don't always get laboratory level performance.

它们甚至可能给出危险的建议。

Like, they can give you dangerous advice.

但如果是用来搞诊断、探索症状，只要目的是为了见医生前做好准备，那就很有用。至于红灯区，我跟他们说绝对别碰：别让AI替你做医疗决策。

But diagnosis and, like, exploring symptoms, as long as you use it in a way to prep to see your doctor can be very The red light, what I tell them never to do, is, like, ask medical management decisions.

别跟它说‘我医生让我这么做的’。

Like, don't say, my doctor said to do this.

这个对吗？

Is this right?

我得了癌症。

Like, I have cancer.

真希望你别得癌症。

God forbid you have cancer.

这个化疗方案是对的吗？

Is this the right chemotherapy option?

很多这类决定都特别复杂，涉及的信息量太大了。

Like, a lot of those decisions are so nuanced, sinking so much information.

这些事AI根本搞不定，而且它们特别会哄人，错得离谱还能让你觉得它说得对。

Those are things that the models don't do well, and they're so sycophantic, they can convince you that they're saying the right thing even when they're wrong.

对。

Yeah.

亚当，我很好奇。

I'm curious, Adam.

在这儿，

Out here

在旧金山，到处都是健身达人和健康狂人，人人都爱用各种设备追踪自己的身体数据，还找像Function Health这样的公司做全套体检，搞个上百项化验，然后把所有数据丢进Claude或者ChatGPT，当它们是自己的第一道医疗顾问。

in San Francisco, there are all these fitness people and health maxers, people who love to track themselves using all manner of devices, and people are getting these full body workups from companies like Function Health that are, you know, sort of concierge medicine things, and they'll get, you know, a 100 labs done, and then they'll upload all that data into Claude or ChatGPT and just sort of treat it as a sort of first line medical professional in their lives.

你觉得这种做法靠谱吗？还是说，反而让人们对根本不该担心的事过度焦虑了？

You think that is a good practice, or is that just making people, you know, way too worried about things that maybe they don't need to be worried about?

对。

Yeah.

这确实让人对根本不该在意的事太紧张了。

So that's making people way too worried about things they don't need to worry about.

这纯粹是闲聊式诊疗。

And this is chatty PT.

大模型都这样。

LLMs in general.

我的意思是，跟大语言模型聊症状的坏处在于，它们太会哄人了，容易把你推入那种网络疑病症的焦虑深渊。

I mean, the dark side of talking to an LLM about your symptoms is they are so sycophantic, they can drive you into, like, the cyberchondria worry hole.

目前还没证据表明，这种大规模常规检测+把数据扔进大模型的做法，真能改善健康结果。

The evidence is not there yet that these sort of large routine testing functional medicine and putting it into an LLM does anything to improve health outcomes.

不过要是大模型劝你多运动、吃得健康，那倒确实挺靠谱的。

Now if if your LLM is telling you to to work out and eat healthier, that's probably pretty good.

好好睡觉。

Sleep.

对。

Yeah.

那像ChatGippityHealth这种工具呢？它能把你Apple Watch或者Fitbit的数据转成ChatGPT能分析的形式。

What about the integrations like ChatGippityHealth, which lets you sort of convert your Apple Watch or Fitbit data into something that ChatGPT can analyze?

还有个新版本叫ChatGPT for clinicians，专门给医生用的。

There's also a new version of ChatGPT for clinical use called ChatGPT for clinicians.

在你看来，这些整合方案或者项目里，有没有哪个更有前景？

Are any of these integrations or projects more promising in your in your view?

还没到那一步，但我觉得将来有可能。

Not yet, but I think it could be at some point.

我的意思是，ChatGPT 健康版能把你病历里的数据拉进来，让你跟自己的病历聊天。

I mean, so ChatGPT for health pulls in your data from the medical record and lets you chat with your medical records.

第一个让人担心的问题是隐私。

Now reason number one for concern is privacy.

这等于把你的整个医疗记录都传给了一个AI公司。

That's obviously gonna have your entire medical history going to a AI company.

而且这些记录也不会经过你手动去重，把能识别身份的信息删掉。

It's also going to not be redacted by you in a way to remove identifiable things.

第二个问题，我觉得要是谈医疗记录数据，那真是乱得一塌糊涂。

Reason number two, I I think, if we're talking about health record data, it's really messy.

里面既有表格数据。

They include tabular data.

还有大量复制粘贴进去的文本。

They include copy forwarded data that's been copied and pasted.

而且如果你看过自己的医疗记录，就会发现里面还有很多错误的信息。

And they also, if you've ever read your health records, they include things that are wrong.

你的健康数据里头，错误和记录不当的地方可真不少。

There's a lot of errors or misdocumented things in your health data.

结果呢，光是把一堆信息直接扔给大模型，这事儿可没那么神奇。

And it turns out that just copying a bunch of information like, there LLMs aren't magical.

你不能把整个病历一股脑儿丢进去，就指望它能给出靠谱的建议。

You can't just copy your entire medical record in and think that you're gonna get good performance.

但我从来不会否定这项技术的潜力。

And I would never bet against the technology.

我相信，未来我们一定能找到方法，把人的健康状况建模出来，真正理解它。

I think that we will get to the point that we have ways to build representations of of humans and understand their health.

但就现在来说，把所有东西全塞进大模型里——就像ChatGPT for Health理论上能做的那样——其实根本没啥优势，也帮不了你更好地了解自己的健康。

But right now, there's, like, no advantage to just dumping everything in an LLM, is what ChatGP2 for health theoretically would allow you to do in a way that would allow you to better understand your health.

我很好奇，你有没有看到犹他州那个试验？他们用AI代理自动续开近200种常规药物的处方。

I'm curious if you saw this trial they're doing in Utah where you can use an AI agent to autonomously renew prescriptions for almost 200 routine drugs.

是的。

Yes.

据说有人工审核，但大部分都是自动处理的。

There's apparently some human review, but, mostly, this is automated.

你觉得这是好主意还是坏主意？

Is that a good idea, bad idea?

从全球范围来看，不行。

Well so globally, no.

我们不该让大模型直接给病人开处方。

We should not be, having LLMs write prescriptions for people.

犹他州这个试验特别针对的是续药。

The trial in Utah in particular is a refill.

也就是说，医生在过去12个月内已经开过一次处方了。

So a doctor has already written a prescription within the last twelve months.

大概就是想节省全科医生的时间，不用他们反复审核和续方。

And I guess the idea is that it saves the primary care doctor time from having to review and refill.

说实话，如果你去问大多数医生，他们确实挺烦这些续药请求的。

I'll tell you, if you talk to most doctors, yes, it is annoying to get refill requests.

不是的。

No.

真正让我们抓狂的不是这个。

That is not the thing that drives us crazy.

这根本不是我们急着想推进的场景。

This is not like a use case that we're screaming for.

我觉得这更多是为了验证一下，看这种模式在现实里能不能行得通。

I think it's being done as a proof of concept of of can this work in the real world.

这个试点项目本身并不危险。

This trial in and of itself is not dangerous.

只是开药续方，而且我听说里面没有阿片类药物。

I prescription refills, and I think there's no opiates.

没有危险药品，而且最初的处方必须是医生开的。

There's no dangerous drugs in it, and a doctor has to have written the original one.

但就算这个系统在这儿能行得通，也不代表我们该让AI自主开新处方。

But even if it does work in this, that does not mean we should be having autonomous AI systems write new prescriptions.

这不安全，现在也不是个好主意。

That is not safe, and it's not a good idea yet.

我觉得这其实是医生或者医学会在搞变相收费。

See, I think this is a case where, like, this is sort of rent seeking behavior on the part of doctors or doctor organizations.

我每次续药，都要跟医生见一面，大概六到八分钟吧。

Like, when I have gotten refills for prescriptions, I meet with a doctor for, you know, six to eight minutes.

他们就问：‘药效果怎么样？’

They say, how's it working?

我就说：‘挺好的。’

I say, great.

他们再问：‘有啥副作用吗？’

They say, are you having any side effects?

我就说：‘没有。’

I say, no.

他们就说：‘行，我给你续个药。’

They say, okay, I'll write you a refill.

整个流程看起来就是为了让你再付一次诊疗费，根本没给你什么真正的医疗建议。

And the whole process just seems totally designed to, like, get me to pay up for another doctor visit and not give me any actual good medical advice.

那我换个角度说，你觉得大家抵制这类系统，是不是就是想让人不断跑医院、花钱看病？

So if I could play devil's advocate, like, do you think that there that the sort of resistance to programs like this are motivated by just wanting to keep people coming to the doctor and paying for those visits?

首先，你那些药续方不都是打电话给药房，他们自动发个请求给医生，医生点个‘是’就完事了，你根本不用跟医生说话吗？

So first, aren't most of your prescription refills just done as in you you call the pharmacy and they send an automated thing to your doctor and they click the yes button and you never talk to them?

不是的。

No.

有些药，他们非得让你亲自来门诊一趟。

For some, they make you actually do an office visit.

可能还让你再测个血压什么的。

And maybe they want you to, you know, take your blood pressure again or whatever.

那我也来替他们说句话。

So I'll I'll do the devil's advocate back.

假设我开了一种很常见的抗抑郁药，患者需要续方。

Let's say I prescribe a fairly common antidepressant, and they they wanted to be refilled.

但我不知道的是，这个病人可能正出现口腔里新长的病变，其实是早期溃疡。

What what I don't know is that this patient may be the the silly question you get in the clinic may be new lesions forming in your mouth, and it's an early ulcer.

如果我们没在24到48小时内发现，可能会发展成史蒂文斯-约翰逊综合征，这可是能致命的并发症。

And if we don't pick it up within twenty four to forty eight hours, you may develop, like, Steven Johnson syndrome, so potentially life threatening complication.

有些药物，比如降压药，风险比较高，所以必须跟进观察。

And the reason there are certain types of drugs, including antihypertensives, is that they can be high risk, and we need follow-up.

那这就是全部了吗？

Now is that everything?

当然不是。

No.

而且，确实应该有更多药可以不用处方就能买到。

And, definitely, there should be more things over the counter.

我觉得，大多数医生根本不会闲着想：‘真希望我能多些复诊机会。’

I like, I don't think that most doctors are sitting around saying, I wish I had more medication follow-up visits.

这些措施之所以存在，是因为有些症状可能非常危险。

And the reason some of these things exist is that there can be very dangerous symptoms.

是的。

Yeah.

所以你得继续去看医生，凯文。

So keep going to the doctor, Kevin.

我们可不能让你长出那些溃疡。

We can't have you developing those lesions.

这事儿对节目来说太重要了。

It's too important to the show.

我再问你一个别的。

Let me ask you about another one.

这个其实完全是件好事。

This one actually seemed like just an unqualified good.

梅奥诊所本周宣布了那个叫Red Mod的AI系统，它能提前三年从常规CT扫描里发现胰腺癌的细微征兆。

The Mayo Clinic announced this week, Red Mod, this AI system that identified subtle changes in routine CT scans up to three years before a pancreatic cancer diagnosis.

这在检测胰腺癌方面的准确率，比人类高出太多了。

And this was, like, many, many, many percentage points better in detecting pancreatic cancer than human beings.

这正是我一直等着AI去实现的事，没想到它真做到了。

So to me, this is like the sort of thing I keep waiting for AI to do, and it seems like it's actually doing it.

尤其是胰腺癌这种特别难发现、存活率又极低的病，这简直太让人振奋了。

And, of course, that's very exciting with something like pancreatic cancer, which is notoriously difficult to detect and has, like, very low survival rates.

是的。

Yeah.

这跟那些自动驾驶AI代理的讨论完全不在一个频道，但真正了不起的进展正在发生。

And this is so, like, completely out of the discourse of, like, autonomous AI agents, there's really exciting stuff happening.

梅奥诊所已经在乳腺癌筛查方面做了不少出色的研究。

So the Mayo Clinic, there have been some great studies on breast cancer detection.

这些算法已经厉害到能在临床流程中，达到比医生更稳定的检出率了。

A lot of these algorithms have gotten so good that they're able to identify breast cancer better than I shouldn't say better than people, but in a workflow that has a high a good detection rate.

还有在做结肠镜检查时，能更早发现可能癌变的息肉。

And then in in picking up, like, potentially cancerous polyps when you get a colonoscopy.

所以现在有很多令人振奋、非常积极的进展。

So there's a lot of exciting and really positive things that are coming.

说实话，最终我们还得看Red Mod在真实世界和临床试验中的表现，但我对这类技术真的很乐观。

And I'm I I mean, at the end of the day, we'll need to see how Red Mod works in the real world and a trial, but I I'm really optimistic about that sort of technology.

你觉得，如果AI能显著延长人的寿命，是因为AI发现了新药，还是因为AI让日常医疗变得更高效、更准确？

Do you think that if AI meaningfully extends life expectancy for for people, it will be because of new AI discovered drugs or because of changes to routine health care that are made more efficient or more accurate by AI?

是后者。

Number two.

我觉得谈AI药物研发时，最难的环节其实不是设计新化合物。

I think that when you talk about AI drug discovery, the part of the pipeline that's so difficult is not necessarily the coming up with the new compounds.

真正卡脖子的是做临床试验和走审批流程，这些或许能加快，但远不如药物发现阶段提升的空间大。

It's running the clinical trials and getting it through the regulatory process, which can probably be sped up, but not as much as the discovery.

你知道吗？在美国，有太多人看不起医生，连最基础的药都买不起，因为缺医少药，连糖尿病都控制不好。

You know, if we get this right, there's so many people in The US who don't have access to a doctor, who don't have access to very basic medications, who can't control their diabetes because of lack of access.

如果咱们能做对了，我真的希望能让更多人获得医疗资源——我这话说出来都得敲敲木头，别灵验了反被说嘴。

And I'm really hopeful that if we do this wisely, we can, you know, get people more access to care, which I'm doing my knock on wood.

希望这样能改善健康状况。

Hopefully, will improve health outcomes.

所以啊，说真的，这些潜在的好处其实没那么让人激动。

So like like all of this, I think the potential benefits are, like, less exciting.

真正重要的是让更多人吃到饭、喝上水，减少心脏病和中风，让更多人做癌症筛查，而不是指望靠什么新发现的AI基因编辑技术就能治愈衰老。

They're they're getting people more people the bread and butter and getting, you know, more people to have less heart attacks, more people to have less strokes, more people to get their cancer screening, and not necessarily like, oh, we cure aging with some sort of new AI discovered CRISPR technology.

不过，你有没有觉得，咱们到现在还没看到第一个由AI发现的神药，这其实挺意外的？

Are you at all surprised, though, that we haven't yet seen the first, like, AI discovered wonder drug?

要说我觉得最了不起的神药，那肯定是GLP类药物，我当住院医的时候就开始用了。

I mean, the biggest wonder drug of my career has been the GLP ones, which, was I started using it when I was a resident.

这药我们用了很久，后来才重新发现它能治糖尿病。

So we had it for a really long time, and we had to, like, repurpose the drug for diabetes.

所以，我不觉得意外。

So, no, I'm not.

说实话，医学和科学本来就是一团乱麻，总有些故事讲的是，哎，我们突然发现了什么了不起的东西，像青霉素那样。

Like, medicine and science is is just kind of messy, and it's there are always those stories about, like, you know, we discover something amazing and makes it like penicillin.

但即使是青霉素，也花了整整二十年才用到病人身上。

But even penicillin took, like, twenty years to get into human beings.

所以，不，我觉得AI最终还是会发现新药的。

So, no, I I think we will see AI discover drugs.

我觉得AI带来的好处，会跟医学进步一样，没那么轰动。

I think it's just the benefits from AI are gonna be like the benefits from medicine.

虽然没大家想的那么激动人心，但依然很重要。

It'll be a lot less exciting than people think, but still important.

现在很多人担心AI进学校、进教育，怕人会因此退化，脑子变懒。

So there's a lot of worry right now about sort of AI in schools, in education, the some of the cognitive atrophy that people are worried about.

要是我们事事都靠AI，那基本功不就丢了吗？

Oh, if we start using AI to do all of our work, we're not gonna have the basic skills.

你担心现在的医学生会这样吗？以前他们得把所有知识记在脑子里，现在动动嘴问个聊天机器人就行，会不会慢慢把医生的本事给磨没了？

Is that something you're worried about for, like, recent medical school graduates where maybe they would have had to hold all this stuff, in their brains a few years ago, and now they can just ask a chatbot, and maybe that's gonna erode some of their skills as a physician?

是的。

Yes.

所以，我目前最担心的短期到中期问题，就是劳动力的技能退化。

So that is the biggest worry that I actually have about sort of the short to medium term is deskilling of the workforce.

我们已经有了一些证据。

We have some evidence.

去年波兰有个研究挺吓人的，他们给医生用的不是语言模型，而是一种能检测息肉的技术。

There was a sort of scary study last year from Poland on a trial where they gave doctors not a language model, but a polyp detecting technology.

研究人员先看了医生们在使用这项技术前检测息肉的能力，用了三个月后再看一次。

And they they looked at their ability to detect polyps, so potentially cancerous lesions in the colon before using it and then after using it for three months.

结果发现，不靠技术的时候，他们检测息肉的准确率下降了六个百分点。

And when not using it, their ability to detect polyps dropped by six percentage points.

这些可都是有经验的医生，用了技术才三个月，检测癌症前兆的能力就掉了六个百分点。

So these are skilled doctors using technology, and they lose six absolute percentage points of their ability to detect potentially cancer in three months.

那你想想，要是刚学这行的人，从一开始就靠机器呢？

And then imagine that you're learning to do it for the first time.

他们还能不能真正掌握这些本事？

Will you ever gain those skills?

所以，像哈佛医学院这样的地方，我觉得全球的医学院都一样，我们最担心的就是：这玩意儿会怎么影响我们培养新一代医生？

So, like, at at Harvard Medical School, like, this is and medical schools, I think, everywhere, this is our big worry, which is how will this affect us to train the new new generation of doctors?

这事儿跟其他行业其实都一样。

And it's like every other field.

你就说写代码调试吧。

Like, you talk about debugging code.

想当个好医生，得经过一整套训练，为啥？因为你得犯错，还得有个比你懂行的人在旁边盯着，不然你犯的错真会害了病人。

In order to become a do new doctor, you go through all this training because you need to make mistakes, and you need to have someone who above you who knows what's going on so those mistakes won't hurt patients.

教育本来就是这么回事，但现在这东西正威胁着整个模式。

And that's just how education works, and it's this threatens that.

说真的，这事儿挺有意思，因为我有图形计算器，要是现在把它拿走，我肯定连抛物线都画不好了。

I mean, it's interesting, though, because it's like, you know, it's probably true that because I had access to a graphing calculator, like, if you took it away from me, I'd be worse at, like, plotting parabolas on a graph.

但解决办法其实很简单——我继续用计算器就行了。

But the solution to that is that I just keep using the calculator.

你懂的？

You know?

所以啊，我不确定这问题到底有那么严重。

So, like, I'm not sure how big of a problem this really is.

我还想说，人类社会里头，中年人抱怨年轻人这事，早就根深蒂固了。

I I'll also say there's something, there's something deeply ingrained in human society that middle aged people complain about young people.

所以每当我们谈技能退化的时候，都得记着这一点。

So I think whenever we talk about deskilling, we have to keep that in mind.

对。

Yeah.

我的意思是，我就算……

I mean, I for

说真的，我希望我的医生都用AI模型。

what it's worth, like, I want my doctors to be using AI models.

我希望他们在给我做诊断前，先去参考一下集体智慧。

I want them to be consulting the hive mind before they weigh in on my specific condition.

知道他们用的是公开的证据或者类似的东西，我作为病人一点都不担心。

It doesn't threaten me as a patient to know that they are using open evidence or something similar.

但我猜，对很多人来说，这听起来会挺奇怪的。

But I'm guessing for a lot of people, that would seem strange.

也许有些医生不敢公开说他们用了AI，怕病人觉得他们不靠谱。

And maybe there are some physicians who don't advertise how much they're using AI because their patients might think less of them.

你觉得这种情况真的存在吗？

Do you think that's happening?

哦，确实有。

Oh, yeah.

我觉得在某些情况、某些地方，肯定存在一种社会压力，让人不敢承认自己用AI，好像承认了就丢了面子似的。

I think there in certain situations and certain places, I bet there's social pressure to say that you're not using AI, that there's some ego on the line.

但我个人没觉得这样。

I don't see that.

不过我是搞AI研究的，估计没人会在我面前这么说。

But, I'm an AI researcher, so I don't think anyone would say that to me.

对我来说，把医生能不能背下最多知识当作好医生的标准，这本身就挺奇怪的。

To me, it just seems weird to, like, to hold as your standard for what makes a good doctor that they have memorized, like, a maximum amount of material.

不然呢，其实我们讨论的核心就是这个意思。

Or, like, that's basically what we're talking about.

这感觉就像伦敦的出租车司机，他们得把每条街道都背下来，全装在脑子里。

It's sort of like, you know, the taxi drivers in London that have to, like, to, like, learn every single street and, like, hold them all in their heads.

这本事确实厉害，但就算他们用导航，我也完全觉得没问题。

It's very impressive, but I'm fine with them using the GPS.

对的。

Yeah.

对的。

Yeah.

我觉得现在情况不太一样了——以前确实拼的是谁背得多，

And I I think it's less about so it is about memorizing.

但以现在AI的发展水平来说，早就不是这样了，

It's more at this point, right, with where AI is now.

现在更重要的是你有扎实的专业积累，甚至可以说是有经验智慧，能判断出AI系统给出的结论哪不对。而这种判断力，至少以目前的情况来看，以后说不定会变，只能靠经手大量病例、不断复盘总结才能练出来。

It's more having sort of that knowledge and call it wisdom to know when the system might be suggesting something wrong, which is something that right now, and this may change, we get by seeing a lot of cases and reflecting on them.

所以现在，要是想达到最佳效果，还是得靠那些受过传统训练、有经验的人，再配上AI系统。

So right now, you're gonna get the best performance if you have experienced human trained in the old fashioned way with an AI system.

但我觉得你们的意思是，总有一天这可能就没那么重要了。

But I think your guys' point is at some point that might not matter.

对吧？

Right?

AI系统迟早会超越我们所有人。

The AI systems might just outperform all of us.

到时候嘛，就像用图形计算器一样，直接用就行——不过我们现在还没到那一步。

And then, yeah, I I guess it's like just use the graphing calculator, but we're not there yet.

如果我们对医疗数据的隐私保护没那么严格，AI模型会不会更强？

Would the AI models be better if we were less protective of privacy for medical data?

我的天，这问题可太有争议了。

I mean, that's a that's such a loaded question.

在回答之前，我得先说一句：病人隐私特别重要，我们得尊重每个人的隐私权，尊重他们对自己数据的掌控权。

So the first thing that I'm gonna say before I answer that is patient privacy is very important, and we should respect people's privacy and their ownership over their data.

但确实如此。

But yeah.

简而言之，它们在某些方面表现不够好，是因为想提升大模型，就得用标注过的医疗数据来训练。

So in short, like, the reason they're not better at certain things is that you need to to get LLMs better, you need to label and then train them on the sort of labeled health data.

在美国，医疗数据的使用本来就受到很多合理限制。

And there's all in The US, there are appropriately many restrictions on how health data can be used.

我猜像OpenAI这样的公司，通过推出ChatGPT医疗版，其实能积累更多自己的数据——尽管他们说不会拿去训练。

I suspect that these companies like OpenAI, by having ChatGPT for health, they will gain some more of their own data, which they say they're not gonna train on.

我相信他们确实不会拿去训练，但至少能用这些数据来评估模型、优化表现。

I trust that they're not gonna train on it, but they'll be able to use that data to at least evaluate their models and try to make them better.

我觉得他们就该拿去训练。

I think they should train on it.

当然，这会严重违反隐私法规，违法了。

I mean, obviously, that'd be a huge illegal violation of privacy.

但AI医生的能力肯定会大幅提升。

But it would also make the the AI doctors better.

是的

Yeah.

好太多了

Much better.

而且我觉得，你

And and I think, you

你知道，很多人其实愿意做这种交换。

know, a lot of people would be sort of willing to make that trade off.

所以我至少觉得，看医生的时候应该有个小勾选框，写着：我同意我的个人健康数据用于训练AI模型。

So I at least think there should be a little checkbox when you go to the doctor that says, like, I'm okay having my personal health data used to train AI models.

我本人

I for

是的

one Yeah.

换来的可能是你的天价医疗账单打七折。

Exchange for, like, 30% off your giant medical bill.

你有优惠券呢。

You got a coupon.

你会有。

You would

拿到优惠券。

get a coupon.

你懂的吧。

You know?

你下次用司美格鲁肽，能打八折。

You got, like, your next Ozempic shot is 20% off.

免费送你。

It's on the house.

没错。

Exactly.

那咱们就先说到这儿吧。

Well, that's good place to leave it.

感谢亚当·罗德曼医生再次做客，也请继续跟我们分享医学界的最新动态。

Doctor Adam Rodman, thanks so much for coming back, and keep us posted on what is going on in medicine.

不客气，各位。

My pleasure, guys.

非常感谢。

Thank you very much.

谢谢医生。

Thanks, doc.

回来之后，我们来聊聊 Talkie，一个只用1930年之前数据训练的大型语言模型。

When we come back, we'll talk about talkie, an LLM trained only on data from before 1930.

凯西，我们这档节目平时讲的都是未来的事儿。

Well, Casey, usually on this show, we are talking about the future.

但今天咱们要倒着走，回到1930年那个年代。

But today, we are going to take a trip back to the past, specifically to the year 1930.

1930年那会儿，世界都发生了啥？

What was happening in the year 1930?

天哪。

Oh my goodness.

当然了，那时候我们正处在大萧条时期。

Well, of course, we were in the middle of the Great Depression.

我奶奶刚满十一岁，正盼着几年后能买到人生第一件商店里的新裙子。

My grandmother had recently turned 11 and was looking forward to getting her first store bought dress in just a few years.

这是一个全新的语言模型，叫Talkie，专门用1931年以前的数据训练的。

So this is a new language model, a vintage vintage LLM called talkie, and it is trained exclusively on data from before 1931.

这个项目是三个人做的：大卫·杜韦诺、尼克·莱文，还有亚历克·拉德福——就是写GPT-1那篇论文的主笔，以前在OpenAI的那位。

This is a research project built by three guys, David Duvenaud, Nick Levine, and Alec Radford, the lead author of the GPT one paper, former OpenAI researcher over there.

这项目太有意思了，这周刷爆了我的信息流，因为它其实是在做一次实验：如果只给大语言模型喂1930年以前的数据，它会变成什么样？

And this is a fascinating project that has been burning up my timeline this week because this is an experiment in what happens if you only feed a large language model data from before a certain cutoff.

是的。

Yeah.

网上当然也挺多那种基于角色的聊天机器人，能让你感觉像在跟过去的人聊天。

And obviously, there are a lot of, you know, kind of character based chatbots on the Internet that will give you the experience of talking to somebody from the past.

但这个项目特别的地方在于，他们刻意只用那个年代及之前的数据来训练。

But what makes this project different is that they try to limit themselves to training data from that time and before.

他们希望这样能避免受到之后时代信息的干扰。

The hope was that it would avoid any kind of contamination from what came after.

你会听到一些非常有趣、甚至可能很有用的想法，关于这种语言模型未来能派上什么用场。

And as you'll hear, have some really interesting and potentially useful ideas about what this kind of LLM might one day be used for.

凯西，你有没有花时间试用过这个模型？

Casey, have you spent any time playing around with this model?

试过。

I have.

我试着问了它一个我觉得最‘1930年代’的问题——比如说，什么是最重要的理念？

I tried to ask it the most nineteen thirties question I could think of, which was, say, what's the big idea?

它怎么回答的？

What did it say?

它说：最重要的理念，就是普及。

It said, the big idea is to popularize.

我就问了：‘你说的普及，普及啥呀，老兄？’

And I said, popularize what, fella?

它回答：‘普及一项运动。’

And it said, popularize a sport.

我说：‘我是同性恋。’

And I said, I'm gay.

然后这事我们就没再往下聊了。

So that's kind of where we let that one drop.

它回了个：‘同性恋。’

And it said, gay.

你开心吗？

You're happy?

对啊。

Yes.

没错。

Exactly.

它

It

说你的心一定很轻松，先生。

said your heart must be light, sir.

是的。

Yeah.

我超爱这个实验。

I love this experiment.

我就喜欢那种怪怪的小众语言模型。

I love, like, weird niche language models.

我最爱的语言模型之一叫Golden Gate Claude，就是那个特别版的Claude，简直对金门大桥着了魔。

My one of my favorite language models of all time was Golden Gate Claude, which was the special version of Claude that was, like, pathologically obsessed with the Golden Gate Bridge.

我觉得Talkie也属于这一类——算不上多实用，但能帮我们看清一些特别重要、有趣的东西，比如语言模型在特定训练方式下会怎样。

I would put talkie in sort of that category of, like, an experimental research model that is maybe not all that useful on its own, but like helps illuminate something interesting and important about these language models and what happens when you train them in specific ways.

所以今天，我们想聊聊Talkie的创作者之一。

So today we wanted to talk with one of the creators of talkie.

我们请来了大卫·杜韦诺德。

We are going to bring in David Duvenaud.

大卫是多伦多大学的副教授，研究人工智能通用治理和灾难性风险防控，同时也是Talkie的联合创始人之一。

David is an associate professor at the University of Toronto who researches AGI governance and catastrophic risk mitigation, and he is one of the cocreators of talkie.

找大卫来聊这个，再合适不过了。

And there's really Duvenaud better person to talk to about it.

没错。

That's true.

大卫·杜韦诺德，欢迎来到《硬核》节目。

David Duvenaud, welcome to Hard Fork.

非常感谢，凯文。

Thank you very much, Kevin.

这个叫Talkie的项目太有意思了。

So this project, talkie, is fascinating.

它是个复古风格的大型语言模型。

It is a vintage LLM.

你和尼克、亚历克当初为啥要搞这个东西？

Explain why you and Nick and Alec made this thing.

这事其实是一年前开始的。

So this all started a year ago.

当时我和尼克对预测特别感兴趣。

Was me and Nick were interested in forecasting.

我们想搞清楚：能不能教会机器预测五年、十年后的事？

Like, specifically, can we teach machines how to forecast, like, five or ten years ahead of time?

就是，未来的大图景到底会什么样？

Like, what is the big picture gonna be?

我们自己当然也有不少想法，但不觉得别人该盲目信我们。

Just because we have our own sort of, like, pet ideas about what the future is gonna be, we don't think people should take our word for it.

而且我们也不觉得，机器做的预测能轻易让人信服，除非它有几十年的准确记录。

And we also don't think that people should trust machine forecasts unless they have a track record going back, like, decades and decades.

所以我们的想法是：如果能造出一个只了解某个时间点之前世界的模型，就能让它去预测五年、十年后的事。

So the idea here is that if we could build a model who really only knew about the world up to a certain date, we could ask it to forecast, like, five or ten years ahead of time.

比如问它：五年后《纽约时报》的头条会是什么？

Like, ask it, what's the New York Times headline gonna be five years from now?

或者会不会再爆发一场大战之类的？

Or is there gonna be another great war or something?

我们可以不断尝试，看看哪些事情是能预测的。

And we can iterate and see, like, what kinds of things are predictable.

到底需要什么条件？

What does it take?

也就是说，事情能提前多久被预见？

Like, how far out can things be foreseen?

然后，希望最终我们能拥有拥有百年预测记录的机器。

And then, hopefully, eventually, we'll have machines that have, like, a hundred year track record of forecasting.

到那时，我们就可以在2026年问它：你觉得八年后、四年后、或者两年后会发生什么？

And then we can ask them, you know, in 2026, what do you think is gonna happen, like, know, two or four, eight years from now?

那时候我们就能知道该多信任这些预测了。

And we'll have an idea of how much to trust those forecasts.

这想法挺有意思的，但我觉得关键得有特别好的数据。

It's a fascinating idea, but it strikes me it requires you to have, like, really good data.

所以这里得是1930年以前的高质量数据。

So in this case, really good pre nineteen thirties data.

我猜获取这些数据肯定比现在随便爬爬Reddit或者那些前沿模型能接触到的资料难多了。

I'm gonna guess that was harder to obtain than just, you know, going out and like crawling Reddit or, you know, and and everything else that the frontier models have access to.

那你们是怎么解决这个问题的？这些1930年以前的数据是从哪儿来的？

So how did you face that challenge and where'd you get this pre nineteen thirties data?

是啊。

Yeah.

我得提一下，其实现在有很多团队都在做特别棒的档案整理工作。

So I should mention, you know, there's a ton of groups doing a ton of awesome archival work here.

我们最早觉得惊艳的一个数据集，是哈佛图书馆扫描的他们馆藏的1%内容。

The first dataset we got excited about was institutional books, which was Harvard Library scanned, like, 1% of their entire collection.

所以他们手上有大量18世纪到20世纪初的资料。

And so they had tons of data from, like, eighteen hundreds, early nineteen hundreds.

其实有很多团队都在做大量的档案整理工作。

Like, there's there's whole bunch of different groups doing tons of work.

要一一列举得花很长时间，而且最近半年，OCR技术的进步简直神速。

It would take a long time to enumerate this, but and also, like, OCR has just gotten a ton better just in the last, like, six months even.

以前就一直有人想把这类数据自动数字化，但质量一直不行，直到最近才真正好起来。

And so there's always been lots of projects to, like, automatically digitally scan this this data, but it just hasn't been very high quality until very recently.

我猜你们选1930年作为分界线，是因为那之后的作品大多还受版权保护，之前就进入公共领域了，对吧？

And I assume that part of the reason you chose the cutoff date of around 1930 is because that's when sort of works become public domain, Anything after that is copyrighted.

除了这个，还有别的原因选这个时间点吗？

Are there any other reasons you chose that specific point in time?

没有。

No.

就这一个原因。

That was entirely it.

我们就是想让所有数据都公开开源，而1930年是最后一个几乎不会惹上法律麻烦的年份。

It's that we wanted to make everything publicly available and open source, and nineteen thirties is just the sort of most recent date that has almost zero legal headaches with releasing data or or anything like that.

是的

Yep.

所以

So

我一直特别感兴趣，看大家用这个模型都做了些什么。

I've been fascinated by seeing, like, what people are trying with this model.

有人让模型做预测，也有人问它最喜欢的作家，或者对哪些历史人物的看法。

People are having it make predictions, but also asking it about its favorite authors or its opinions of, you know, of major historical figures.

对你来说，哪些实验最有趣？

What have been the experiments that have been most interesting to you?

对

Yeah.

我看到大家最常玩的，就是问：2026年会是什么样？

The the fun things that I've seen people do is, I mean, a lot of people like to ask, like, what's 2026 gonna be like?

模型的回答还挺有哲理的，比如说，我们终于明白战争是坏事了。

And the model has sort of very philosophical answers about how, like, well, we we we have figured out that war is bad.

我们会迎来一个和平得多的文明。

We'll much have, like, a much more peaceful civilization.

有时候它又说，末日要来了。

Or sometimes it says, like, it's the end times.

这模型太不稳定了，还没聪明到能系统性地深入思考问题。

I mean, it's a very inconsistent model, and it's not quite smart enough to really, like, you know, think things through in a systematic way.

它就是给你一种感觉而已。

It it kind of just gives you vibes.

是的。

Yeah.

这其实引出了一个挺有意思的问题，关于这类大模型的本质。

Now it it that that brings up a sort of interesting, like, wrinkle of, like, the kind of LLM this is.

因为如果你现在去问一个前沿模型预测未来，它不只是在猜一串统计上最可能的词。

Because if you were to ask, like, a frontier model today to predict the future, it would not only be trying to guess a statistically likely sequence of words.

对吧？

Right?

它还得做一些推理才行。

It would also be, like, doing some reasoning.

它可没在做这个。

It talkie is not doing that.

对吧？

Right?

所以从它的设计方式来看，我们自然会认为它在预测未来方面，比现在的模型差得多。

So that that just sort of seems like by the way that it is built, we would expect it to be less good at forecasting as as the models we have today.

没错。

Yep.

当然了。

Absolutely.

这还只是刚起步的水平。

This is a very baby steps model.

目前这种基础的微调，还有像超级预测那样的辅助框架，其实都能显著提升人的推理能力。

The basic fine tuning for reasoning and the scaffolds like, the super forecasting scaffolds that we know just improve anyone's reasoning.

比如说，你想一想各种不同的可能性，然后给每个可能性分配一个概率。

Like, you know, think of the different distinct possibilities and assign them each sub probabilities.

这模型现在还不够聪明，没法好好执行这种复杂的多步骤指令。

The model's just not really smart enough to follow these kind of detailed multi steps instructions yet.

所以我们只是先发布了最基础的版本，但很明显，后续完全可以一步步加上这些改进。

So, again, we just wanted to release the first thing that we did, but it's like a there's a clear path to adding all these refinements.

所以你们以后打算加上推理功能吗？

You so you do plan to add reasoning as you go?

当然了。

Oh, absolutely.

对。

Yeah.

明白了。

Okay.

还有人指出，Talkie背后的模型似乎知道一些它不该知道的事，比如希特勒上台、罗斯福当总统这些事，都是在它数据截止之后才发生的。

People have also pointed out that the model behind talkie seems to know about some things that it probably shouldn't know about, like the rise of Hitler, the presidency of FDR, things that that didn't happen until after its data cut off.

这能证明训练数据被更近期的信息污染了吗？

Is that proof that there's been some kind of contamination of the training data with more recent data?

肯定有污染。

Oh, there's definitely contamination.

这事儿我们会一直反复处理、不断优化，算是个长期问题。

And this is sort of like one of the ongoing, like, things that we're gonna have to just keep revisiting again and again and refining.

我们有个分类器，专门找那些时间错乱的内容。

So we have a classifier that tries to look for things that are anachronistic.

尤其是如果你要用它来做预测或评估预测，这个问题必须彻底解决。

And especially if you wanna use this for forecasting or to evaluate forecasting, it's really important that we really nail this issue.

我们想了很多办法，比如设一些‘警报词’，让模型绝对不能给它们任何可能性。

So we have all sorts of ideas for, like, canaries and things that we think, the model should just never assign any likelihood to.

比如说，广岛和长崎，你想想看。

Like, think of, I don't know, Nagasaki and Hiroshima.

二战之前，这两个地方几乎不可能出现在同一句话里，除非是特别特别偶然的情况。

Like, before World War two, like, those those two towns would just never show up in the same sentence ever almost, except for, like, some weird weird coincidence.

所以只要模型觉得这些特定名字有可能出现在同一句话里，你就能判断出有没有重要事件的信息泄露。

So you can just tell whether there's been, leakage about important events if the model just thinks that there's any chance that you'll see those particular names together.

总之，我们已经做了不少努力来防止这种泄露。

So this anyway, like, we've done we've made a bunch of efforts to avoid leakage.

但现在确实存在泄露，所以你暂时别拿它来评估你的预测模型。

We know there's leakage right now, so you shouldn't use it to evaluate your forecasting scaffold yet.

但问题是，如果模型只从档案资料里扫描和OCR识别的书籍获取数据，它怎么可能会接触到这些信息呢？

But, like, how is it getting that data if it's only being fed scanned, OCR ed books from archival sources?

因为档案资料里的日期经常出错。

Because archival sources have wrong dates in them all the time.

或者文本的日期本身就模糊不清，比如出了新版，或者后加了前言之类的。

Or it's kind of unclear what the, like, date of a text is because there's, like, an updated edition, or sometimes there's, like, a preface that's been added later.

有时候甚至在正文中间，有人会插进一条未来的备注，比如历史学家后来加的注释，什么什么的。

Or sometimes even just in the middle of the text, they'll be like, someone inserted some future notice, like, you know, note in like, later on, like, historians note, da da da da da.

所以这些细小的修改根本没法全查出来，但出版信息里的原始日期却还照旧保留着。

And so it's just really hard to check all these little edits that people make, and then they still maintain the original publication date on the metadata.

我明白了。

I see.

我问了Talkie它了解我什么，它说有个叫凯文·奥哈拉的人，1840年出生在都柏林，上过基督教兄弟会学校，后来当了老师。

I asked talkie what it knew about me, and it said, Kevin O'Hara, which is not my name, was born in Dublin in 1840 and having been educated at the School of the Christian Brothers, became a teacher in it.

之后他转行做了记者，有好几年都在《民族报》工作。

He afterwards adopted the profession of journalism and was for some years connected with the staff of The Nation newspaper.

它还说我还写过几首流行歌曲，比如《莫莉·阿斯托》和《爱尔兰移民》。

It also said I had written several popular songs, including Molly Astor and the Irish Immigrant.

显然这些大半都是错的，但它把我跟新闻业扯上了关系，这让我挺在意的，可能也说明数据里真有污染。

Now, obviously, most of that is wrong, but it did connect me to journalism, which I found interesting and maybe, like, some other evidence of some data contamination.

但问题是，这玩意儿是不是偷偷连了互联网？

But, like, is is this thing accessing the Internet in some way?

不然它怎么知道我——或者至少是模型里跟我挂钩的那个凯文·奥哈拉——是干新闻的？

Or, like, how would it have known that I or at least Kevin O'Hara, this character sort of connected to me in the model, was a journalist?

这问题问得真好。

You know, that's a great question.