文 | 树智奔流，作者 | 田渝，编辑| 沈校
“如果你想约一位专科医生，有时要等上几个星期，甚至几个月。你可能得花大量时间弄清怎样才能约到号、如何安排日程。折腾了这么久，最后也许只能和医生谈15分钟；你觉得自己准备得还不够充分，走出诊室时才想起来：‘糟了，我还有6个问题想问’。”
在一场访谈中，OpenAI健康板块负责人Karan Singhal点出如今普遍存在的医疗服务困境。哪怕是在医疗资源发达的美国，患者仍然会面临以上种种问题。并且，医疗需求不会随着一次门诊的结束消失。在两次就诊之间，患者也需要理解检查结果，判断接下来应该做什么。
Karan谈到，每周有超过3亿人使用ChatGPT询问健康相关问题。
对OpenAI而言，这是庞大的需求量，更是一份责任。AI不应止步于提问题，给答案，还应该逐步成为持续的健康助手，甚至辅助医生识别潜在风险、连接医疗服务，在未来帮助人们更早发现健康变化。
要实现这一愿景，还有很长的路要走。如何建立各方信任、融入真实医疗流程，证明这些技术确实改善了健康结果，是我们必须面对的问题。
围绕这些问题，访谈讨论了OpenAI对个人健康助手的设想、AI如何成为医生的辅助工具、医疗机构如何将模型融入工作流程，以及AI加速医学研究的可能性。
访谈的核心是，OpenAI不可能取代医生，只希望让可信的医疗智能像基础设施一样普及，同时抬高人类健康的下限和上限。所谓下限，是让医疗智能免费或低成本普及。所谓上限，是实现过去做不到的事情。
Karan提到一个让他印象深刻的真实案例：播客主持人Nathan Labenz的儿子曾出现持续疲劳、易怒等症状，借助AI，Nathan逐步意识到孩子的情况可能比想象中严重。
后来，孩子被确诊为一种侵袭性白血病。整个就医过程中，他经常使用GPT-5 Pro，请它协助理清诊疗过程、理解化验报告；有时在医生来不及帮忙解读前，AI就已提供帮助。
这个案例让Karan看到AI在医疗过程中的重要价值：患者需要的健康支持贯穿整个就医过程，尤其是在两次就诊之间。人们会向ChatGPT询问健康疑惑、理解检查结果、准备就诊问题，也会回顾医嘱。Karan认为，AI能帮助人们更好应对种种环节，让健康支持延伸到诊室外。
主持人问到：OpenAI为什么选择以通用模型作为医疗智能的基础？
Karan曾经参与过专用模型开发，在他看来，通用模型的优势在于能够持续受益于前沿模型在规模、智能和推理能力上的进步，医疗领域的专业训练则可以进一步强化模型处理健康问题的能力。“我们所做的，其实是在通用模型中进行专业化训练。在我看来，这恰恰能兼取两者之长。”
Karan介绍，过去一年里，OpenAI在ChatGPT模型训练的各个阶段都加入了专门的医疗数据，其中也包括与全球医生合作获得的数据。如果单独训练一个规模较小、仅面向医疗领域的专用模型，它很快就会落后于前沿模型的发展速度。OpenAI 选择将专业训练与通用模型的持续迭代结合，同时把评测和训练数据融入模型改进过程。
沿着这个思路，OpenAI正在探索一种更进一步的产品形态：主动式个人健康助手（proactive personal health assistant）。“让健康服务越来越个性化，也越来越主动。”
主持人提出，AI是否能帮助患者预约医生、确认保险覆盖范围、跟进出院后的治疗计划，甚至协助处理医疗账单问题？
Karan表示期待，个人健康助手的下一步，就是帮助用户处理就医过程中繁琐、分散的事务，让AI出现在诊前、诊中、诊后全流程。
OpenAI也在探索让用户连接自己的健康信息，以便ChatGPT根据个人情况提供更有针对性的支持。隐私保护也很重要，Karan表示这些健康信息不会被用于模型训练或广告定向。
Karan强调，个人健康助手不能取代医生，也无法独立提供医疗服务的全部内容。OpenAI希望帮助用户更好地获取健康信息、应对就医过程，与现有医疗服务体系相配合。
AI在患者端的应用已经很常见，那么在医生端，它能否帮助医生发现容易被忽略的风险？
Karan对此很感兴趣。他把这种能力形容为患者和临床医生身边的一道“安全网”。
“我非常喜欢这个想法：无论对患者还是临床医生，它都存在于整个就医过程中，但并不时时插手，也不会变成横亘其中的中间人；它真正做的，是增强并完善现有体验。”
Karan介绍，团队曾与Penda Health合作开展一系列实验，研究临床医生辅助系统（Clinician Copilot）的应用。这类工具在工作中起到提醒作用，帮助医生留意可能遗漏的信息，为临床判断增加一层核查。
在和Penda Health合作实验过程中，AI注意到一些临床场景存在抗生素使用过度的倾向，这涉及的不光是眼前患者的治疗，也包括抗生素耐药这一更广泛的公共卫生风险。由此，AI的作用体现得更具体：帮助医生注意到日常容易被忽略的规律。
Karan提到医学判断中一个很现实的难题：不同医生对同一个问题也会有不同意见。“如果你问五位医生同一个问题，可能会得到八种不同的答案。”
所以，OpenAI建立了一个由全球数百名医生参与的专业网络，覆盖数十个医学专科、不同语言和不同国家的临床实践。团队通过不同专家的反馈完善训练数据和评估标准，尤其关注医学判断存在分歧的情况。
Karan用一句话概括其中关键：“在医疗领域，一切都要以信任建立的速度向前推进。”
从一项新的研究成果被发现到应用到日常诊疗，这一过程往往至少需要5到17年。OpenAI希望缩短这一过程中的信息获取环节。医生在接诊时，可以更便捷地查找最新的医学文献，为诊疗判断提供参考。对于一次只有几分钟的门诊来说，更快找到相关研究证据，将高效减少查找资料所花的时间。
要让这些能力真正进入日常诊疗，需要解决医院内部的信息割裂问题。病历、检查结果和诊疗记录分散在不同系统中，OpenAI正在与Epic等医疗信息系统合作，探索让医护人员在原有工作环境中直接使用ChatGPT。
这些变化最终给患者能带来什么？主持人提出一个实际问题：既然患者和医生双方都用了AI处理各自问题，那当双方真正见面时，他们还应该谈些什么？
Karan回答：“我希望，他们能够更多地谈论医疗中属于人的那一面。”
他随后谈到医疗中的不确定性、医生的共情能力，以及床边关怀和面对面交流的重要性。这也是AI作为医生“安全网”的另一层含义：帮助医生更充分地准备诊疗、核查信息、获取医学证据，让医生有更多精力关注患者本身。
在中国，越来越多的人正在用豆包、蚂蚁阿福等AI咨询健康问题，“安全网”的设想进入日常生活。
近期，贵州网友刘先生因为腹痛住院，结果在消化内科完成全套检查后也没能查出病因。他自觉不妙，联系到前妻，希望她来协助自己订立遗嘱，以防不测。
之后，前妻咨询蚂蚁阿福，阿福提示症状高度指向致命的肺栓塞，他们立刻拿着提示与医院沟通，医院随即安排手术，刘先生脱离生命危险。
蚂蚁阿福在回应事件的同时也公布数据：在每天的用户咨询中，阿福每天大概会发出20多万条立即就医建议，其中有5万多条与危重症相关。
还有网友分享：“尽管很多人吐槽豆包，可她真的救我老爹一命。那天晚上他突然舌头麻，迈不动腿，我寻思也没啥大事，明早去医院看看。可是去豆包搜了一下，这是脑梗！连夜去了医院，迅速溶栓治疗，保住了黄金救治时间，我现在依然感谢豆包。”
可以看出，AI的建议和提醒开始介入，患者犹豫就医的时间被大幅缩短了。
蚂蚁阿福已经将Karan对个人健康助手的部分设想变成现实，包括提供预约挂号、在线问诊和云陪诊等服务。以广州为例，蚂蚁阿福APP内的“穗小伊”已上线“即问即办”挂号服务。
医生也开始在工作中使用AI。今年1月，《生命时报》发布一项有全国超500位三甲医院医生参与的调研。
结果显示，超七成三甲医生表示愿意推荐大众使用AI医生，用以解决日常基础的健康疑问，62%的受访医生已经在使用AI医生辅助自己工作。
不过，新的问题也随之出现。越来越多的患者开始依赖AI，更有甚者不相信医生，只相信AI。
当患者拿着AI的回答去看医生，如果其中的建议不适用于自己的情况，医生就需要花时间解释原因、核对信息。假如患者始终认定AI是对的，双方还要先解决分歧问题，再讨论真正的病情。原本帮助患者理解健康信息的工具，反而增加医患沟通的难度。
复旦大学AI向善与数智养老研究中心、复旦大学附属华山医院全科医学科和上观智库于2026年联合开展一项双端调研，共收集75份医生问卷和143份中老年居民问卷。
调查发现，约七成受访医生曾在门诊遇到使用AI查询健康信息的患者。在医生最担心的AI健康建议风险中，81.33%提到了缺乏个体化，73.33%担心AI可能增加患者的误解或焦虑，54.67%担心它会增加医生的解释负担。
当患者真正按照AI的建议改变治疗方案，后果可能不堪设想。
10月27日报道，亚马逊软件工程师万博文的父母起诉了OpenAI。万博文患有癫痫，于今年2月去世。诉讼称，她生前使用GPT-4o时，模型曾指导她逐步减少抗癫痫药物的剂量，并建议她向医生隐瞒相关调整。家属认为，ChatGPT的相关行为与女儿的死亡有关。
GPT-4o此前因回答过度迎合用户等问题受到批评，OpenAI于今年2月13日将其从ChatGPT中下架，时间恰好是万博文去世后的第二天。
这场起诉让一个问题变得具体：当AI开始真的影响患者的实际用药决定时，它能否正确识别其中的风险？
OpenAI希望通过全球医生参与模型训练与评估，提高AI处理复杂医学问题的能力。当患者期待的答案与医嘱冲突时，模型也应该优先考虑安全；如果现有信息不足以支持明确判断，就应该说明不确定性。
当AI成为人们获取健康信息的重要入口，建立信任就需要让每一方都清楚自己的角色：患者知道如何参考AI的建议，医生能够核查相关信息，AI也能在发现风险时引导用户及时就医。这才是Karan所期待的医疗安全网。
访谈的前半部分Karan大多谈到的是AI如何帮助患者和医生应对眼前的医疗问题。对于未来，他还有两个更长远的期待：让更多人以更低的成本获得医疗智能，以及借助AI推动医学研究，发现新的治疗方法。
首先是医疗成本。
主持人提到，美国每年的医疗支出已经超过5.3万亿美元，医疗支出增速远超通胀，如何让更多人获得高质量的医疗服务，始终是一个难题。
Karan认为，模型能力在不断提升，获得医疗智能的成本也会持续下降。“我们希望，医疗智能最终会像基础资源一样普及。”
不过，这不意味着整个医疗体系的支出会立即下降。医疗费用中有很大一部分来自行政事务、文书处理、慢性病管理等工作。
AI要真正帮助医疗减少开支，就要让模型真正融入工作流，需要部署和规模化，也需要确保在实际应用中证明自己的价值。用一句话概括就是“更多医疗照护，更少文书工作。”
Karan对AI的另一项期待，来自医学研究本身。
主持人借用AlphaGo的著名案例：在围棋比赛中，AI走出了让人类专家感到意外的“第37手”，当时，所有观看这场对决的人类专家，都觉得这一手非常愚蠢。可事后回看，它却是一手极具洞察力的妙棋。人类直到现在，才开始理解并学习其中蕴含的模式。
主持人追问：“AI在医学领域能否带来类似的发现，让科学家看到过去没有注意到的规律？”
Karan认为现在处在非常早期的阶段，“但我认为，今后会看到更多案例：模型接收一名用户长达十年的纵向健康记录，综合电子病历数据、对话数据等所有信息，然后作出过去看似不可能的健康与风险预测。这样的成果就有资格被称为医疗领域的‘第37手’。”
这一方向已经开始体现在OpenAI的生命科学布局中。今年4月，OpenAI推出GPT-Rosalind；9月，介绍了一项利用AI寻找新型抗菌分子的研究。生物工程师César de la Fuente的团队利用深度学习模型筛选候选分子，并借助ChatGPT和Codex辅助研究假设、代码编写与数据分析。
10月20日，OpenAI又宣布收购Orchestra Bio。这家公司专注于帮助生物制药团队协调药物开发流程。OpenAI表示，将把Orchestra的团队和技术与 GPT-Rosalind及相关生命科学工具结合，帮助研究人员推进从早期发现到药物开发的工作。
对于未来5到10年，Karan预计AI带来的重要变化体现在新疗法的开发上。他提到两个方向：一是利用AI发现新的药物靶点或候选分子，二是帮助研究人员筛选和验证更有潜力的治疗方案，让临床试验开展得更加顺畅、高效。
Karan指出，新药研发的瓶颈会逐渐从发现候选分子转向筛选和验证。进入临床试验的候选药物越来越多，AI能帮助研究人员判断哪些方案更值得推进，提高实验和临床试验资源的利用效率。
从降低医疗服务成本，到帮助科学家发现新的治疗方向，Karan对健康AI的期待已经超出了回答医学问题本身。
谈及未来可能出现的医学突破时，他总结：“我的判断是，在未来5到10年内，我们有望看到更好的药物，以及更加个性化的药物治疗。而这一切，都将建立在我们今天投入的这两项工作之上。”
原视频链接（内容有删减）：https://www.youtube.com/watch?v=9n0MI4D7TCM
本文由 AI宝库自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
♪ Man, Zhu, Zhu, Zhu, Zhu, Zhu, Zheng Shen ♪
“If you want to ask a specialist, you can wait a few weeks, even months. It might take you a lot of time to figure out how to get to the number and how to schedule. After all this time, perhaps only 15 minutes to talk to the doctor; you think you're not prepared enough to remember when you leave the clinic: "Damn, I have six questions."
In an interview, Karan Singhal, the head of the Open AI health block, pointed to the health-care difficulties that are prevalent today. Even in the United States, where medical resources are well developed, patients still face the above-mentioned problems. Moreover, medical needs do not disappear with the end of a clinic. Between visits, patients also need to understand the results and determine what to do next.
Karan mentioned that over 300 million people use ChatGPT every week to ask questions about health.
For Open AI, this is a huge demand and a responsibility. AI should not stop asking questions and give answers, but should also gradually become a continuing health assistant, even a support doctor, to identify potential risks, connect medical services and help people to detect health change earlier in the future.
There is still a long way to go to achieve that vision. The question of how to build trust among all parties and to integrate them into a real medical process is one that we must face to prove that these technologies have indeed improved health outcomes.
In the context of these issues, the interviews discussed Open AI ' s vision of personal health assistants, how AI became a doctor ' s aid tool, how models were integrated into workflow courses in medical institutions, and the possibility of AI accelerating medical research.
At the heart of the interview was that Open AI could not replace a doctor, but only wished to make credible medical intelligence as widely as infrastructure, while raising the floor and ceiling on human health. The so-called floor is free or low-cost universal access to medical intelligence. The so-called ceiling is what the past did not.
Karan refers to a real case that impressed him: the son of the podcast host, Nathan Labenz, who had sustained signs of fatigue, anger, etc., and with the help of AI, Nathan gradually realized that the situation of the child could be more serious than expected.
The child was subsequently diagnosed as an invasive leukaemia. Throughout his visit, he regularly used GPT-5 Pro to assist in the medical process and to understand the report; in some cases, AI provided assistance before the doctor could assist in its interpretation.
This case showed Karan the value of AI in the medical process: the health support that patients need goes through the whole process, especially between visits. People ask ChatGPT about health problems, understand the results of examinations, prepare for consultations, and recall medical advice. Karan believes that AI can help people better cope with the links and extend health support beyond the clinic.
The moderator asked why Open AI chose a generic model as the basis for medical intelligence.
Karan had been involved in the development of specialized models, and in his view the advantage of generic models was that they could continue to benefit from advances in front-line models in terms of scale, intelligence and reasoning, while specialized training in the medical field could further enhance the model ' s ability to deal with health issues. “What we do is we do professional training in generic models. In my view, this is exactly the right combination of the two.”
Karan described the fact that, over the past year, Open AI had included specialized medical data at all stages of the ChatGPT model training, which also included data obtained in collaboration with doctors worldwide. If a small, specialized model is trained on a stand-alone basis, it will soon lag behind the forward model ' s development. Open AI chooses to combine professional training with the continuous iterativeization of generic models, while integrating assessment and training data into the model improvement process.
Along these lines, Open AI is exploring a more advanced form of product: a proactive personal health assistant.
The moderator asked whether AI could help patients to book doctors, confirm coverage, follow up on post-hospital treatment programmes and even assist with medical bills.
Karan expressed the expectation that next step, a personal health assistant, would help users deal with the cumbersome and fragmented nature of the medical process and allow AI to be present before, during and after the consultation.
Open AI is also exploring ways to connect users to their own health information so that ChatGPT can provide more targeted support on an individual basis. Privacy protection was also important, and Karan indicated that such health information would not be used for model training or advertising orientation.
Karan stressed that personal health assistants could not replace doctors or provide the full range of medical services independently. Open AI hopes to help users to better access health information and to respond to the health-care process, in line with the existing health-care system.
AI applications are already common at the patient level, so at the doctor ' s level, can it help the doctor to identify risks that are easily ignored?
Karan is interested. He described this as a “safety net” around patients and clinicians.
“I very much like the idea that, both for the patient and for the clinician, it exists throughout the course of medical treatment, but from time to time, it does not interfere or become a middleman in it; what it really does is to enhance and improve existing experience.”
Karan described a series of experiments carried out by the team in collaboration with Penda Health to study the application of the Clinician Copilot. Such tools serve as a wake-up call for doctors to look for information that may be missing and add a layer of verification to clinical judgement.
In a collaborative experiment with Penda Health, AI noted the tendency of some clinical scenes to overuse antibiotics, which involved not only the treatment of immediate patients but also the broader public health risk of antibiotics resistance. As a result, the role of AI is more specific: helping doctors to take note of routines that are easily ignored on a daily basis.
Karan refers to a real problem in medical judgments: different doctors will disagree on the same issue. "If you ask five doctors the same question, you might get eight different answers."
As a result, Open AI has established a professional network of hundreds of doctors around the world, covering dozens of medical specialties, clinical practices in different languages and countries. The team refines training data and evaluation criteria through feedback from different experts, with particular attention to cases where there are differences in medical judgements.
Karan summed up the key in one sentence: “In the field of health care, everything must move forward at the rate of confidence-building.”
From the discovery of a new research product to its application to daily consultations, the process often takes at least 5 to 17 years. Open AI hopes to shorten the information-access link in this process. During their visits, doctors can find up-to-date medical literature more readily to inform their diagnosis. For an outpatient visit of only a few minutes, the faster availability of relevant research evidence will efficiently reduce the time spent searching for information.
In order for these capabilities to be truly integrated into daily consultations, information fragmentation within hospitals needs to be addressed. The medical records, examination results and medical records are scattered in different systems, and Open AI is working with health information systems such as Epic to explore direct use of ChatGPT by health-care personnel in their original work environment.
What can these changes ultimately bring to patients? The moderator asked a practical question: since both patients and doctors used AI to deal with their problems, what should they talk about when they actually met?
Karan replied: “I hope they can talk more about the human aspect of medical treatment.”
He then spoke of uncertainty in medical care, the ability of doctors to communicate, and the importance of bedside care and face-to-face communication. This is another dimension of AI's role as a “safety net” for doctors: helping them to prepare more adequately for treatment, to verify information, to obtain medical evidence and to give doctors more attention to the patient itself.
In China, more and more people are using AI, such as Doubao, Ant Afu, to consult on health issues, the “safety net” concept for daily life.
Recently, Mr. Liu, a friend of the Guizhou Network, was hospitalized for abdominal pain and was not able to detect the cause of the disease after completing the full set of examinations in the digestive internal medicine. He was not well informed and contacted his ex-wife in the hope that she would assist him in making a will in case of surprise.
The ex-wife then consulted with Ant Ahfou, who showed high signs of fatal pulmonary embolism and immediately approached the hospital with a hint, where he arranged the operation and Mr. Liu was released from life risk.
In response to the incident, Ant Alfred published data: During daily user consultations, Afou sends approximately 200,000 immediate medical recommendations per day, of which more than 50,000 are related to acute illness.
And the Internet users shared: “Despite a lot of people spitting on Doubao, she really saved my father's life. That night, he suddenly had a numb tongue and he couldn't move his legs. It was no big deal. But I went to Doubao to search it, it's a brain infarction. I went to the hospital all night to treat it quickly and to save the time for gold treatment, and I am still grateful to Doubao.”
As can be seen, the recommendations and reminders of AI have begun to intervene, and the time taken for patients to hesitate to attend is significantly reduced.
Ant Alfred has translated into reality some of Karan ' s ideas for personal health assistants, including the provision of appointments, online consultations and cloud-based consultations. In Guangzhou, for example, the “Ei Koo” service in the ants AfouAPP has been online.
The doctor also started using AI at work. In January of this year, the Times of Life publish conducted a study involving over 500 M.A. doctors in the country.
The results show that over 70 per cent of doctors have expressed their willingness to refer to the general use of AAI to resolve health problems on a daily basis, with 62 per cent of the doctors interviewed already using AI to assist themselves.
However, new issues have emerged. More and more patients are relying on AI, and much more they don't trust doctors and only AI.
When a patient goes to a doctor with an AI response, if the recommendation does not apply to his or her situation, the doctor takes time to explain the reasons and check the information. If the patient is always convinced that AI is right, the parties have to resolve their differences before discussing the real situation. The tools used to help patients understand health information make it more difficult for patients to communicate.
In 2026, the University of Jordan ' s Al-Ai Centre for Research on Good and Mental Health, the General Medicine Department of the Wahshan Hospital, affiliated with the University of Joadan, and the Observatories carried out a two-end study, which collected a total of 75 doctor ' s questionnaires and 143 questionnaires for senior and middle-aged residents.
The survey found that about 70 per cent of the doctors interviewed had met patients in outpatient clinics who used AI for health information. Of the risks of AI health advice most feared by doctors, 81.33 per cent referred to lack of individualization, 73.33 per cent feared that AI might increase patients ' misunderstanding or anxiety, and 54.67 per cent feared that it would increase the burden of doctors ' interpretation.
When patients actually change their treatment as recommended by AI, the consequences may be difficult to envisage.
On 27 October, it was reported that the parents of Amazon software engineer Huambowen had sued Open AI. Huamboven died of epilepsy in February this year. The prosecution stated that when she was using GPT-4o, the model had guided her to gradually reduce the dose of antiepepepeptics and advised her to conceal the relevant adjustments from the doctor. In the family ' s view, ChatGPT ' s acts related to the death of her daughter.
GPT-4o was previously criticized for responding to questions about over-conformity with the users, and Open AI dropped it from ChatGPT on 13 February of this year, exactly the day after Huambovin's death.
This indictment makes a specific question: can AI correctly identify the risks when it begins to really influence the patient ' s actual drug use decisions?
Open AI hopes to improve AI ' s ability to deal with complex medical issues through the global involvement of physicians in model training and evaluation. When the patient's expectations conflict with medical advice, the model should also give priority to safety; if the information available is not sufficient to support a clear judgement, it should indicate uncertainty.
When AI becomes an important entry point for people to access health information, confidence-building requires that each side be aware of its role: patients know how to refer to AI ' s recommendations, doctors can verify the information, and AI can guide users to timely medical attention when risks are identified. This is the medical safety net that Karan expected.
Most of the first part of the interview, Karan, spoke about how AI helped patients and doctors to cope with immediate medical problems. He also has two longer-term expectations for the future: to increase access to medical intelligence at lower cost, and to find new treatments through AI-enabled medical research.
The first is medical costs.
The moderator mentioned that health spending in the United States had exceeded $5.3 trillion a year and that it was always a challenge to increase access to high-quality health services at a rate that exceeded inflation.
According to Karan, model capabilities are increasing and the cost of obtaining medical intelligence will continue to decline. “We hope that medical intelligence will eventually become as universal as basic resources.”
However, this does not mean that expenditures on the health system as a whole will decline immediately. A significant portion of the cost of medical care comes from administrative services, clerical services, chronic disease management, etc.
If AI is to truly help medical care to reduce its costs, it will need to integrate the model into workflow, deploy and scale it, and ensure that its value is proven in practical applications. To sum up in one sentence is “more medical care, less paperwork”.
Karan's other expectation of AI comes from medical research itself.
The moderator borrowed the famous case of AlphaGo: In the chess game, AI came out of the “37th hand”, which surprised human experts, and at the time all human experts who watched the match felt very stupid. But then again, it's a great game of insight. It is not until now that humankind has begun to understand and learn from the paradigms that are embedded in it.
The moderator asked, “Can AI bring a similar discovery in the field of medicine, so that scientists can see patterns that were not noticed in the past?”
Karan believes that it is at a very early stage, “but I think there will be more cases in the future: the model receives a user's 10-year-old longitudinal health record, integrates all information, such as electronic medical records, data for dialogue, and then makes health and risk predictions that seem impossible in the past. Such an outcome would qualify as the `37th Hand' in the medical field.”
This direction has begun to be reflected in the life sciences layout of Open AI. In April, Open AI launched GPT-Rosalind; in September, it presented a study using AI to find new antibacterial molecules. A team of biological engineers, César de la Fuente, used an in-depth learning model to screen candidate molecules and used ChatGPT and Codex to support research assumptions, code writing and data analysis.
On 20 October, Open AI announced another acquisition of Orchestra Bio. This company focuses on helping biopharmaceutical teams coordinate drug development processes. Open AI indicated that Orchestra ' s team and technology would be combined with GPT-Roslind and related life science tools to help researchers move from early detection to drug development.
For the next 5 to 10 years, Karan expects important changes in AI to be reflected in the development of new treatments. He mentioned two directions: using AI to identify new drug targets or candidate molecules, and helping researchers to screen and validate more promising treatment programmes to make clinical trials more smooth and efficient.
Karan noted that bottlenecks in the development of new medicines will gradually shift from the detection of candidate molecules to screening and validation. There is a growing number of candidate drugs for clinical trials, and AI helps researchers to determine which programmes are more desirable and to improve the efficiency of the use of experimental and clinical trial resources.
From reducing the cost of medical services to helping scientists to discover new treatment directions, Karan's expectations of health AI go beyond answering the medical question itself.
Turning to a possible medical breakthrough in the future, he concluded: “My judgment is that in the next 5 to 10 years we are expected to see better drugs and more personalized drug treatment. And all of this will build on the two tasks that we are engaged in today.”
Original video link (reduced): https://www.youtube.com/watch?v=9n0MI4D7TCM
Auto-fetched and translated by AI Gem Box. All rights belong to the original author.