WEBVTT

1
00:00:00.080 --> 00:00:01.919
- Hello, everyone. This is Molly Gamble with Becker's

2
00:00:01.919 --> 00:00:03.759
- Healthcare. Thank you for tuning in to the

3
00:00:03.759 --> 00:00:05.139
- Becker's Healthcare podcast.

4
00:00:05.519 --> 00:00:07.599
- Today, we're going to talk about beyond the

5
00:00:07.599 --> 00:00:08.099
- doomerism,

6
00:00:08.720 --> 00:00:11.279
- what safe health care AI looks like in

7
00:00:11.279 --> 00:00:11.779
- practice.

8
00:00:12.464 --> 00:00:15.905
- Joining me for today's conversation is Munjal Shah,

9
00:00:15.905 --> 00:00:19.905
- cofounder and CEO of Hippocratic AI. Munjal, thanks

10
00:00:19.905 --> 00:00:22.404
- for joining me today and being my participant,

11
00:00:22.625 --> 00:00:24.324
- my guest for this big topic.

12
00:00:24.945 --> 00:00:26.704
- Oh, yeah. I'm, really excited to be here.

13
00:00:26.704 --> 00:00:27.845
- Thanks for having me.

14
00:00:28.289 --> 00:00:28.789
- Likewise.

15
00:00:29.250 --> 00:00:31.410
- For our audience who hasn't caught up with

16
00:00:31.410 --> 00:00:33.250
- us before, Munjal, can you share a bit

17
00:00:33.250 --> 00:00:35.409
- more about yourself and your work in health

18
00:00:35.409 --> 00:00:37.109
- care just to get them up to speed?

19
00:00:38.049 --> 00:00:40.129
- Yeah. I'm a serial entrepreneur. This is actually

20
00:00:40.129 --> 00:00:41.784
- my fourth company. So,

21
00:00:42.225 --> 00:00:44.384
- I'm a glutton for startup punishment, but,

22
00:00:45.744 --> 00:00:46.244
- and

23
00:00:46.784 --> 00:00:48.784
- what we've built here at Hippocratic is really

24
00:00:48.784 --> 00:00:49.344
- we build,

25
00:00:50.304 --> 00:00:52.784
- agents that call patients on the phone and

26
00:00:52.784 --> 00:00:54.965
- conduct clinical patient facing activities,

27
00:00:55.450 --> 00:00:57.689
- largely within the scope of a registered nurse

28
00:00:57.689 --> 00:01:00.170
- and and below. So we don't diagnose and

29
00:01:00.170 --> 00:01:02.570
- we don't prescribe, but, but we do everything

30
00:01:02.570 --> 00:01:03.070
- else.

31
00:01:03.530 --> 00:01:04.030
- Okay.

32
00:01:04.329 --> 00:01:06.650
- And we are sitting down to discuss this

33
00:01:06.650 --> 00:01:09.450
- at a pretty interesting time. The conversation about

34
00:01:09.450 --> 00:01:10.189
- AI safety

35
00:01:10.704 --> 00:01:12.724
- right now is is so loud,

36
00:01:13.265 --> 00:01:15.424
- and a lot of this discussion and these

37
00:01:15.424 --> 00:01:15.924
- headlines

38
00:01:16.224 --> 00:01:18.644
- and consternation, it's happening in altitude

39
00:01:19.744 --> 00:01:21.444
- among Frontier Labs, policymakers

40
00:01:22.799 --> 00:01:25.280
- that are somewhat removed from the realities of

41
00:01:25.280 --> 00:01:26.340
- delivering care.

42
00:01:26.799 --> 00:01:29.939
- You've seen leaders from OpenAI and Anthropic specifically

43
00:01:30.079 --> 00:01:31.619
- call for shared regulation

44
00:01:31.920 --> 00:01:33.619
- to ensure greater safety.

45
00:01:34.640 --> 00:01:36.659
- You, on the other hand, have built Hippocratic

46
00:01:37.064 --> 00:01:39.944
- specifically around deploying AI safely in health care.

47
00:01:39.944 --> 00:01:42.024
- This has been part of your your mission,

48
00:01:42.024 --> 00:01:44.825
- your charter since day one. So maybe to

49
00:01:44.825 --> 00:01:46.825
- to start us off, how is the work

50
00:01:46.825 --> 00:01:48.765
- you and your partners are doing with clinical

51
00:01:48.825 --> 00:01:49.564
- AI agents

52
00:01:50.040 --> 00:01:52.680
- just fundamentally different from what we're hearing out

53
00:01:52.680 --> 00:01:54.140
- of the Frontier Labs lately?

54
00:01:55.240 --> 00:01:56.760
- Yeah. You know, we took a very different

55
00:01:56.760 --> 00:01:57.260
- approach.

56
00:01:57.640 --> 00:01:59.240
- I think when we started the company, you

57
00:01:59.240 --> 00:02:00.920
- know, we named it Hippocratic. We made the

58
00:02:00.920 --> 00:02:03.105
- tagline do no harm. And we said... I

59
00:02:03.265 --> 00:02:05.185
- actually said to the investors, I said, I'm

60
00:02:05.185 --> 00:02:07.905
- not going to ship until it's safer than

61
00:02:07.905 --> 00:02:08.485
- a human.

62
00:02:09.425 --> 00:02:11.665
- And, they're like, great. When when is that?

63
00:02:11.665 --> 00:02:13.525
- And I was like, I'm... I don't know.

64
00:02:13.985 --> 00:02:15.745
- I'm like, and if you're not comfortable with

65
00:02:15.745 --> 00:02:17.580
- that, don't give me your money.

66
00:02:17.979 --> 00:02:18.780
- And I think, you know, today,

67
00:02:21.020 --> 00:02:22.620
- and and if you look at the Frontier

68
00:02:22.620 --> 00:02:23.900
- Labs, you know, they're like, well, we need

69
00:02:23.900 --> 00:02:25.259
- to all slow down and make it safe.

70
00:02:25.259 --> 00:02:27.020
- I'm like, you coulda just made it safe

71
00:02:27.020 --> 00:02:27.520
- yourself.

72
00:02:28.540 --> 00:02:30.219
- Like, that's what we did. We just didn't

73
00:02:30.219 --> 00:02:32.055
- ship. It took us eighteen months to actually

74
00:02:32.055 --> 00:02:34.215
- cross the threshold of being safe enough to

75
00:02:34.215 --> 00:02:37.674
- even begin our first patient facing calls. And,

76
00:02:38.854 --> 00:02:40.954
- it's just a a personal responsibility.

77
00:02:41.495 --> 00:02:43.194
- Like, I don't I don't think you need

78
00:02:43.495 --> 00:02:45.300
- to have some... You know, it's like it's

79
00:02:45.300 --> 00:02:47.560
- like peer pressure in in middle school. Like,

80
00:02:47.699 --> 00:02:49.219
- you don't need to do the right thing

81
00:02:49.219 --> 00:02:50.739
- only because if the other guy does the

82
00:02:50.739 --> 00:02:52.739
- right thing, you just need to do the

83
00:02:52.739 --> 00:02:55.379
- right thing. And so we approach this problem

84
00:02:55.379 --> 00:02:57.800
- very differently from a place of personal responsibility

85
00:02:58.659 --> 00:03:01.305
- and commitment because... Look. I mean, we're all

86
00:03:01.305 --> 00:03:03.625
- patients. We all have family members who are

87
00:03:03.625 --> 00:03:04.125
- patients.

88
00:03:04.745 --> 00:03:07.545
- Like, you you just wanna build something that

89
00:03:07.545 --> 00:03:10.284
- truly helps people and doesn't hurt them, and

90
00:03:10.825 --> 00:03:12.504
- and the only way to ensure that is

91
00:03:12.504 --> 00:03:13.965
- is to take the care upfront.

92
00:03:15.219 --> 00:03:16.900
- Let's unpack that a little bit. When you

93
00:03:16.900 --> 00:03:17.400
- say

94
00:03:18.180 --> 00:03:20.419
- safe, let's make it more concrete. You mentioned

95
00:03:20.419 --> 00:03:22.900
- it took you eighteen months to to cross

96
00:03:22.900 --> 00:03:24.740
- a threshold into what you would deem as

97
00:03:24.740 --> 00:03:26.599
- a safe product or safe AI,

98
00:03:27.254 --> 00:03:29.495
- with the name Hippocratic and do no harm

99
00:03:29.495 --> 00:03:30.235
- in... Intact.

100
00:03:30.854 --> 00:03:32.935
- But when you say you're engineering for safety

101
00:03:32.935 --> 00:03:34.875
- back in that that time and still today,

102
00:03:35.094 --> 00:03:37.655
- what does that actually involve? Can you speak

103
00:03:37.655 --> 00:03:39.114
- to that in a bit more detail?

104
00:03:39.580 --> 00:03:41.439
- Yeah. You know, safety is a multilayer

105
00:03:41.740 --> 00:03:43.580
- thing. In fact, you know, Jensen said this

106
00:03:43.580 --> 00:03:45.500
- best when this whole debate was occurring, and

107
00:03:45.500 --> 00:03:46.159
- he said,

108
00:03:46.699 --> 00:03:47.919
- safety is about engineering.

109
00:03:48.780 --> 00:03:50.780
- If you want safety, you engineer safety. But

110
00:03:50.780 --> 00:03:52.400
- the way you do it is redundancy

111
00:03:53.405 --> 00:03:55.165
- in every way you can think of. So

112
00:03:55.165 --> 00:03:57.485
- we start first where every employee takes the

113
00:03:57.485 --> 00:03:57.985
- Hippocratic

114
00:03:58.365 --> 00:03:58.865
- Oath,

115
00:03:59.405 --> 00:04:01.645
- which is a version of the original one,

116
00:04:01.965 --> 00:04:04.205
- but one we made for for kind of

117
00:04:04.205 --> 00:04:04.944
- our company.

118
00:04:05.310 --> 00:04:06.750
- But it says, hey. Look. I'll I'll make

119
00:04:06.750 --> 00:04:08.509
- sure I put patients first. Why? Because you

120
00:04:08.509 --> 00:04:09.810
- want people in the right psychology.

121
00:04:11.229 --> 00:04:13.009
- Second, we decided that,

122
00:04:13.870 --> 00:04:15.709
- what we would do is limit our scope.

123
00:04:15.709 --> 00:04:17.310
- You know? People say they wanna be safe.

124
00:04:17.310 --> 00:04:18.990
- I'm like, look. The best measure of being

125
00:04:18.990 --> 00:04:19.730
- safe is

126
00:04:20.324 --> 00:04:22.324
- self editing. You know? And we said we're

127
00:04:22.324 --> 00:04:24.324
- not gonna diagnose. We're not gonna prescribe, but

128
00:04:24.324 --> 00:04:26.084
- we also went deeper. We said we're not

129
00:04:26.084 --> 00:04:28.084
- gonna do kids under two. We're not gonna

130
00:04:28.084 --> 00:04:30.425
- do high risk pregnancies. We're not gonna do

131
00:04:30.564 --> 00:04:33.125
- mental health. We're not gonna do hospice care.

132
00:04:33.125 --> 00:04:35.144
- Like, there's just areas that we knew

133
00:04:35.910 --> 00:04:37.189
- it just wasn't the right,

134
00:04:38.629 --> 00:04:40.870
- you know, area for AI to be used.

135
00:04:40.870 --> 00:04:43.189
- And then we said, alright. We're not gonna

136
00:04:43.189 --> 00:04:45.529
- build one LLM. We're gonna build 31 LLMs

137
00:04:45.750 --> 00:04:48.230
- and have 30 of the LMs double check

138
00:04:48.230 --> 00:04:49.990
- the main LLM all the time. In fact,

139
00:04:49.990 --> 00:04:51.995
- some of them double check another LM that

140
00:04:51.995 --> 00:04:53.294
- double checks the main LM.

141
00:04:54.154 --> 00:04:56.154
- And we said you need this redundancy of

142
00:04:56.154 --> 00:04:57.675
- systems. I mean, that's how you design an

143
00:04:57.675 --> 00:04:59.354
- airplane. You don't just have one way to

144
00:04:59.354 --> 00:05:01.514
- control the rudder. You know, you have fly

145
00:05:01.514 --> 00:05:03.754
- by wire. You also have a mechanical connection

146
00:05:03.754 --> 00:05:05.774
- in some planes. You have a wired connection.

147
00:05:05.995 --> 00:05:07.720
- Like, you just because if one goes down,

148
00:05:07.720 --> 00:05:09.080
- you wanna make sure the other two were

149
00:05:09.080 --> 00:05:11.399
- there. And so we took a similar approach.

150
00:05:11.399 --> 00:05:12.620
- Then we said, hey.

151
00:05:13.080 --> 00:05:15.319
- That's not enough. Let's make sure it'll transfer

152
00:05:15.319 --> 00:05:17.240
- the call to a human if needed if

153
00:05:17.240 --> 00:05:19.240
- you're having chest pains or shortness of breath.

154
00:05:19.240 --> 00:05:22.044
- And so we built an escalation engine that'll

155
00:05:22.044 --> 00:05:24.125
- actually transfer to humans. We've now done that

156
00:05:24.125 --> 00:05:25.264
- 38,000

157
00:05:26.365 --> 00:05:28.704
- times. So we've done, you know, 300,000,000

158
00:05:28.764 --> 00:05:30.144
- plus patient interactions,

159
00:05:31.485 --> 00:05:31.985
- but,

160
00:05:32.939 --> 00:05:35.580
- in those interactions, 31,000 times, we had to...

161
00:05:35.580 --> 00:05:37.180
- Or 38,000 times, we had to transfer it

162
00:05:37.180 --> 00:05:38.480
- to a human. And so,

163
00:05:39.500 --> 00:05:41.420
- and then we said, let's output test. Let's

164
00:05:41.420 --> 00:05:43.520
- hire 7,000 nurses to make,

165
00:05:44.300 --> 00:05:46.300
- test calls to the AI, and let's not

166
00:05:46.300 --> 00:05:47.199
- ship until,

167
00:05:48.165 --> 00:05:49.524
- and in those test calls, one out of

168
00:05:49.524 --> 00:05:50.884
- 50 times, they got a human on the

169
00:05:50.884 --> 00:05:52.165
- other line, and we said, alright. We're not

170
00:05:52.165 --> 00:05:53.865
- shipping till we're safe as the human.

171
00:05:55.845 --> 00:05:58.165
- And and then we monitor the first thousand

172
00:05:58.165 --> 00:06:00.745
- calls, every single call within an hour,

173
00:06:01.819 --> 00:06:03.419
- and call the patient back. If there's a

174
00:06:03.419 --> 00:06:05.259
- problem, we sample half a percent of the

175
00:06:05.259 --> 00:06:07.680
- calls after it's live. Like, we've just done

176
00:06:07.819 --> 00:06:09.120
- every last thing.

177
00:06:10.699 --> 00:06:12.560
- You know, this is belt, suspenders,

178
00:06:13.975 --> 00:06:17.274
- whatever comes after suspenders on your jacket. Like,

179
00:06:17.574 --> 00:06:19.415
- this is this is every last thing we

180
00:06:19.415 --> 00:06:21.355
- could think of to just ensure safety.

181
00:06:21.814 --> 00:06:24.535
- Yeah. It sounds like restraint and redundancy has

182
00:06:24.535 --> 00:06:25.035
- been

183
00:06:25.495 --> 00:06:28.180
- two really important things among others. But I

184
00:06:28.180 --> 00:06:29.699
- I guess just to turn and flip the

185
00:06:29.699 --> 00:06:31.220
- seats a little bit, Munjal, like, if you

186
00:06:31.220 --> 00:06:33.079
- were sitting in a CMIOs

187
00:06:33.459 --> 00:06:36.259
- or CIO seat, what are the questions you'd

188
00:06:36.259 --> 00:06:38.180
- want a vendor to answer before letting their

189
00:06:38.180 --> 00:06:40.040
- agents anywhere near patients?

190
00:06:41.300 --> 00:06:42.740
- Yeah. You know, I think that,

191
00:06:44.235 --> 00:06:47.274
- there's this trope out there that's constantly, like,

192
00:06:47.274 --> 00:06:48.654
- all the models are commodities.

193
00:06:49.595 --> 00:06:50.735
- I don't think so.

194
00:06:51.675 --> 00:06:53.915
- Like, there's a reason Harvey in legal is

195
00:06:53.915 --> 00:06:55.055
- doing really well

196
00:06:56.729 --> 00:07:00.009
- despite Claude legal launching because people want a

197
00:07:00.009 --> 00:07:02.490
- built for purpose system. Because when you get

198
00:07:02.490 --> 00:07:03.949
- a built for purpose system,

199
00:07:04.649 --> 00:07:07.069
- it closes the last two, 3%,

200
00:07:07.610 --> 00:07:10.089
- last five, 10%, which in safety is a

201
00:07:10.089 --> 00:07:10.750
- big deal.

202
00:07:11.305 --> 00:07:13.165
- Like, you can't tell your boss, hey. I

203
00:07:13.225 --> 00:07:14.504
- I bought this thing. It was half the

204
00:07:14.504 --> 00:07:16.665
- price, but I'm gonna kill 2% more people.

205
00:07:16.665 --> 00:07:18.285
- Like like, that's not okay.

206
00:07:18.904 --> 00:07:20.985
- It might be okay in, you know, taking

207
00:07:20.985 --> 00:07:22.745
- your order at Taco Bell with the voice

208
00:07:22.745 --> 00:07:24.745
- AI. Like, okay. 2% of the time, I

209
00:07:24.745 --> 00:07:27.069
- gave you three tacos instead of two. Okay.

210
00:07:27.069 --> 00:07:28.670
- Fine. Like, you know, it's not the end

211
00:07:28.670 --> 00:07:30.770
- of the world. But I think for us,

212
00:07:31.069 --> 00:07:33.330
- the first question they need to ask is,

213
00:07:33.710 --> 00:07:35.009
- you know, what's the architecture?

214
00:07:35.629 --> 00:07:37.970
- Is it a multimodel safety architecture?

215
00:07:38.270 --> 00:07:40.189
- How has it been tested? Did they do

216
00:07:40.189 --> 00:07:41.090
- output testing?

217
00:07:41.595 --> 00:07:43.375
- So some of the points I just made.

218
00:07:43.915 --> 00:07:45.435
- But one of the ones that I find

219
00:07:45.435 --> 00:07:47.214
- everybody is missing at the moment

220
00:07:48.154 --> 00:07:50.975
- is they've forgotten why this revolution happened.

221
00:07:51.835 --> 00:07:54.654
- It's called the large language model.

222
00:07:55.729 --> 00:07:58.129
- And yet, so many times today, people are

223
00:07:58.129 --> 00:08:00.050
- not asking how big are the models you're

224
00:08:00.050 --> 00:08:00.550
- using.

225
00:08:01.649 --> 00:08:02.149
- So,

226
00:08:02.930 --> 00:08:04.629
- you know, most people don't know this, but...

227
00:08:04.930 --> 00:08:07.410
- Or have forgotten, but we we measure models

228
00:08:07.410 --> 00:08:09.555
- by the number of parameters they have. The

229
00:08:09.555 --> 00:08:12.115
- number of parameters to use is quite literally

230
00:08:12.115 --> 00:08:14.355
- the number of connections between the neurons in

231
00:08:14.355 --> 00:08:15.334
- the neural network.

232
00:08:16.035 --> 00:08:18.194
- And just like the human brain, the more

233
00:08:18.194 --> 00:08:19.975
- neurons an animal has,

234
00:08:20.595 --> 00:08:22.754
- the more connections it has and the smarter

235
00:08:22.754 --> 00:08:24.055
- it is. Right? Meaning,

236
00:08:24.500 --> 00:08:26.340
- you know, your mouse brain is not as

237
00:08:26.340 --> 00:08:28.420
- smart as your dog brain, is not as

238
00:08:28.420 --> 00:08:29.879
- smart as your dolphin brain,

239
00:08:30.819 --> 00:08:33.700
- and obviously onto the chimpanzee and and humans.

240
00:08:33.700 --> 00:08:34.440
- And so

241
00:08:35.299 --> 00:08:37.379
- this is actually the breakthrough that happened where

242
00:08:37.379 --> 00:08:40.365
- we got the first models at 7,000,000,030

243
00:08:40.365 --> 00:08:43.165
- parameters that really created breakthroughs, and then CHAT

244
00:08:43.165 --> 00:08:44.225
- GPT came

245
00:08:44.684 --> 00:08:46.524
- and was a much higher number. But, you

246
00:08:46.524 --> 00:08:49.325
- know, today, the the fable you use is

247
00:08:49.325 --> 00:08:51.485
- 10... Is rumored to be. They haven't published

248
00:08:51.485 --> 00:08:53.345
- the numbers, which is why we lost track,

249
00:08:53.639 --> 00:08:55.580
- but it's rumored to be 10,000,000,000,000

250
00:08:55.639 --> 00:08:56.139
- parameters.

251
00:08:57.000 --> 00:08:58.759
- Now why do you care in health care

252
00:08:58.759 --> 00:09:00.920
- about the size of the brain that's being

253
00:09:00.920 --> 00:09:02.840
- used in the model? In fact, a lot

254
00:09:02.840 --> 00:09:04.519
- of companies are running around saying, we have

255
00:09:04.519 --> 00:09:06.279
- small language models. You know they're really better.

256
00:09:06.279 --> 00:09:07.799
- I'm like, no. They're not. The whole darn

257
00:09:07.799 --> 00:09:10.355
- breakthrough was large language models, guys.

258
00:09:10.815 --> 00:09:12.894
- And the reason is that we found the

259
00:09:12.894 --> 00:09:15.154
- large language models catch subtlety.

260
00:09:15.934 --> 00:09:17.774
- So if I ask you, you know, Molly,

261
00:09:17.774 --> 00:09:19.375
- are you short of breath? And you don't

262
00:09:19.375 --> 00:09:21.134
- say yes. You don't say no. You say,

263
00:09:21.134 --> 00:09:23.554
- feels like there's an elephant on my chest.

264
00:09:24.779 --> 00:09:27.019
- Our model better darn well say that's a

265
00:09:27.019 --> 00:09:27.519
- yes.

266
00:09:28.059 --> 00:09:29.519
- Shoot. You know? And,

267
00:09:30.460 --> 00:09:32.139
- what we found is the small models will

268
00:09:32.139 --> 00:09:34.059
- miss that. So there's a lot of subtly,

269
00:09:34.059 --> 00:09:35.120
- a lot of ambiguity,

270
00:09:36.299 --> 00:09:38.725
- that's critical in health care that you want.

271
00:09:38.725 --> 00:09:40.965
- And so you want this intelligence. In fact,

272
00:09:40.965 --> 00:09:41.865
- the new mythos,

273
00:09:42.404 --> 00:09:44.325
- one that's finding all those bugs, is rumored

274
00:09:44.325 --> 00:09:45.684
- to be 27,000,000,000,000

275
00:09:45.684 --> 00:09:46.184
- parameters.

276
00:09:46.965 --> 00:09:49.524
- And so, you know, at Hippocratic, we're about

277
00:09:49.524 --> 00:09:51.225
- now 5,100,000,000,000

278
00:09:51.285 --> 00:09:52.345
- parameters, which

279
00:09:53.070 --> 00:09:54.830
- I would use more except you have to

280
00:09:54.830 --> 00:09:56.430
- make it fast enough for voice, and so

281
00:09:56.430 --> 00:09:58.610
- that's why we use 30 in parallel. So,

282
00:09:59.230 --> 00:10:01.389
- is we we basically push the number of

283
00:10:01.389 --> 00:10:03.889
- parameters up by firing them in parallel because

284
00:10:04.190 --> 00:10:06.269
- the voice latency means you you can't take

285
00:10:06.269 --> 00:10:08.345
- too long. And Fable takes a long time.

286
00:10:08.345 --> 00:10:09.705
- Like, you put in something. The other day,

287
00:10:09.705 --> 00:10:10.924
- I put in something, and,

288
00:10:11.304 --> 00:10:12.504
- you know, I had enough time to go

289
00:10:12.504 --> 00:10:14.345
- get a cappuccino and come back before it

290
00:10:14.345 --> 00:10:16.424
- was done. Like, you can't use that in

291
00:10:16.424 --> 00:10:18.024
- voice, so you you have an inherent trade

292
00:10:18.024 --> 00:10:20.924
- off there. But the size of the model

293
00:10:21.269 --> 00:10:23.930
- matters because it's basically the amount of intelligence

294
00:10:23.990 --> 00:10:25.129
- you're bringing to bear.

295
00:10:25.670 --> 00:10:27.910
- Mhmm. You just touched on something, which is

296
00:10:27.910 --> 00:10:29.750
- that this work... This... The the work of

297
00:10:29.750 --> 00:10:30.250
- safety

298
00:10:30.710 --> 00:10:33.350
- comes intention to other goals to speed and

299
00:10:33.350 --> 00:10:35.370
- or... And urgency being one of them.

300
00:10:36.034 --> 00:10:39.394
- But Munjal, I mean, nearly every AI company

301
00:10:39.394 --> 00:10:41.334
- in health care right now, it... It's making

302
00:10:41.634 --> 00:10:43.954
- some version of of the safety claim. And

303
00:10:43.954 --> 00:10:44.934
- for a leader

304
00:10:45.394 --> 00:10:46.375
- trying to separate

305
00:10:46.914 --> 00:10:50.054
- real rigor, like you're discussing, from good marketing,

306
00:10:50.750 --> 00:10:53.230
- are there any sniff test or litmus test

307
00:10:53.230 --> 00:10:54.990
- that you would put forward as a way

308
00:10:54.990 --> 00:10:57.629
- of verifying what what you're... They're actually hearing

309
00:10:57.629 --> 00:10:58.450
- about safety?

310
00:11:00.509 --> 00:11:02.129
- The thing about safety is

311
00:11:03.164 --> 00:11:06.205
- it... It's only measurable backward looking at some

312
00:11:06.205 --> 00:11:07.664
- point. Right? So,

313
00:11:08.125 --> 00:11:10.445
- like, you are safe because you have been

314
00:11:10.445 --> 00:11:10.945
- safe.

315
00:11:11.725 --> 00:11:12.544
- And so,

316
00:11:13.325 --> 00:11:15.804
- you know, we've now done these 300,000,000 patient

317
00:11:15.804 --> 00:11:16.304
- interactions,

318
00:11:17.049 --> 00:11:19.549
- and we've not had one major safety incident.

319
00:11:20.410 --> 00:11:22.090
- So I'm like, you can bet on somebody

320
00:11:22.090 --> 00:11:23.389
- who maybe

321
00:11:23.769 --> 00:11:26.649
- hasn't proven that, but you don't. Like, why

322
00:11:26.649 --> 00:11:28.590
- take that chance if you don't have to?

323
00:11:31.129 --> 00:11:31.629
- Second,

324
00:11:32.334 --> 00:11:33.394
- I think that

325
00:11:33.934 --> 00:11:35.634
- there's another element,

326
00:11:36.014 --> 00:11:37.774
- and you can get into the architecture, and

327
00:11:37.774 --> 00:11:39.455
- you can get into the model size, and

328
00:11:39.455 --> 00:11:41.154
- you can get into some of those properties.

329
00:11:41.214 --> 00:11:41.714
- But

330
00:11:42.014 --> 00:11:44.254
- in the end, to me, there's only two

331
00:11:44.254 --> 00:11:46.495
- things that really will help you. One, did

332
00:11:46.495 --> 00:11:47.875
- we do output testing?

333
00:11:48.250 --> 00:11:49.690
- Did we test what's coming out of the

334
00:11:49.690 --> 00:11:51.290
- model, not just say it's safe because of

335
00:11:51.290 --> 00:11:52.269
- what went in?

336
00:11:52.809 --> 00:11:54.490
- You know, drugs are done this way. They're

337
00:11:54.490 --> 00:11:55.930
- not... The FDA is not like, oh, yeah.

338
00:11:55.930 --> 00:11:57.930
- You you built that drug. You're just following

339
00:11:57.930 --> 00:11:59.930
- the process, so, therefore, we approve it. Like,

340
00:11:59.930 --> 00:12:01.290
- they're like, no. No. No. We're gonna test

341
00:12:01.290 --> 00:12:01.904
- the output.

342
00:12:02.305 --> 00:12:04.545
- Mhmm. So look for the output testing. Look

343
00:12:04.545 --> 00:12:07.024
- for the history of it. But, you know,

344
00:12:07.024 --> 00:12:08.785
- this is a new field, so there's a

345
00:12:08.785 --> 00:12:10.404
- ton of history for anybody.

346
00:12:12.785 --> 00:12:14.545
- And then I think that, you know, the

347
00:12:14.545 --> 00:12:15.365
- other element,

348
00:12:16.460 --> 00:12:18.779
- if you're a leader just trying to parse

349
00:12:18.779 --> 00:12:19.759
- through it all,

350
00:12:21.980 --> 00:12:24.320
- is, you know, have your own teams

351
00:12:25.019 --> 00:12:26.080
- go at it clinically

352
00:12:27.740 --> 00:12:29.519
- and look for corner cases

353
00:12:30.174 --> 00:12:31.455
- that are there. I mean, I'll I'll tell

354
00:12:31.455 --> 00:12:32.975
- you an example. We had a patient calling

355
00:12:32.975 --> 00:12:33.475
- in

356
00:12:34.174 --> 00:12:34.575
- on,

357
00:12:35.455 --> 00:12:37.455
- he was, calling in a schedule with a

358
00:12:37.455 --> 00:12:37.955
- neurologist.

359
00:12:39.054 --> 00:12:40.254
- He said, oh, I need an appointment with

360
00:12:40.254 --> 00:12:42.174
- a neurologist, and and the AI is like,

361
00:12:42.174 --> 00:12:43.615
- hey. Why why do you need an appointment?

362
00:12:43.615 --> 00:12:44.514
- And he's like,

363
00:12:45.559 --> 00:12:46.679
- or or what? Why do you need an

364
00:12:46.679 --> 00:12:47.879
- appointment, and what type of appointment do you

365
00:12:47.879 --> 00:12:49.240
- need? And he's like, oh, well, I was

366
00:12:49.240 --> 00:12:51.179
- struck by lightning yesterday at work.

367
00:12:52.519 --> 00:12:54.519
- Like, so you would think scheduling doesn't need

368
00:12:54.519 --> 00:12:57.240
- a clinically safe thing, but, actually, every now

369
00:12:57.240 --> 00:12:58.220
- and then, it does.

370
00:12:58.774 --> 00:13:00.375
- And you don't... You know? And that that

371
00:13:00.375 --> 00:13:01.975
- one's not so easy because it's not like

372
00:13:01.975 --> 00:13:04.075
- the answer is he should've called 911.

373
00:13:04.375 --> 00:13:06.715
- It's the next day. Like, he's he's okay.

374
00:13:07.495 --> 00:13:07.995
- Mhmm.

375
00:13:08.375 --> 00:13:09.975
- But at the same time, you better not

376
00:13:09.975 --> 00:13:11.975
- book him an appointment with a neurologist six

377
00:13:11.975 --> 00:13:12.955
- weeks from now.

378
00:13:13.409 --> 00:13:15.589
- Like, somebody's gotta squeeze this guy in tomorrow.

379
00:13:15.649 --> 00:13:16.549
- Right? So,

380
00:13:18.049 --> 00:13:20.309
- I think that there's a lot of different

381
00:13:20.370 --> 00:13:21.829
- elements, but in the end,

382
00:13:22.529 --> 00:13:24.769
- you really wanna build a safer system. You

383
00:13:24.769 --> 00:13:26.850
- really need to realize almost every call in

384
00:13:26.850 --> 00:13:27.669
- health care

385
00:13:28.725 --> 00:13:30.904
- can... Has a risk of being a clinical

386
00:13:30.964 --> 00:13:32.745
- call. So even when you think

387
00:13:33.524 --> 00:13:36.245
- it's a nonclinical use case, you're you're not

388
00:13:36.245 --> 00:13:38.245
- always right, and so you you want this

389
00:13:38.245 --> 00:13:40.325
- capability to be there at any time to

390
00:13:40.325 --> 00:13:41.065
- jump in.

391
00:13:41.769 --> 00:13:42.509
- Mhmm. Mhmm.

392
00:13:42.889 --> 00:13:45.289
- The stakes are just so much... They're already

393
00:13:45.289 --> 00:13:47.450
- higher in this... In in health care. And

394
00:13:47.450 --> 00:13:49.710
- then even there, there can still be, like,

395
00:13:50.089 --> 00:13:52.409
- being struck by lightning. My goodness. Like, what

396
00:13:52.409 --> 00:13:52.909
- a

397
00:13:53.289 --> 00:13:56.004
- cascading... The stakes are just even higher than

398
00:13:56.004 --> 00:13:57.144
- you could even imagine.

399
00:13:57.685 --> 00:13:59.205
- And and this is where you need the

400
00:13:59.205 --> 00:14:02.404
- judgment. Right? Because Right. You you think struck

401
00:14:02.404 --> 00:14:05.285
- by lightning means scheduled this way is anywhere

402
00:14:05.285 --> 00:14:06.024
- in the protocol?

403
00:14:06.899 --> 00:14:08.740
- Like like, there's no business rule that was,

404
00:14:08.740 --> 00:14:11.159
- if struck by lightning, then do this. Like,

405
00:14:11.379 --> 00:14:14.100
- you know, even the escalation protocols don't have

406
00:14:14.100 --> 00:14:15.959
- probably struck by lightning. So

407
00:14:16.339 --> 00:14:18.740
- the model just had to have judgment. And,

408
00:14:18.740 --> 00:14:20.659
- basically, the bigger the model, the better the

409
00:14:20.659 --> 00:14:21.159
- judgment.

410
00:14:22.875 --> 00:14:24.954
- Working off of of that story, I mean,

411
00:14:24.954 --> 00:14:26.174
- so much of the public,

412
00:14:27.674 --> 00:14:30.654
- narrative or commentary or attention

413
00:14:31.274 --> 00:14:33.834
- right now is really looking at advanced AI

414
00:14:33.834 --> 00:14:35.054
- as something to fear.

415
00:14:35.540 --> 00:14:37.779
- It sounds like from these stories and the

416
00:14:37.779 --> 00:14:39.960
- work that you've been putting forward at Hippocratic

417
00:14:40.100 --> 00:14:42.100
- AI, you and your team are just clearly

418
00:14:42.100 --> 00:14:44.120
- operating from a different premise altogether.

419
00:14:44.980 --> 00:14:46.899
- What would you say, Manjal, is the counter

420
00:14:46.899 --> 00:14:49.605
- case you'd make to that type of doomerism?

421
00:14:49.985 --> 00:14:51.745
- And maybe more to the point, what are

422
00:14:51.745 --> 00:14:54.245
- you seeing in live deployments that actually supports

423
00:14:54.945 --> 00:14:55.445
- optimism?

424
00:14:56.225 --> 00:14:56.725
- Yeah.

425
00:14:57.825 --> 00:14:59.205
- This thing drives me nuts.

426
00:15:00.225 --> 00:15:02.065
- You know, the... For... I'm I'm gonna start

427
00:15:02.065 --> 00:15:03.285
- at the highest level.

428
00:15:04.009 --> 00:15:06.009
- Why are we listening to doomers? These people

429
00:15:06.009 --> 00:15:07.309
- are owing 5,000

430
00:15:07.610 --> 00:15:09.450
- in the history of the world. They've never

431
00:15:09.450 --> 00:15:11.629
- been right because we're all here.

432
00:15:12.570 --> 00:15:14.669
- I'm like, oh, but this time for sure,

433
00:15:15.370 --> 00:15:17.610
- like, you know, it's it's... There's must be

434
00:15:17.610 --> 00:15:20.144
- something inherent about human beings beings that we

435
00:15:20.144 --> 00:15:22.841
- want to believe the doomer, framework because literally

436
00:15:22.841 --> 00:15:25.239
- these people have always been wrong. And I

437
00:15:25.239 --> 00:15:27.636
- I I thought about it deeply. I was

438
00:15:27.636 --> 00:15:30.034
- like, why are they always wrong? I mean,

439
00:15:30.034 --> 00:15:32.432
- as it's on the surface, I could see

440
00:15:32.432 --> 00:15:34.930
- why some of these arguments, you know, might

441
00:15:34.930 --> 00:15:36.769
- play out. It's... I actually think it's because

442
00:15:36.769 --> 00:15:39.590
- you can never predict the secondary and tertiary

443
00:15:39.889 --> 00:15:40.389
- implications,

444
00:15:41.250 --> 00:15:43.250
- only the first. So you're like like, let's

445
00:15:43.250 --> 00:15:46.070
- say you and I had owned farms on...

446
00:15:46.784 --> 00:15:48.784
- When the tractor came out in the industrial

447
00:15:48.784 --> 00:15:50.945
- revolution. People might have said, oh my god.

448
00:15:50.945 --> 00:15:52.725
- The farmer's gonna do so amazing.

449
00:15:53.345 --> 00:15:55.264
- You guys own all the good fertile land.

450
00:15:55.264 --> 00:15:57.345
- Like, you're gonna make a fortune now. You're

451
00:15:57.345 --> 00:15:59.184
- gonna make food for so much less, and

452
00:15:59.184 --> 00:16:01.044
- you're gonna make so much more of it.

453
00:16:01.250 --> 00:16:02.769
- Did the farmers end up super rich in

454
00:16:02.769 --> 00:16:05.970
- society from the tractor? No. Because they all

455
00:16:05.970 --> 00:16:08.049
- made a ton more food, and the cost

456
00:16:08.049 --> 00:16:09.029
- of food dropped.

457
00:16:09.490 --> 00:16:10.929
- And, actually, they ended up some of the

458
00:16:10.929 --> 00:16:13.909
- worst off. And so... But people could anticipate

459
00:16:14.375 --> 00:16:15.975
- kind of the first phase of it, but

460
00:16:15.975 --> 00:16:18.714
- not the second implication and the third implication.

461
00:16:18.774 --> 00:16:19.914
- And I think we keep

462
00:16:20.294 --> 00:16:22.875
- missing that. The second thing that I've seen

463
00:16:23.575 --> 00:16:26.235
- is we're wrong about the job loss thesis.

464
00:16:27.139 --> 00:16:28.419
- In fact, I'll I'll tell you one of

465
00:16:28.419 --> 00:16:30.440
- the key learnings that we've seen with Hippocratic

466
00:16:30.580 --> 00:16:33.559
- is we have this framework called Copilot, Autopilot,

467
00:16:33.700 --> 00:16:34.200
- Infinipilot.

468
00:16:34.580 --> 00:16:36.340
- And Copilot is things you do with the

469
00:16:36.340 --> 00:16:38.340
- human in the loop. Autopilot is things you

470
00:16:38.340 --> 00:16:40.200
- do today that we can now do cheaper,

471
00:16:40.259 --> 00:16:42.684
- can staff where you couldn't staff. And infinite

472
00:16:42.684 --> 00:16:44.284
- pilots, things you never thought to do till

473
00:16:44.284 --> 00:16:46.684
- you had an infinite supply of clinicians, you

474
00:16:46.684 --> 00:16:47.184
- know,

475
00:16:47.565 --> 00:16:49.504
- at at a very low cost. And

476
00:16:50.445 --> 00:16:52.684
- almost all of our... We knew Copilot was

477
00:16:52.684 --> 00:16:54.605
- gonna be hard because the human in the

478
00:16:54.605 --> 00:16:56.759
- loop, you only get 10% efficiency gain, and

479
00:16:56.759 --> 00:16:59.000
- it kind of evaporates on you. So we

480
00:16:59.000 --> 00:17:01.240
- focused on autopilot. Actually, every one of our

481
00:17:01.240 --> 00:17:03.240
- autopilot use cases has been a bit of

482
00:17:03.240 --> 00:17:03.820
- a struggle.

483
00:17:04.440 --> 00:17:07.079
- Like, we've gotten people there, but it's kinda

484
00:17:07.079 --> 00:17:09.160
- ground warfare. Like, we have had to fight

485
00:17:09.160 --> 00:17:11.259
- and fight to make the use cases work.

486
00:17:12.255 --> 00:17:14.894
- And every one of the infinite pilot, the

487
00:17:14.894 --> 00:17:16.654
- things you never thought to do that are

488
00:17:16.654 --> 00:17:18.515
- emergent have worked brilliantly.

489
00:17:19.295 --> 00:17:19.795
- Mhmm.

490
00:17:20.255 --> 00:17:20.755
- And

491
00:17:21.295 --> 00:17:23.375
- that's where I realized, okay, the job loss

492
00:17:23.375 --> 00:17:24.894
- thesis is false. So I'll give you one

493
00:17:24.894 --> 00:17:27.134
- example. You know, we recently did a loss

494
00:17:27.134 --> 00:17:27.795
- to follow-up,

495
00:17:29.480 --> 00:17:31.799
- for a large health system where we called

496
00:17:31.799 --> 00:17:34.039
- 1,700 patients that had not gone in for

497
00:17:34.039 --> 00:17:36.220
- their low dose CT scans for lung cancer.

498
00:17:37.399 --> 00:17:38.679
- And we didn't get all of them, but

499
00:17:38.679 --> 00:17:40.119
- they were gonna be at zero because they

500
00:17:40.119 --> 00:17:42.140
- gave up. That's what loss to follow-up means.

501
00:17:42.599 --> 00:17:43.000
- And,

502
00:17:43.714 --> 00:17:44.994
- they had tried to text. They had tried

503
00:17:44.994 --> 00:17:46.515
- an email. They had tried to letter. And

504
00:17:46.515 --> 00:17:48.835
- we got 02:50 to schedule. We got a

505
00:17:48.835 --> 00:17:51.255
- 180 to show up. We found nine significant

506
00:17:51.315 --> 00:17:53.335
- findings, three cancers, like,

507
00:17:54.515 --> 00:17:56.674
- you know, like... And one of them was

508
00:17:56.674 --> 00:17:58.349
- caught so early. It only needed a re

509
00:17:58.429 --> 00:18:00.429
- resection. It did not actually need a,

510
00:18:02.589 --> 00:18:04.529
- chemo or radiation. I mean,

511
00:18:05.630 --> 00:18:06.849
- you know, what a great

512
00:18:07.549 --> 00:18:08.049
- outcome

513
00:18:08.990 --> 00:18:11.069
- that we only could've done with the abundance

514
00:18:11.069 --> 00:18:12.769
- of of AI. And so

515
00:18:13.174 --> 00:18:15.414
- I think that the company should make great

516
00:18:15.414 --> 00:18:16.855
- products and not put them out till they're

517
00:18:16.855 --> 00:18:17.355
- safe.

518
00:18:17.815 --> 00:18:19.994
- They should also, by the way, take liability,

519
00:18:20.615 --> 00:18:21.835
- which we do at Hippocratic.

520
00:18:22.454 --> 00:18:24.474
- We're saying, look. We'll stand behind our products.

521
00:18:24.694 --> 00:18:26.634
- And I think any vendor that doesn't

522
00:18:27.539 --> 00:18:30.519
- is is a vendor one shouldn't use because

523
00:18:30.820 --> 00:18:33.700
- taking liability means I'll make my product better

524
00:18:33.700 --> 00:18:34.200
- because

525
00:18:34.820 --> 00:18:37.720
- I'm forced to financially as well as morally.

526
00:18:38.100 --> 00:18:38.600
- So,

527
00:18:39.265 --> 00:18:40.625
- anyways, those are those are some of the

528
00:18:40.625 --> 00:18:42.304
- things I think that, you know, we really

529
00:18:42.304 --> 00:18:44.304
- should look at the balance of payments because

530
00:18:44.304 --> 00:18:46.065
- we do that with every other technology in

531
00:18:46.065 --> 00:18:46.804
- health care.

532
00:18:47.184 --> 00:18:49.025
- Mhmm. I think it's a really good point

533
00:18:49.025 --> 00:18:51.924
- of our... Just human psychology and and biases

534
00:18:52.144 --> 00:18:53.765
- we might have to our, like, not

535
00:18:54.480 --> 00:18:57.440
- not anticipating those secondary and tertiary effects like

536
00:18:57.440 --> 00:18:58.980
- you had said in your opening.

537
00:18:59.359 --> 00:19:01.759
- The part of that response, Minjal, it's very

538
00:19:01.759 --> 00:19:03.839
- important to to consider. I love the farmer's

539
00:19:03.839 --> 00:19:04.339
- analogy.

540
00:19:05.359 --> 00:19:07.460
- As we head into 2027,

541
00:19:07.765 --> 00:19:09.605
- maybe we can wind down here and leave

542
00:19:09.605 --> 00:19:11.765
- our listeners with one thing that you'd want

543
00:19:11.765 --> 00:19:13.765
- the health care c suite to understand about

544
00:19:13.765 --> 00:19:15.384
- adopting AI safely.

545
00:19:15.845 --> 00:19:17.365
- And then the second part of that... My

546
00:19:17.365 --> 00:19:19.845
- question is what you think will separate leaders

547
00:19:19.845 --> 00:19:21.845
- who get this right from the ones who

548
00:19:21.845 --> 00:19:22.585
- who don't.

549
00:19:23.789 --> 00:19:25.710
- Yeah. So I think one part is,

550
00:19:26.750 --> 00:19:27.250
- don't

551
00:19:27.710 --> 00:19:29.950
- think there's such a thing as clinical and

552
00:19:29.950 --> 00:19:30.450
- nonclinical.

553
00:19:31.870 --> 00:19:33.170
- We call this omnidopic.

554
00:19:34.110 --> 00:19:35.789
- It says, hey. When a patient calls in

555
00:19:35.789 --> 00:19:37.710
- to schedule, if you have them on the

556
00:19:37.710 --> 00:19:38.210
- phone,

557
00:19:39.575 --> 00:19:40.875
- you better darn well

558
00:19:41.174 --> 00:19:41.914
- talk about

559
00:19:42.535 --> 00:19:45.174
- the discharge follow-up that they needed to do

560
00:19:45.174 --> 00:19:45.674
- today,

561
00:19:46.615 --> 00:19:48.555
- even if that's not why they called in.

562
00:19:48.615 --> 00:19:50.134
- And you can only do that if you

563
00:19:50.134 --> 00:19:52.170
- have a clinical and nonclinical

564
00:19:52.549 --> 00:19:55.190
- single platform, single framework. You should look for

565
00:19:55.190 --> 00:19:57.269
- that in the market and try to build

566
00:19:57.269 --> 00:19:59.190
- that because it'll take advantage of kind of

567
00:19:59.190 --> 00:19:59.849
- the opportunity,

568
00:20:00.630 --> 00:20:02.869
- much better. I think that the second element

569
00:20:02.869 --> 00:20:03.369
- is,

570
00:20:04.390 --> 00:20:05.049
- you know,

571
00:20:05.944 --> 00:20:06.444
- architect...

572
00:20:06.744 --> 00:20:08.904
- Look for things that are architected for safety

573
00:20:08.904 --> 00:20:09.884
- from day one.

574
00:20:11.144 --> 00:20:13.704
- Everybody thinks, oh my god. These frontier models

575
00:20:13.704 --> 00:20:16.265
- are amazing, and they do everything. I think

576
00:20:16.265 --> 00:20:19.545
- they've forgotten one very important insight we've gathered,

577
00:20:19.545 --> 00:20:20.285
- which is...

578
00:20:20.609 --> 00:20:22.950
- I call this the jagged intelligence insight.

579
00:20:23.490 --> 00:20:25.569
- These models can solve a 150 year old

580
00:20:25.569 --> 00:20:26.309
- math proofs.

581
00:20:26.690 --> 00:20:28.049
- But the other day, by the way, one

582
00:20:28.049 --> 00:20:29.990
- of these frontier models we were using,

583
00:20:31.170 --> 00:20:33.490
- we were testing out actually was, like, it

584
00:20:33.490 --> 00:20:35.670
- offered a woman a pap smear over telemed.

585
00:20:36.355 --> 00:20:37.875
- Said, ma'am, for your pap smear, would you

586
00:20:37.875 --> 00:20:39.095
- like telemed or in person?

587
00:20:39.875 --> 00:20:41.394
- And I'm like, pretty sure you can't do

588
00:20:41.394 --> 00:20:43.095
- a pap smear over telemed.

589
00:20:43.634 --> 00:20:44.134
- But,

590
00:20:45.554 --> 00:20:48.034
- these models are really smart and really dumb

591
00:20:48.034 --> 00:20:49.549
- at the same time. It's just we don't

592
00:20:49.549 --> 00:20:51.470
- realize the dumb things because we don't have

593
00:20:51.470 --> 00:20:54.049
- evals that tell us those those troughs.

594
00:20:54.589 --> 00:20:56.509
- And so you really wanna look for things

595
00:20:56.509 --> 00:20:56.910
- that,

596
00:20:57.470 --> 00:20:58.910
- that go... But I think you will see

597
00:20:58.910 --> 00:21:01.089
- the specialist companies in general.

598
00:21:02.164 --> 00:21:04.244
- Really, the question you're gonna wanna ask them

599
00:21:04.244 --> 00:21:05.944
- is what are all the troughs they fixed,

600
00:21:06.964 --> 00:21:09.125
- versus just, you know, what's your absolute score

601
00:21:09.125 --> 00:21:10.884
- on how you can solve, you know, the

602
00:21:10.884 --> 00:21:12.505
- next math proof. And so

603
00:21:12.884 --> 00:21:14.644
- I think that that's a... That's an element.

604
00:21:14.644 --> 00:21:16.025
- And then the last part is

605
00:21:17.140 --> 00:21:19.240
- patients are getting used to talking AI.

606
00:21:19.619 --> 00:21:21.539
- Mhmm. This question that you used to ask

607
00:21:21.539 --> 00:21:23.140
- about, oh my god. Are they really gonna

608
00:21:23.140 --> 00:21:23.880
- talk to it?

609
00:21:24.579 --> 00:21:26.259
- They're doing it all the time now. They're

610
00:21:26.259 --> 00:21:27.319
- talking to MetaMuse.

611
00:21:28.019 --> 00:21:29.539
- I mean, the other day, I I ordered

612
00:21:29.539 --> 00:21:31.460
- flowers for a family member on one of

613
00:21:31.460 --> 00:21:33.075
- the on on MetaMuse,

614
00:21:33.855 --> 00:21:34.355
- and,

615
00:21:35.054 --> 00:21:37.534
- they're talking to it in in Taco Bell

616
00:21:37.534 --> 00:21:39.294
- drive thrus. They're talking to it on their

617
00:21:39.294 --> 00:21:42.575
- bank. Like, they are very much getting used

618
00:21:42.575 --> 00:21:43.794
- to it. And so

619
00:21:44.255 --> 00:21:46.619
- you're soon gonna see in the next, I

620
00:21:46.619 --> 00:21:48.000
- bet, in '27,

621
00:21:48.619 --> 00:21:51.339
- every single consumer you'll start seeing will know

622
00:21:51.339 --> 00:21:53.180
- the difference, say, hey. This is not like

623
00:21:53.180 --> 00:21:55.740
- talking to IVR. It's way smarter. It really

624
00:21:55.740 --> 00:21:56.799
- answers my question.

625
00:21:57.420 --> 00:21:59.200
- It really completes my action,

626
00:22:00.384 --> 00:22:03.105
- and they'll just be comfortable. In fact, even

627
00:22:03.105 --> 00:22:05.505
- prefer it over many other forms of interaction,

628
00:22:05.505 --> 00:22:07.345
- and I think that folks need to get

629
00:22:07.345 --> 00:22:08.164
- ahead of that,

630
00:22:08.865 --> 00:22:10.545
- because the patients are doing it all day

631
00:22:10.545 --> 00:22:12.644
- long. I mean, what, 25%

632
00:22:13.359 --> 00:22:15.919
- of all searches on ChattGPT now are health

633
00:22:15.919 --> 00:22:17.519
- care? So people are like, are they gonna

634
00:22:17.519 --> 00:22:19.359
- really use AI in health care? I'm like,

635
00:22:19.359 --> 00:22:21.299
- they're already using it.

636
00:22:22.079 --> 00:22:24.000
- Right. And, by the way, there's almost a

637
00:22:24.000 --> 00:22:28.045
- billion weekly act of uniques. That means 250,000,000

638
00:22:28.265 --> 00:22:30.205
- people are doing health care searches

639
00:22:31.144 --> 00:22:34.105
- today. Like, what do you mean? Like, it's

640
00:22:34.105 --> 00:22:36.825
- already happening. And so I think they need

641
00:22:36.825 --> 00:22:38.365
- to get ready for this new world,

642
00:22:39.529 --> 00:22:41.929
- and think about what sort of... How does

643
00:22:41.929 --> 00:22:44.190
- that change access? How does that change follow-up?

644
00:22:44.809 --> 00:22:46.730
- And then the next phase is now already

645
00:22:46.730 --> 00:22:48.970
- coming, which is they're not even gonna call

646
00:22:48.970 --> 00:22:51.609
- you. Their agent, their MetaMuse is gonna call

647
00:22:51.609 --> 00:22:52.109
- you

648
00:22:52.684 --> 00:22:54.625
- and book the appointment and

649
00:22:54.924 --> 00:22:56.144
- follow-up on things.

650
00:22:57.325 --> 00:22:59.664
- And are you ready to have an interface

651
00:22:59.804 --> 00:23:01.744
- where you can talk to their agent,

652
00:23:02.204 --> 00:23:04.144
- and their agent will talk to the patient?

653
00:23:05.740 --> 00:23:07.339
- I love this. So for our listeners, it's

654
00:23:07.339 --> 00:23:10.160
- just twenty twenty seven omnitopics, so not distinguishing

655
00:23:10.380 --> 00:23:13.039
- or not separating the clinical versus nonclinical.

656
00:23:13.819 --> 00:23:15.740
- I love that challenge for those interactions. If

657
00:23:15.740 --> 00:23:17.279
- you have a patient on the line,

658
00:23:17.654 --> 00:23:20.375
- everything in... Is somewhat... Is is in clinical

659
00:23:20.375 --> 00:23:20.955
- in nature.

660
00:23:21.654 --> 00:23:23.755
- And then look for things that are built

661
00:23:23.815 --> 00:23:24.555
- with safety

662
00:23:25.174 --> 00:23:27.434
- day one. Beware of that jagged intelligence.

663
00:23:28.055 --> 00:23:29.819
- AI can be so brilliant and so

664
00:23:30.299 --> 00:23:32.539
- so dumb. Like you remind us with that

665
00:23:32.539 --> 00:23:35.419
- PAP smear anecdote, Mujal. And then patients getting

666
00:23:35.419 --> 00:23:37.339
- used to talking to AI. Sometimes it feels

667
00:23:37.339 --> 00:23:39.179
- like there's a lag to how we're still

668
00:23:39.179 --> 00:23:41.835
- seeing AI and moving into this industry. But

669
00:23:41.994 --> 00:23:43.534
- to your point, consumers,

670
00:23:43.835 --> 00:23:45.914
- patients, they're not only using it and talking

671
00:23:45.914 --> 00:23:47.934
- to it, but sometimes at length,

672
00:23:48.554 --> 00:23:51.034
- much much longer and with more detail than

673
00:23:51.034 --> 00:23:52.015
- you might assume.

674
00:23:52.794 --> 00:23:54.420
- Munjal, I just wanna thank you for your

675
00:23:54.420 --> 00:23:56.900
- time and your insights today. Anything you wanted

676
00:23:56.900 --> 00:23:58.279
- to add in closing here?

677
00:23:59.700 --> 00:24:01.720
- No. I, you know, I I think that...

678
00:24:02.660 --> 00:24:04.900
- Look. These things are really great helps in

679
00:24:04.900 --> 00:24:06.580
- health care. You know, I had a family

680
00:24:06.580 --> 00:24:09.619
- member diagnosed with stage four cancer actually this

681
00:24:09.619 --> 00:24:11.315
- week, and it was awful.

682
00:24:12.095 --> 00:24:13.934
- But I will tell you the large language

683
00:24:13.934 --> 00:24:17.075
- models are... Have real... I mean, in one...

684
00:24:17.294 --> 00:24:19.294
- In literally the last seventy two hours, I

685
00:24:19.294 --> 00:24:20.355
- was able to

686
00:24:21.054 --> 00:24:23.509
- look at every clinical trial, look at, you

687
00:24:23.509 --> 00:24:26.409
- know, every treatment plan, understand what it means,

688
00:24:26.629 --> 00:24:28.809
- and prepare for my meeting with the oncologist,

689
00:24:30.309 --> 00:24:31.690
- later on today, actually,

690
00:24:33.029 --> 00:24:34.730
- in a way that I could have never.

691
00:24:35.269 --> 00:24:36.950
- Like, you know, I could have never... It

692
00:24:36.950 --> 00:24:39.125
- would have taken me a month of research

693
00:24:39.125 --> 00:24:40.345
- to get what I got.

694
00:24:41.365 --> 00:24:41.605
- And,

695
00:24:42.724 --> 00:24:45.204
- I truly... And even... It prepared me a

696
00:24:45.204 --> 00:24:46.404
- 71

697
00:24:46.404 --> 00:24:47.464
- bulleted checklist

698
00:24:48.085 --> 00:24:48.904
- to to,

699
00:24:49.764 --> 00:24:51.684
- go through and make sure... And even took

700
00:24:51.684 --> 00:24:53.284
- the 10 that I gotta make sure I

701
00:24:53.284 --> 00:24:54.919
- ask the oncologist today.

702
00:24:55.539 --> 00:24:58.359
- Like, it's really helped me and my family

703
00:24:58.419 --> 00:25:00.759
- zero in on what's needed. And,

704
00:25:02.419 --> 00:25:03.240
- this is,

705
00:25:03.779 --> 00:25:05.539
- you know, this is a gift we never

706
00:25:05.539 --> 00:25:07.299
- had before, and I I think the whole

707
00:25:07.299 --> 00:25:09.285
- world already kind of is starting to get

708
00:25:09.285 --> 00:25:11.225
- that if they've been through a health crisis

709
00:25:11.605 --> 00:25:13.225
- since these LMS came out.

710
00:25:13.684 --> 00:25:14.184
- Mhmm.

711
00:25:16.164 --> 00:25:18.484
- In response to some of the doomerism is,

712
00:25:18.484 --> 00:25:20.965
- unfortunately, when health care hits close to home

713
00:25:20.965 --> 00:25:22.244
- like it has for you and your family,

714
00:25:22.244 --> 00:25:23.625
- which I'm sorry to hear,

715
00:25:23.980 --> 00:25:25.980
- Munjal, about your family member, and I I

716
00:25:25.980 --> 00:25:27.519
- wish them and you the best.

717
00:25:28.059 --> 00:25:29.440
- But nothing can quite counterdumerism,

718
00:25:29.899 --> 00:25:32.779
- like, being reminded of what's capable now, with

719
00:25:32.779 --> 00:25:35.099
- the technology like this at the fingertips of

720
00:25:35.099 --> 00:25:35.679
- of consumerisms

721
00:25:35.980 --> 00:25:37.839
- and patients and families and organizations.

722
00:25:38.625 --> 00:25:40.164
- So thanks for sharing that.

723
00:25:40.865 --> 00:25:43.184
- Listeners, you can tune into more podcasts from

724
00:25:43.184 --> 00:25:45.984
- Becker's Healthcare by visiting our podcast page at

725
00:25:45.984 --> 00:25:48.085
- beckershospitalreview.com.

726
00:25:48.473 --> 00:25:51.033
- Munjal, thank you again. Always a pleasure hearing

727
00:25:51.033 --> 00:25:52.553
- from you and learning more about the work

728
00:25:52.553 --> 00:25:54.633
- you and Hippocratic AI are building and doing

729
00:25:54.633 --> 00:25:55.373
- with organizations.

730
00:25:56.393 --> 00:25:57.853
- Thank you for having me, Molly.