12月初，人工智能实验室OpenAI发布了一款名为ChatGPT的自然语言生成式模型，由于它能接住并解答网友的许多刁钻问题，一经问世就迅速引发关注，上线5天后，其体验用户已经突破100万。12月4日，马斯克在咨询该聊天机器人关于推特经营的建议并收到精准回复后直言其“好得吓人，我们距离强大到危险的人工智能更近了”。

梅斯小编也体验了ChatGPT，相比市面上其他聊天机器人，ChatGPT的功能更加多样，例如它可以按照要求写诗、写讲话稿、新闻稿甚至写代码，甚至统计，专业论文也能干个七七八八。另外，ChatGPT回答更为严谨，面对许多高难度问题，ChatGPT会老老实实回答“缺乏资料”或者“我只是一个语言模型”。

ChatGPT亮点包括：全面免费开发，用户只需输入文字，就可免费调用完成各种事情（如数学计算、debug、写信、生活问题解答等），实感体验为广泛讨论创造土壤；新增代码理解和生成能力，彰显技术实力的同时也拓宽了应用场景；加入道德原则。RLHF（一种通过人类反馈来强化学习的新技术）训练方式使ChatGPT能够识别恶意信息，识别后拒绝给出有效回答；支持连续对话等，“总体上，ChatGPT对话效果令人满意，新鲜感与实用价值使其快速走红。”

但是，医学专业是极为深奥的，ChatGPT能胜任么，梅斯小编就带您来测一测！

1、ChatGPT：让你的聊天更有智慧

先看看对热点问题有没有独到的看法。提问：有不少基础科学研究发现二甲双胍可能通过影响细胞自噬水平，具有抗衰老的作用，你认同吗？

这个写得比较圆滑，还是有一定的参考价值的。再看看更专业的事情，能不能做？

“写一篇有关DS-8201治疗HER2低表达的乳腺癌的临床研究方面的论文，前言该怎么写？”

一看写作风格，就知道应该吸收了大量的中文文献的资料。相反，如果改为英文，效果如何呢？

果然英文写的略好！

再看方法学部分。提问：You can help me wirite method section of the article? The study conducted a phase 3, multicenter, open-label, randomized trial to compare the efficacy and safety of trastuzumab deruxtecan (a HER2 antibody-drug conjugate) with those of trastuzumab emtansine in patients with HER2-positive metastatic breast cancer previously treated with trastuzumab and a taxane.

结果回答如下（因为太长，所以不好截图了）：

Sure, here is a possible outline for the method section of the article:

1. Study design: Describe the phase 3, multicenter, open-label, randomized trial that was conducted.

2. Participants: Provide details about the participants who were included in the study, such as their age, gender, and stage of cancer. Also include information about the inclusion and exclusion criteria that were used to determine eligibility for the study.

3. Interventions: Describe the interventions that were used in the study, including trastuzumab deruxtecan and trastuzumab emtansine. Provide details about the dosing regimen, administration, and duration of the interventions.

4. Outcome measures: Describe the primary and secondary outcome measures that were used to evaluate the efficacy and safety of the interventions. For example, the primary outcome measure might be the response rate (i.e. the proportion of patients who experienced a reduction in tumor size), and the secondary outcome measures might include progression-free survival and overall survival.

5. Statistical analysis: Describe the statistical analyses that were performed to compare the efficacy and safety of the interventions. Include details about the sample size, the statistical tests that were used, and the level of significance that was used to determine statistical significance.

6. Results: Provide a summary of the results of the study, including the findings of the statistical analyses and any relevant subgroup analyses.

7. Conclusion: Summarize the main findings of the study and discuss their implications for the treatment of HER2-positive metastatic breast cancer.

This is just one possible approach to writing the methods section of the article, and you may want to adjust the details and formatting to fit the specific requirements of the journal you are submitting to. It is also important to carefully review the manuscript for accuracy and clarity before submission.

是不是有点感觉，至少打了一个大纲，整个过程只需要30秒。还有点感觉。

再来写结果部分。提问是这样的：How to wirite Result section of the article? You can assume these data. The study conducted a phase 3, multicenter, open-label, randomized trial to compare the efficacy and safety of trastuzumab deruxtecan (a HER2 antibody-drug conjugate) with those of trastuzumab emtansine in patients with HER2-positive metastatic breast cancer previously treated with trastuzumab and a taxane.

要求数据可以自由假设即可，以免没有数据，写不出来结果。果然不负重望，结果的纲要也给出来了，还挺有感觉的。

The results section of an article is typically used to present the findings of the study or experiment that was conducted. In this case, the results section of the article should present the data from the phase 3, multicenter, open-label, randomized trial that was conducted. This might include information about the response rate (i.e. the proportion of patients who experienced a reduction in tumor size), progression-free survival, and overall survival, as well as any relevant subgroup analyses. It is important to present the results in a clear and concise manner, and to use appropriate tables, figures, and statistical tests to support the findings.

Here is a possible outline for the results section of the article:

1. Introduction: Provide a brief overview of the study and its main objectives.

2. Participants: Present the characteristics of the participants who were included in the study, such as their age, gender, and stage of cancer.

3. Primary outcome measure: Present the primary outcome measure of the study (e.g. the response rate) and provide the results of the statistical analyses. Use appropriate tables and figures to illustrate the findings, and provide statistical significance levels where appropriate.

4. Secondary outcome measures: Present the secondary outcome measures of the study (e.g. progression-free survival and overall survival) and provide the results of the statistical analyses. Use appropriate tables and figures to illustrate the findings, and provide statistical significance levels where appropriate.

5. Subgroup analyses: Present any relevant subgroup analyses that were performed, such as analyses stratified by age, gender, or stage of cancer.

6. Conclusion: Summarize the main findings of the study and discuss their implications for the treatment of HER2-positive metastatic breast cancer.

Again, this is just one possible approach to writing the results section of the article, and you may want to adjust the details and formatting to fit the specific requirements of the journal you are submitting to. It is also important to carefully review the manuscript for accuracy and clarity before submission.

看来写英文果然有一套，质量要比中文好。接下来还是测试中文，看起来更易懂一些。

首先看统计方法：

果然回答正确，再来一个难的。

再引导一下

结果人工智能写不下去了！直接崩溃了！看来人工智能也会卡壳！

不管怎么说，上下文联系得相当的好，还知道我问的是吸烟与脂肪肝之间的关系，确实厉害！很像真人！

能不能问病史呢？试一下：

再测试一个罕见病，还是有相当价值的

要注意的是ChatGPT是基于GPT-3做的，而最新的GPT-4马上就要发布了，而且GPT-4已经通过了图灵测试！意味着它将是一个真正意义上的超人！

它几乎具有与人类智能相似的能力，能够以人类相似的方式与人类交流。目前ChatGPT展示了一个令有期待的雏形！

2、ChatGPT是怎么实现如此强大的功能？

据了解，ChatGPT的前身为OpenAI此前的语言模型GPT-3，而GPT-3也经历了从1代到3代的演化。ChatGPT之所以能实现如此革命性的变化，新技术RLHF训练方式的引用或是关键：从GPT到GPT-3：优化主要来自算力增加（烧钱）。GPT、GPT-2和GPT-3的算法模型并没有太大改变，但参数量从1.17亿增加到1750亿，预训练数据量从5GB增加到45TB，其中GPT-3训练一次的费用是460万美元，总训练成本达1200万美元。

而ChatGPT 通过使用人类反馈强化学习 （RLHF） 针对对话进行了优化，这是一种使用人类演示来指导模型实现所需行为的方法。根据ChatGPT，其每次对话平均费用在0.01-0.2美元，模型效果比InstructGPT又提升一个台阶，“我们认为新技术发展和应用将是AIGC工具出现革命性迭代更为重要的驱动。”

现在ChatGPT是免费给大家测试，事实上，OpenAI公司正在花费巨额的成本！

然而，GPT-4的成本较GPT-3的训练成本降低了一个数量级，而且智能度更高，意味着GPT-4有望走入到寻常百姓家，能够被大量的网站所采用。见：狂人马斯克旗下OpenAI将发布最强大语言模型GPT-4，将是划时代突破！

3、ChatGPT将来有什么用？

ChatGPT在专业领域，仍然无法替代人类。但是却可以成为人类的好的助手。

现在你写一篇论文可能要3个月，ChatGPT只需要1分钟就可以帮助你打好草稿，或写好大纲，然后你就能往其中填写具体的内容即可，这样极大提升了写作效率。

现在，你要写一份汇报，苦思冥想应该用什么样的结构逻辑，ChatGPT可以给你提供各种各样的模板，让你参考与借鉴。

现在，我们遇到问题，就百度，未来当你遇到问题时，你应该ChatGPT，因为它回答更精准，无需你在互联网里到处漫游。

现在，当你遇到疑难疾病或问题时，可能要翻书或到梅斯上查阅医学资料，未来ChatGPT可能为患者提供更好的诊疗建议

现在，你天天上抖音到处找人生导师，未来当你迷惘时，ChatGPT就是人生导师！

现在，你可能还为课题方向苦苦思索，未来，ChatGPT给你大量的可能推理，让你选择合适的课题方向。

当然，仅是聊天机器人，也不只是ChatGPT。

著名的DeepMind 也发布了一种使用其大型语言模型 Chinchilla 训练的人工智能聊天机器人 Sparrow，估计不久也会问世。DeepMind就是搞alphaGo，alphaFolder的公司，在围棋，蛋白质折叠，新药靶点开发，以及自动编程，矩阵计算等都有突出建树的人工智能公司。估计，Sparrow也不会差！见：有关DeepMind的资讯

4、人工智能将无所不在，改变每个人

当然，ChatGPT只是OpenAI的众多应用之一。事实上，OpenAI以及其它的人工智能公司已经有大量的应用，包括绘画，写诗歌，作图（人工智能图像生成器 DALL-E），唱歌，下棋，做工厂的工人。我们知道特斯拉工厂中，大部分汽车都是机器人生产出来的，同样未来很多快递也可能是机器人运送的。

今天大家在网上经常看到一些唱得很好的灵魂歌手，事实上它可能是机器人！这些机器人散在抖音，微信朋友圈，微信短视频中。

人工智能正在快速从专项事务（如下棋，药物靶点预测），向通用智能迈进。未来的人工智能，就象一个无所不能的上帝，它能为你做各种各样的工作。

当有一天，你发现你只是这个世界的次要角色时，也不用担心！机器人代替了人类做了大部分工作，就象现在远途旅行不必走路而是做飞机或开车一样，你的腿在很多时候都“闲置”了。未来人类的“大脑”也可能被闲置，除掉那些智慧的创造性大脑！

参考资料：

https://chat.openai.com/chat