Transcriptly

How AI Is Changing Healthcare for Patients & Doctors | Dr. Fei-Fei Li & Dr. Andrew Huberman

Huberman Lab Clips10:06Added Aug 31, 2026

Dr. Fei-Fei Li and Dr. Andrew Huberman discuss how artificial intelligence is transforming scientific discovery and healthcare, focusing on AI's ability to s...

Watch on YouTube →
Contributed by 刘嘉琪

Transcript

Transcript format

  1. AI in Healthcare & Biology

  2. 00:00Since uh you're here, I'm going to go
  3. 00:02next to something that I think most
  4. 00:03everybody would agree would be a
  5. 00:05wonderful thing if it existed and it's
  6. 00:08already starting to happen, which is the
  7. 00:10use of AI to augment health discovery,
  8. 00:13treatment of disease and so on. So using
  9. 00:16the AlphaGo example from before and
  10. 00:18people surely still remember the cat
  11. 00:21example, those just follow certain
  12. 00:23rules. Alph Go is very complicated set
  13. 00:25of rules, but if you learn them, there's
  14. 00:27a constrained set of rules.
  15. 00:29With the cat, it seems unconstrained,
  16. 00:32like infinite possibilities, but it's
  17. 00:33constrained enough that machines and
  18. 00:35humans can learn it really well.
  19. 00:37When you start getting into medicine,
  20. 00:40there are rules of medicine. There are
  21. 00:42rules of science. You have a question,
  22. 00:44you pose a hypothesis, you test the
  23. 00:46hypothesis, you try and rule out your
  24. 00:47hypo and so on like the the scientific
  25. 00:49method. And in medicine, every field has
  26. 00:52its methods. We observe, we observe
  27. 00:55disease, we observe who recovers, we
  28. 00:57have a case report, we do a randomized
  29. 00:58control trial. So there are rules and
  30. 01:00the internet knows these rules. So LLMs
  31. 01:02can be used to mine health information
  32. 01:04very well because there are constrained
  33. 01:06rules. But I think you and I both know
  34. 01:09because I also consider you a biologist
  35. 01:11that the rules of biology are still
  36. 01:14revealing themselves to us. Which is not
  37. 01:16to say that the dermatologists,
  38. 01:17neurosurgeons, and oncologists don't
  39. 01:19know what they're doing, but they're
  40. 01:21doing what they're doing within a
  41. 01:22constrained set of rules that they
  42. 01:23learned. And even if they continue to
  43. 01:25learn and update them,
  44. 01:27it's every month it seems now that a
  45. 01:29discovery comes out that violates the
  46. 01:31rule. Like I learned that action
  47. 01:33potentials are unitary. They always look
  48. 01:35the same. You either fire or not.
  49. 01:37But there was a paper not but 12 years
  50. 01:39ago that showed that the shape of an
  51. 01:41action potential can vary quite a lot.
  52. 01:42It was published in Nature. Mhm.
  53. 01:45Everyone saw it and then no one wanted
  54. 01:46to deal with it.
  55. 01:47It's just too much. It changes the rule.
  56. 01:50Neurons are supposed to be either graded
  57. 01:52or all are one. And the all I mean it's
  58. 01:54in every single textbook. So now if I
  59. 01:56take a bunch of neural activity and I
  60. 01:58give it the rule, oh well you know
  61. 02:00action potentials can be big, they can
  62. 02:01be small in the same neuron. It
  63. 02:03completely confuses everything we
  64. 02:05understand about neuroscience
  65. 02:07and it just our understanding of the
  66. 02:08brain just breaks down to zero. Yeah.
  67. 02:10But if you gave AI the rule that it
  68. 02:12could be, you know, a hundred different
  69. 02:14shapes of this signal, well, AI could
  70. 02:17probably do a lot more than even the
  71. 02:20very very best graduate student at dare
  72. 02:22I say Stanford or to be fair MIT or
  73. 02:24Caltech. I don't think it can do it and
  74. 02:28it can do it like in the duration of
  75. 02:29this question, which admittedly is a bit
  76. 02:31long. So, I'd like to get your thoughts
  77. 02:33on how is it that humans in health care,
  78. 02:36the general public and AI can
  79. 02:38collaborate to help solve disease and
  80. 02:41ideally come up with new rules for
  81. 02:43discovery so that we can finally
  82. 02:45understand our biology at a level that
  83. 02:47can really change the course of humanity

  84. AI for Scientific Discovery

  85. 02:50for the better.
  86. 02:50Yeah. No, Andrew, this is probably
  87. 02:53perhaps you touch one of the most
  88. 02:55exciting usage of AI, which is
  89. 02:57scientific discovery. And in the case of
  90. 03:00biio medicine, you know, scientific
  91. 03:02discovery directly connects to human
  92. 03:04health and diseases, I think we're we're
  93. 03:08ready for complete re rewriting of how
  94. 03:11scientific discovery can be done because
  95. 03:14for ages, I don't even know how long, it
  96. 03:17relies on smart humans retaining what
  97. 03:20they have learned from other smart
  98. 03:22humans and and and doing things at the
  99. 03:25speed of our own muscles, I guess, you
  100. 03:28know. Most likely of course there's like
  101. 03:32super colliders and and all that but by
  102. 03:34and large the the ways of doing
  103. 03:37scientific discovery
  104. 03:40human brain or scientists brain are the
  105. 03:43only central character in this process.
  106. 03:48Now we have a new tool whose brain that
  107. 03:52can retain humongous amount of
  108. 03:56information can help us synthesize
  109. 03:58knowledge can go across disciplines in
  110. 04:01ways that you and I cannot go. So for
  111. 04:04example we happen to be both in the
  112. 04:06vision neuroscience AI domain.
  113. 04:10I know nothing about you know oactory
  114. 04:14zero like I don't even know how to spell
  115. 04:16most of probably the these words in that
  116. 04:18our colleagues know right so it's so
  117. 04:21hard for our brain but now we have a
  118. 04:24tool that can break open so so I think
  119. 04:27that we need to change we need to use
  120. 04:30this tool we absolutely I I was just
  121. 04:33thinking 150 or I don't know exactly
  122. 04:36when years ago we electricity
  123. 04:40changed everything in in in our life,
  124. 04:44right? I'm sure that's a moment we were
  125. 04:47thinking about how
  126. 04:50the changes, the opportunities, the
  127. 04:52scary moment. I think we have to come to

  128. Using AI for Personal Diagnosis

  129. 04:55reckon that scientific discovery is one
  130. 04:59of the most exciting opportunity for AI
  131. 05:04and for health, right? How information
  132. 05:06can be synthesized, how information can
  133. 05:08be presented not only to clinicians but
  134. 05:12also to patients and how patients can
  135. 05:15participate in that process from
  136. 05:17diagnosis to treatment is also there's
  137. 05:21just so much we can do now.
  138. 05:24Yeah. I mean AI I won't say AI is better
  139. 05:27than all doctors but AI was able to
  140. 05:30disambiguate vertigo from low blood
  141. 05:33pressure for me a few months back and
  142. 05:36one of the people who got it wrong is a
  143. 05:39ENT who works on the vestibular system.
  144. 05:41What information did you provide just
  145. 05:43your subjective?
  146. 05:44My subjective experience over a day or
  147. 05:47two.
  148. 05:47Okay. Um, turns out it was a medication
  149. 05:49that a doctor had prescribed me that I
  150. 05:51had a like a mild but adverse event. And
  151. 05:54it's a weird thing to step and feel like
  152. 05:56the whole world's dropping down and then
  153. 05:58kind of spinning and I thought, "Oh my
  154. 05:59goodness, this is like feels like
  155. 06:00vertigo." But I remember dizzy and
  156. 06:02lightheaded or different. So I started
  157. 06:04like looking into that and then and um
  158. 06:06sure enough it was a it was a blood
  159. 06:08pressure issue. It brought brought my
  160. 06:10blood pressure, excuse me, down too low.
  161. 06:12And but I consulted we know some smart
  162. 06:15doctors. Um none of these were at
  163. 06:16Stanford. I will say that this is the
  164. 06:18truth. But it was
  165. 06:19we should just be intellectually honest.
  166. 06:21But it's just remarkable. And when I ran
  167. 06:22it back to them, they were like, "That's
  168. 06:24really incredible." You know, had you
  169. 06:25not been on the phone with me and in my
  170. 06:26clinic, I would have been able to do
  171. 06:28some additional testing to be fair. But
  172. 06:30this was zero cost. It took a morning to
  173. 06:33know if I drank some uh electrolytes at
  174. 06:37what I would have thought would be
  175. 06:38excessive level that by two hours later,
  176. 06:42I would be fine. Now, of course, there's
  177. 06:43the possibility of a placebo effect
  178. 06:45here, but two hours later, I was fine.
  179. 06:48And so, it's also very consoling to the
  180. 06:50patient
  181. 06:51to have this. And so, it's not to say
  182. 06:53don't go to a doctor, but it it's

  183. Robotic Surgery & AI Collaboration

  184. 06:55incredible. I mean, this exists now.
  185. 06:56The doctor can use this tooling. By the
  186. 06:58way, I have a very interesting example.
  187. 07:01You know that we have to reschedu this
  188. 07:03uh our conversation because my father
  189. 07:06was going through a surgery right at
  190. 07:08Stanford uh with an incredible surgeon.
  191. 07:11But the surgery was done by a robot, the
  192. 07:14Davinci robot system because it was a
  193. 07:17liver surgery and the surgeon,
  194. 07:19incredible surgeon was driving the
  195. 07:21robot. So it was a deep human machine
  196. 07:25collaboration. After the surgery, I
  197. 07:27asked the surgeon, I said, "Do you
  198. 07:30imagine if say you've done a million,
  199. 07:33which is impossible for a surgeon, but
  200. 07:35human surgeon, but let's collect all of
  201. 07:38human surgeons uh for for this liver,
  202. 07:41this type of liver surgery data. Can we
  203. 07:45possibly train a automatic AI to do
  204. 07:49this?" The answer was not clear. So we
  205. 07:53went a little bit down the rabbit hole
  206. 07:55because liver is a very complicated
  207. 07:57organ. It's extremely vascular. It has a
  208. 08:00lot of vessels and everybody's liver is
  209. 08:04very different. So given the reality of
  210. 08:08how many patients undergo liver surgery
  211. 08:11per year, even if you aggregate um the
  212. 08:15world's liver patient um surgeries, you
  213. 08:18might not have enough data to train
  214. 08:21these algorithm. So this speaks of a
  215. 08:24very important fact that um AI learns
  216. 08:27from patterns. When the patterns are not
  217. 08:30abundant,
  218. 08:32then we have to be careful. We have to
  219. 08:33know how to use AI or how not to use AI.
  220. 08:37You know in this case that having a
  221. 08:39human collaborating with the robot is
  222. 08:42way better than a underlearned robot
  223. 08:45doing the surgery by itself. But the
  224. 08:48same issue might be true for surgeons
  225. 08:49because how many surgeries a surgeon can
  226. 08:52get trained on. So these are
  227. 08:54opportunities that humans and AI can
  228. 08:57totally collaborate with and might
  229. 09:00reveal the best result. Right now the
  230. 09:03future remains to be seen. Can we create
  231. 09:05a artificial simulation of a liver that
  232. 09:08we can now train infinite possibility?
  233. 09:11These are all incredibly open scientific
  234. 09:15possibilities that is waiting ahead of
  235. 09:18us. But then there are
  236. 09:21uh situations like your situation where
  237. 09:24the vertigo versus low blood pressure
  238. 09:27probably have been reported so many
  239. 09:29times that in the database there's
  240. 09:33enough of that that AI has learned that
  241. 09:36so we can then now take advantage of
  242. 09:38that for people who don't have immediate
  243. 09:41access to doctors.
  244. 09:43Amazing. Is your father's surgery went
  245. 09:45okay?
  246. 09:45It did. It actually
  247. 09:47lost
  248. 09:4910x less blood than a typical surgery uh
  249. 09:54thanks to the laparoscopic capability of
  250. 09:57a robot.