WEBVTT 00:00:00.000 --> 00:00:11.420 On today's AI briefing, if the reported figures hold, we think this is a real platform financed ahead of its customers, with a loss roughly five times revenue. 00:00:13.120 --> 00:00:18.200 This is the AI briefing for Saturday, October 31. I'm Michael. 00:00:18.720 --> 00:00:23.460 Like everything here, this briefing is written and read by AI from public sources. 00:00:24.200 --> 00:00:26.860 Today is about money and who has to borrow it. 00:00:26.860 --> 00:00:31.580 Anthropic has filed to go public and the reported numbers are striking 00:00:31.580 --> 00:00:35.820 16.3 billion dollars of revenue in the first half 00:00:35.820 --> 00:00:39.520 against an eighty point seven billion dollar net loss 00:00:39.520 --> 00:00:44.580 open AI has reportedly lined up about 11 billion dollars 00:00:44.580 --> 00:00:50.380 well short of a 30 billion dollar target lenders are reportedly shutting neoclouds 00:00:50.380 --> 00:00:53.460 like core weave and nscale out of the bond market 00:00:53.460 --> 00:00:58.360 Nvidia is weighing whether to cost in a customer's loan and the biggest tech 00:00:58.360 --> 00:01:02.440 companies are spending so heavily on capex that three of them have gone into 00:01:02.440 --> 00:01:07.640 negative free cash flow will take the main stories first then a few quick 00:01:07.640 --> 00:01:09.740 hits 00:01:09.740 --> 00:01:14.540 Stevens take on the top of the day neocloud implosion could really be a 00:01:14.540 --> 00:01:16.780 disaster 00:01:16.780 --> 00:01:22.140 first story and it's an update on anthropic the company has now filed its 00:01:22.140 --> 00:01:28.300 S1, the document a company files before listing its shares. Shaquille Hashim at Transformer 00:01:28.300 --> 00:01:35.740 reports $16.3 billion of revenue in the first half of this year, up 14 times. The reported net 00:01:35.740 --> 00:01:45.060 loss was $80.7 billion. Ed Zitron says the filing shows $430 billion of non-cancellable commitments. 00:01:45.060 --> 00:01:54.700 For Gary Marcus, a $2 trillion valuation needs over $500 billion of revenue, and his target year is 2031. 00:01:55.480 --> 00:01:59.860 Strut Ekuri says the core business is profitable, but shows no figures. 00:02:00.540 --> 00:02:05.600 One caution, none of these numbers has yet been traced to the filing or confirmed elsewhere. 00:02:06.280 --> 00:02:13.260 If they hold, we see a real platform financed ahead of its customers, losing about five times its revenue. 00:02:13.260 --> 00:02:21.420 What we'd read first is how that loss splits between recurring cash compute and non-cash or one-time items. 00:02:23.080 --> 00:02:29.720 If Anthropic shows the scale of the losses, our second story shows how hard the money may be getting to raise. 00:02:30.360 --> 00:02:37.120 OpenAI has reportedly secured about $11 billion in commitments, according to reports gathered by TechMeme. 00:02:37.120 --> 00:02:42.780 The original target was $30 billion at a valuation of $1.4 trillion. 00:02:43.840 --> 00:02:46.220 OpenAI is still trying to raise more. 00:02:46.820 --> 00:02:52.840 Ed Zitron puts the haul at $11 billion with a valuation at $1.25 trillion. 00:02:54.120 --> 00:02:57.000 His word for investor interest was tepid. 00:02:57.500 --> 00:02:58.980 How solid is any of this? 00:02:59.320 --> 00:03:00.220 Not very yet. 00:03:00.760 --> 00:03:03.660 The tech meme items are headlines without much context. 00:03:04.160 --> 00:03:06.820 Zitron's post gives no sourcing for its figures. 00:03:07.120 --> 00:03:13.120 and that description of demand isn't attributed to anyone so this is reported not confirmed with 00:03:13.120 --> 00:03:19.920 that caveat here's what we think a reported 11 billion dollars against a 30 billion dollar target 00:03:19.920 --> 00:03:27.760 is the clearest sign yet that frontier lab capital has a ceiling that holds even at a lower valuation 00:03:27.760 --> 00:03:34.000 the round is still open though so the final number could move for early stage investors the exposure 00:03:34.000 --> 00:03:40.960 sits one step removed if the round closes near 11 billion dollars the most exposed are secondaries 00:03:40.960 --> 00:03:46.880 meaning sales of existing shares and late rounds in lab adjacent companies those are the deals 00:03:46.880 --> 00:03:53.440 priced off frontier lab marks and they look most at risk of repricing a question worth asking might 00:03:53.440 --> 00:04:00.160 be how much a company's last price leaned on those marks the market view reads the same facts more 00:04:00.160 --> 00:04:07.360 calmly there is 11 billion dollar committed and the round is still open the valuation was trimmed 00:04:07.360 --> 00:04:15.760 from 1.4 to 1.25 trillion dollars on that reading this is price discovery rather than a failed raise 00:04:17.920 --> 00:04:24.400 from the labs to the companies that rent them gps third story and it's an update on the neoclouds 00:04:24.400 --> 00:04:29.920 the wall street journal reports that core weave nebius and nscale have been shut out of the bond 00:04:29.920 --> 00:04:36.400 market with lenders demanding tougher terms tech crunch adds that Nvidia is considering costing 00:04:36.400 --> 00:04:42.540 a loan for nscales West Virginia data center the bond market exclusion rests on the journals 00:04:42.540 --> 00:04:49.840 reporting the Nvidia part is unconfirmed and speculative we think this is the first concrete 00:04:49.840 --> 00:04:56.200 repricing of the build-outs debt landing on the company's most reliant on borrowed GPS and if it 00:04:56.200 --> 00:05:02.060 happens a chip vendor costing its customers loan is circular financing for 00:05:02.060 --> 00:05:06.100 startups renting from these three a squeeze on their supplier could arrive 00:05:06.100 --> 00:05:11.620 as price rises or contract changes that makes capacity spread across several 00:05:11.620 --> 00:05:17.640 providers worth more the fourth and last of the main stories moves up to the 00:05:17.640 --> 00:05:22.880 biggest balance sheets of all tech crunch drawing on an S&P global report 00:05:22.880 --> 00:05:27.480 says Alphabet, Amazon, and Meta are now in negative free cash flow. 00:05:28.140 --> 00:05:32.740 Each is guiding to CapEx in the range of $145 to $220 billion. 00:05:33.680 --> 00:05:37.540 Microsoft stayed positive at $18.6 billion. 00:05:38.520 --> 00:05:43.900 The Big Five have raised $326 billion in debt in 2026. 00:05:44.640 --> 00:05:46.800 That is triple last year's amount. 00:05:47.420 --> 00:05:51.140 Tech meme adds that S and P downgraded Meta's outlook to negative, 00:05:51.140 --> 00:05:57.320 citing the heavy capex and negative free cash flow. Meta's stock fell 15% in October. 00:05:57.920 --> 00:06:01.300 On the evidence, this one stands on firmer ground than the rest of today. 00:06:01.860 --> 00:06:07.180 The debt and cash flow trends come from S&P's own data, and the Meta downgrade is confirmed. 00:06:07.880 --> 00:06:13.140 One claim attached to it about AI mode cannibalization remains unverified. 00:06:13.580 --> 00:06:15.780 We'll come back to AI mode in a moment. 00:06:15.780 --> 00:06:20.160 Our read is that negative free cash flow at Alphabet. 00:06:20.740 --> 00:06:25.060 Amazon and Meta means CapEx is outrunning even the strongest balance sheets. 00:06:25.760 --> 00:06:26.980 Debt is filling the gap. 00:06:27.480 --> 00:06:30.060 That echoes the Neocloud story, one level up. 00:06:30.440 --> 00:06:35.500 But Microsoft's staying positive at $18.6 billion cuts against this. 00:06:35.880 --> 00:06:38.860 It shows the spend is a choice, not a necessity. 00:06:39.400 --> 00:06:42.560 The market view leans on that same Microsoft number. 00:06:42.980 --> 00:06:44.740 It shows the spend can be funded. 00:06:45.780 --> 00:06:54.820 From that angle, large, well-capitalized companies borrowing to build capacity is ordinary capital allocation, not distress. 00:06:56.460 --> 00:06:58.600 Two quick ones before the close. 00:06:59.120 --> 00:07:00.980 First, a follow-up on Google. 00:07:01.480 --> 00:07:09.560 Chief Business Officer Philip Skindler said AI mode query growth had an impact because of the lower monetization we were seeing in AI mode. 00:07:10.260 --> 00:07:15.380 Headlines tie that to a 15% move worth $500 billion, which overstates it. 00:07:15.740 --> 00:07:16.480 That was the market. 00:07:17.120 --> 00:07:17.840 Why it matters. 00:07:18.400 --> 00:07:21.560 Google itself says AI answers monetize worse. 00:07:22.160 --> 00:07:26.000 And Bassett is raising about $2 billion in two tranches. 00:07:26.900 --> 00:07:31.740 One values it at $26 billion, the other at $30 billion. 00:07:32.640 --> 00:07:35.440 In June, it was valued at $13 billion. 00:07:36.500 --> 00:07:37.340 Why it matters. 00:07:37.340 --> 00:07:44.680 equity is still pricing AI companies sharply higher even as debt tightens 00:07:44.680 --> 00:07:49.880 so a question worth asking in diligence this week how exposed are startups 00:07:49.880 --> 00:07:55.080 compute supply and pricing to a neocloud that cannot refinance and one to watch 00:07:55.080 --> 00:08:01.460 what makes up Anthropix reported 80.7 billion dollar loss that decides whether 00:08:01.460 --> 00:08:07.100 frontier lab losses shrink with scale or grow with it the s1 income statement 00:08:07.100 --> 00:08:10.620 its footnotes and any amended filing should settle it. 00:08:12.380 --> 00:08:14.920 That's the briefing. Stay with me. 00:08:15.420 --> 00:08:18.560 Right after this, Bella takes you through today's deep dive 00:08:18.560 --> 00:08:21.300 on when the chatbot plays doctor who pays. 00:08:21.800 --> 00:08:26.100 You can also read it or listen to it on its own on the AI Briefing webpage, 00:08:26.600 --> 00:08:29.800 where you'll find the written edition in every past episode. 00:08:30.260 --> 00:08:32.000 The link is in the show notes. 00:08:35.060 --> 00:08:36.040 Thanks, Michael. 00:08:36.600 --> 00:08:41.040 This is the AI briefing weekly deep dive for Saturday, October 31, 00:08:41.580 --> 00:08:43.860 when the chatbot plays doctor, who pays. 00:08:44.380 --> 00:08:48.440 I'm Bella, and with me is Michael, whose job this week is to push back. 00:08:49.000 --> 00:08:53.080 Like the briefing, it is written and read by AI from public sources. 00:08:54.920 --> 00:08:57.040 It matters because of where it lands. 00:08:57.180 --> 00:09:01.240 It joins a coordinated California proceeding that already bundles 00:09:01.240 --> 00:09:05.020 at least a dozen product liability cases over GPT-40, 00:09:05.020 --> 00:09:08.380 and it moves the theory from mental health harm into medicine, 00:09:08.800 --> 00:09:12.720 where licensing laws, drug labels, and a standard of care exist. 00:09:13.420 --> 00:09:17.420 So the question is who carries liability when a model gives medical advice? 00:09:17.920 --> 00:09:20.380 The provider, the app builder, or the user? 00:09:22.260 --> 00:09:27.800 So why is this the week's topic, rather than one more tragic lawsuit on a growing pile? 00:09:27.800 --> 00:09:36.460 OpenAI's own figure is that more than 230 million people ask chat GPT health and wellness questions every week. 00:09:37.040 --> 00:09:40.580 That's self-reported, but it says the exposure isn't niche. 00:09:42.240 --> 00:09:44.460 It matters because of where it lands. 00:09:45.040 --> 00:09:52.440 It joins a coordinated California proceeding that already bundles at least a dozen product liability cases over GPT-40, 00:09:52.440 --> 00:09:55.800 and it moves the theory from mental health harm into medicine, 00:09:56.220 --> 00:10:00.140 where licensing laws, drug labels, and a standard of care exist. 00:10:00.840 --> 00:10:04.840 So the question is who carries liability when a model gives medical advice? 00:10:05.340 --> 00:10:07.820 The provider, the app builder, or the user? 00:10:09.700 --> 00:10:15.220 So why is this the week's topic, rather than one more tragic lawsuit on a growing pile? 00:10:15.220 --> 00:10:23.880 OpenAI's own figure is that more than 230 million people ask chat GPT health and wellness questions every week. 00:10:24.220 --> 00:10:28.000 That's self-reported, but it says the exposure isn't niche. 00:10:29.680 --> 00:10:31.900 It matters because of where it lands. 00:10:32.540 --> 00:10:39.860 It joins a coordinated California proceeding that already bundles at least a dozen product liability cases over GPT-40, 00:10:39.860 --> 00:10:45.980 and it moves the theory from mental health harm into medicine where licensing laws drug labels 00:10:45.980 --> 00:10:51.740 and a standard of care exist so the question is who carries liability when a model gives medical 00:10:51.740 --> 00:11:00.440 advice the provider the app builder or the user so why is this the week's topic rather than one 00:11:00.440 --> 00:11:07.660 more tragic lawsuit on a growing pile open AI's own figure is that more than 230 million people 00:11:07.660 --> 00:11:13.120 ask chat GPT health and wellness questions every week that's self-reported 00:11:13.120 --> 00:11:20.020 but it says the exposure isn't niche it matters because of where it lands it 00:11:20.020 --> 00:11:24.280 joins a coordinated California proceeding that already bundles at least a dozen 00:11:24.280 --> 00:11:29.440 product liability cases over GPT for oh and it moves the theory from mental 00:11:29.440 --> 00:11:34.120 health harm into medicine where licensing laws drug labels and a standard 00:11:34.120 --> 00:11:39.700 of care exist. So the question is who carries liability when a model gives medical advice, 00:11:40.200 --> 00:11:47.220 the provider, the app builder, or the user. Picture a seed stage app for managing chronic 00:11:47.220 --> 00:11:54.420 medications built on a frontier model's API that produces a dangerous taper. OpenAI's business 00:11:54.420 --> 00:11:59.000 terms say the customer is responsible for all activities that occur under its account, 00:11:59.000 --> 00:12:02.480 including the activities of end users who come through its app. 00:12:03.220 --> 00:12:06.100 The customer also agrees to indemnify OpenAI, 00:12:06.700 --> 00:12:11.440 which means covering OpenAI's costs for third-party claims within that scope. 00:12:12.760 --> 00:12:17.880 Liability under the agreement is capped at what the customer paid in the prior 12 months. 00:12:19.440 --> 00:12:24.820 But that contract doesn't bind the injured user who can still sue the provider directly. 00:12:24.820 --> 00:12:35.380 In the character, AI case, a federal court reportedly let claims against Google go forward on the theory that it supplied a component, the underlying model. 00:12:35.880 --> 00:12:41.060 So the provider isn't immune. The terms mostly decide who pays whom afterwards. 00:12:41.600 --> 00:12:45.780 The user is the third party here, and that's where the record matters. 00:12:45.780 --> 00:12:50.900 Picture a seed stage app for managing chronic medications 00:12:50.900 --> 00:12:55.720 built on a frontier model's API that produces a dangerous taper. 00:12:56.540 --> 00:12:59.540 OpenAI's business terms say the customer is responsible 00:12:59.540 --> 00:13:02.060 for all activities that occur under its account, 00:13:02.520 --> 00:13:05.540 including the activities of end users who come through its app. 00:13:06.220 --> 00:13:09.160 The customer also agrees to indemnify OpenAI, 00:13:09.740 --> 00:13:11.920 which means covering OpenAI's costs 00:13:11.920 --> 00:13:14.500 for third-party claims within that scope. 00:13:15.780 --> 00:13:20.940 Liability under the agreement is capped at what the customer paid in the prior 12 months. 00:13:22.540 --> 00:13:27.860 But that contract doesn't bind the injured user, who can still sue the provider directly. 00:13:28.500 --> 00:13:38.440 In the Character, AI case, a federal court reportedly let claims against Google go forward on the theory that it supplied a component, the underlying model. 00:13:38.920 --> 00:13:40.680 So the provider isn't immune. 00:13:41.160 --> 00:13:44.120 The terms mostly decide who pays whom afterwards. 00:13:44.120 --> 00:13:48.720 The user is the third party here, and that's where the record matters. 00:13:50.520 --> 00:13:58.760 Picture a seed stage app for managing chronic medications, built on a Frontier Model's API, that produces a dangerous taper. 00:13:59.580 --> 00:14:08.580 OpenAI's business terms say the customer is responsible for all activities that occur under its account, including the activities of end users who come through its app. 00:14:08.580 --> 00:14:17.540 The customer also agrees to indemnify OpenAI, which means covering OpenAI's costs for third-party claims within that scope. 00:14:18.680 --> 00:14:23.980 Liability under the agreement is capped at what the customer paid in the prior 12 months. 00:14:25.500 --> 00:14:30.920 But that contract doesn't bind the injured user who can still sue the provider directly. 00:14:30.920 --> 00:14:41.480 In the character, AI case, a federal court reportedly let claims against Google go forward on the theory that it supplied a component, the underlying model. 00:14:41.980 --> 00:14:43.680 So the provider isn't immune. 00:14:44.220 --> 00:14:47.160 The terms mostly decide who pays whom afterwards. 00:14:47.700 --> 00:14:51.880 The user is the third party here, and that's where the record matters. 00:14:51.880 --> 00:14:57.000 Picture a seed stage app for managing chronic medications 00:14:57.000 --> 00:15:01.820 built on a frontier model's API that produces a dangerous taper. 00:15:02.640 --> 00:15:05.640 OpenAI's business terms say the customer is responsible 00:15:05.640 --> 00:15:08.160 for all activities that occur under its account, 00:15:08.620 --> 00:15:11.640 including the activities of end users who come through its app. 00:15:12.320 --> 00:15:15.260 The customer also agrees to indemnify OpenAI, 00:15:15.840 --> 00:15:18.020 which means covering OpenAI's costs 00:15:18.020 --> 00:15:20.600 for third-party claims within that scope. 00:15:21.880 --> 00:15:27.040 Liability under the agreement is capped at what the customer paid in the prior 12 months. 00:15:28.640 --> 00:15:33.960 But that contract doesn't bind the injured user, who can still sue the provider directly. 00:15:34.600 --> 00:15:44.540 In the Character, AI case, a federal court reportedly let claims against Google go forward on the theory that it supplied a component, the underlying model. 00:15:44.700 --> 00:15:50.220 So the provider isn't immune. The terms mostly decide who pays whom afterwards. 00:15:50.220 --> 00:15:54.820 The user is the third party here, and that's where the record matters. 00:15:56.580 --> 00:16:09.420 In October 2025, OpenAI clarified its usage policies to bar tailored advice that requires a license, such as legal or medical advice, without appropriate involvement by a licensed professional. 00:16:10.260 --> 00:16:13.240 It then added that the model behavior has also not changed. 00:16:13.940 --> 00:16:18.700 For OpenAI, the policy supports a misuse defense like the one it used in RAIN, 00:16:19.100 --> 00:16:22.340 where it cited misuse, unauthorized use of the product. 00:16:23.140 --> 00:16:26.920 For plaintiffs, it's a written rule the product didn't enforce, 00:16:26.920 --> 00:16:29.500 which reads like the company knew the risk. 00:16:31.160 --> 00:16:34.080 On precedent, the honest scorecard is thin. 00:16:34.720 --> 00:16:39.880 In Garcia vs. Character Technologies, a Florida federal court treated the app as 00:16:39.880 --> 00:16:44.620 a product for the purposes of plaintiff's claims arising from defects in the character, 00:16:45.260 --> 00:16:48.720 AI app rather than ideas or expressions within the app. 00:16:49.360 --> 00:16:55.600 That was only a refusal to dismiss, and the case settled in January on undisclosed terms. 00:16:56.880 --> 00:17:01.400 In Walters, a defamation case, OpenAI won summary judgment, 00:17:01.920 --> 00:17:06.240 so plaintiffs are getting past early motions but nobody has won at trial. 00:17:06.240 --> 00:17:14.620 in october 2025 open ai clarified its usage policies to bar tailored advice that requires 00:17:14.620 --> 00:17:20.800 a license such as legal or medical advice without appropriate involvement by a licensed professional 00:17:20.800 --> 00:17:27.540 it then added that the model behavior has also not changed for open ai the policy supports a 00:17:27.540 --> 00:17:33.680 misuse defense like the one it used in rain where it cited misuse unauthorized use of the product 00:17:33.680 --> 00:17:40.780 For plaintiffs, it's a written rule the product didn't enforce, which reads like the company knew the risk. 00:17:42.420 --> 00:17:45.380 On precedent, the honest scorecard is thin. 00:17:46.020 --> 00:17:55.900 In Garcia vs. Character Technologies, a Florida federal court treated the app as a product for the purposes of plaintiffs' claims arising from defects in the character, 00:17:56.540 --> 00:18:00.020 AI app rather than ideas or expressions within the app. 00:18:00.020 --> 00:18:06.900 That was only a refusal to dismiss, and the case settled in January on undisclosed terms. 00:18:08.240 --> 00:18:17.540 In Walters, a defamation case, OpenAI won summary judgment, so plaintiffs are getting past early motions but nobody has won at trial. 00:18:17.540 --> 00:18:25.920 in october 2025 open ai clarified its usage policies to bar tailored advice that requires 00:18:25.920 --> 00:18:32.100 a license such as legal or medical advice without appropriate involvement by a licensed professional 00:18:32.100 --> 00:18:38.840 it then added that the model behavior has also not changed for open ai the policy supports a 00:18:38.840 --> 00:18:44.980 misuse defense like the one it used in rain where it cited misuse unauthorized use of the product 00:18:44.980 --> 00:18:52.080 For plaintiffs, it's a written rule the product didn't enforce, which reads like the company knew the risk. 00:18:53.720 --> 00:18:56.680 On precedent, the honest scorecard is thin. 00:18:57.320 --> 00:19:07.220 In Garcia vs. Character Technologies, a Florida federal court treated the app as a product for the purposes of plaintiffs' claims arising from defects in the character, 00:19:07.840 --> 00:19:11.320 AI app rather than ideas or expressions within the app. 00:19:11.320 --> 00:19:18.200 That was only a refusal to dismiss, and the case settled in January on undisclosed terms. 00:19:19.540 --> 00:19:28.840 In Walters, a defamation case, OpenAI won summary judgment, so plaintiffs are getting past early motions but nobody has won at trial. 00:19:28.840 --> 00:19:37.220 in october 2025 open ai clarified its usage policies to bar tailored advice that requires 00:19:37.220 --> 00:19:43.400 a license such as legal or medical advice without appropriate involvement by a licensed professional 00:19:43.400 --> 00:19:50.140 it then added that the model behavior has also not changed for open ai the policy supports a 00:19:50.140 --> 00:19:56.280 misuse defense like the one it used in rain where it cited misuse unauthorized use of the product 00:19:56.280 --> 00:20:03.380 For plaintiffs, it's a written rule the product didn't enforce, which reads like the company knew the risk. 00:20:05.020 --> 00:20:07.980 On precedent, the honest scorecard is thin. 00:20:08.620 --> 00:20:18.500 In Garcia vs. Character Technologies, a Florida federal court treated the app as a product for the purposes of plaintiffs' claims arising from defects in the character, 00:20:19.140 --> 00:20:22.620 AI app rather than ideas or expressions within the app. 00:20:22.620 --> 00:20:29.500 That was only a refusal to dismiss, and the case settled in January on undisclosed terms. 00:20:30.840 --> 00:20:40.140 In Walters, a defamation case, OpenAI won summary judgment, so plaintiffs are getting past early motions but nobody has won at trial. 00:20:41.800 --> 00:20:45.160 That's a fair challenge, and the facts don't settle it. 00:20:45.160 --> 00:20:50.660 If courts did accept that framing, legislators wouldn't need A-specific medical rules. 00:20:50.660 --> 00:20:55.460 existing licensing law would already reach where every reading lands in the 00:20:55.460 --> 00:20:59.820 same place is the cost it flows to whoever is closest to the user and has 00:20:59.820 --> 00:21:06.380 the least bargaining power in most stacks that's the startup I'd want to 00:21:06.380 --> 00:21:11.000 know how insurers are responding but I haven't seen a source on that so it's a 00:21:11.000 --> 00:21:16.220 question not a fact the policy reading is where I'm most doubtful unlicensed 00:21:16.220 --> 00:21:21.640 practice statutes were written for people presenting themselves as doctors. OpenAI says 00:21:21.640 --> 00:21:27.520 ChatGPT has never been a substitute for professional legal or medical advice, and a court could 00:21:27.520 --> 00:21:31.440 easily find a disclaimed general tool isn't practicing anything. 00:21:33.220 --> 00:21:38.820 That's a fair challenge, and the facts don't settle it. If courts did accept that framing, 00:21:39.320 --> 00:21:44.820 legislators wouldn't need A-specific medical rules. Existing licensing law would already reach. 00:21:45.400 --> 00:21:48.100 Where every reading lands in the same place is the cost. 00:21:48.580 --> 00:21:52.540 It flows to whoever is closest to the user and has the least bargaining power. 00:21:53.200 --> 00:21:55.020 In most stacks, that's the startup. 00:21:56.980 --> 00:22:03.600 I'd want to know how insurers are responding, but I haven't seen a source on that, so it's a question, not a fact. 00:22:04.100 --> 00:22:06.440 The policy reading is where I'm most doubtful. 00:22:07.040 --> 00:22:11.240 Unlicensed practice statutes were written for people presenting themselves as doctors. 00:22:12.100 --> 00:22:17.740 OpenAI says ChatGPT has never been a substitute for professional legal or medical advice, 00:22:17.740 --> 00:22:22.820 and a court could easily find a disclaimed general tool isn't practicing anything. 00:22:24.640 --> 00:22:27.920 That's a fair challenge, and the facts don't settle it. 00:22:28.500 --> 00:22:33.400 If courts did accept that framing, legislators wouldn't need A-specific medical rules. 00:22:33.920 --> 00:22:36.200 Existing licensing law would already reach. 00:22:36.800 --> 00:22:39.480 Where every reading lands in the same place is the cost. 00:22:39.480 --> 00:22:43.940 It flows to whoever is closest to the user and has the least bargaining power. 00:22:44.660 --> 00:22:46.400 In most stacks, that's the startup. 00:22:48.320 --> 00:22:54.960 I'd want to know how insurers are responding, but I haven't seen a source on that, so it's a question, not a fact. 00:22:55.480 --> 00:22:57.800 The policy reading is where I'm most doubtful. 00:22:58.420 --> 00:23:02.640 Unlicensed practice statutes were written for people presenting themselves as doctors. 00:23:03.480 --> 00:23:09.120 OpenAI says ChatGPT has never been a substitute for professional legal or medical advice. 00:23:09.480 --> 00:23:14.200 and a court could easily find a disclaimed general tool isn't practicing anything. 00:23:16.040 --> 00:23:19.300 That's a fair challenge, and the facts don't settle it. 00:23:19.840 --> 00:23:24.780 If courts did accept that framing, legislators wouldn't need a specific medical rules. 00:23:25.300 --> 00:23:27.580 Existing licensing law would already reach. 00:23:28.160 --> 00:23:30.860 Where every reading lands in the same place is the cost. 00:23:30.860 --> 00:23:35.320 It flows to whoever is closest to the user and has the least bargaining power. 00:23:36.040 --> 00:23:37.800 In most stacks, that's the startup. 00:23:39.480 --> 00:23:46.380 And I think the design framing will keep surviving dismissal, which on its own forces spending on safety tooling. 00:23:47.040 --> 00:23:48.420 Your worry is fair, though. 00:23:48.960 --> 00:23:54.660 If this case loses on causation, people could wrongly read it as a green light for the whole category. 00:23:56.620 --> 00:23:58.920 I'd pair that with a contract question. 00:23:59.480 --> 00:24:05.820 Has the founder read the indemnity in their model provider's terms, and do they know what it exposes them to? 00:24:06.420 --> 00:24:09.100 Where we still disagree is this case itself. 00:24:09.480 --> 00:24:15.820 I think the unlicensed practice count is the overreach, and the causation fight could make it a weak test. 00:24:17.520 --> 00:24:24.280 And I think the design framing will keep surviving dismissal, which on its own forces spending on safety tooling. 00:24:24.940 --> 00:24:26.320 Your worry is fair, though. 00:24:26.860 --> 00:24:32.560 If this case loses on causation, people could wrongly read it as a green light for the whole category. 00:24:34.560 --> 00:24:36.800 I'd pair that with a contract question. 00:24:36.800 --> 00:24:42.020 has the founder read the indemnity in their model providers terms and do they 00:24:42.020 --> 00:24:47.180 know what it exposes them to where we still disagree is this case itself I 00:24:47.180 --> 00:24:52.340 think the unlicensed practice count is the overreach and the causation fight 00:24:52.340 --> 00:24:58.220 could make it a weak test and I think the design framing will keep surviving 00:24:58.220 --> 00:25:03.500 dismissal which on its own forces spending on safety tooling your worry is 00:25:03.500 --> 00:25:08.500 fair though if this case loses on causation people could wrongly read it 00:25:08.500 --> 00:25:14.740 as a green light for the whole category I'd pair that with a contract question 00:25:14.740 --> 00:25:19.940 has the founder read the indemnity in their model providers terms and do they 00:25:19.940 --> 00:25:25.160 know what it exposes them to where we still disagree is this case itself I 00:25:25.160 --> 00:25:30.260 think the unlicensed practice count is the overreach and the causation fight 00:25:30.260 --> 00:25:31.620 could make it a weak test. 00:25:33.300 --> 00:25:36.980 And I think the design framing will keep surviving dismissal, 00:25:37.400 --> 00:25:40.080 which on its own forces spending on safety tooling. 00:25:40.740 --> 00:25:42.120 Your worry is fair, though. 00:25:42.660 --> 00:25:44.680 If this case loses on causation, 00:25:45.120 --> 00:25:48.360 people could wrongly read it as a green light for the whole category. 00:25:50.240 --> 00:25:53.380 I'd add one faster signal, the contracts. 00:25:53.960 --> 00:25:56.240 If providers tighten their terms on health use 00:25:56.240 --> 00:25:59.260 or startups start negotiating indemnity limits, 00:25:59.260 --> 00:26:02.180 the market is telling us who it expects to lose. 00:26:02.880 --> 00:26:06.900 Court rulings take years, and terms of service can change much faster. 00:26:09.120 --> 00:26:10.260 Third, the transcripts. 00:26:11.140 --> 00:26:13.600 Discovery should produce the full conversation, 00:26:14.120 --> 00:26:17.400 including whether the model ever told her to talk to her doctor. 00:26:17.980 --> 00:26:21.360 That's where these allegations either hold up or soften. 00:26:23.180 --> 00:26:26.360 I'd add one faster signal, the contracts. 00:26:26.360 --> 00:26:35.160 If providers tighten their terms on health use or startups start negotiating indemnity limits, the market is telling us who it expects to lose. 00:26:35.840 --> 00:26:39.880 Court rulings take years, and terms of service can change much faster. 00:26:41.720 --> 00:26:43.360 Third, the transcripts. 00:26:44.120 --> 00:26:50.380 Discovery should produce the full conversation, including whether the model ever told her to talk to her doctor. 00:26:50.900 --> 00:26:54.300 That's where these allegations either hold up or soften. 00:26:56.360 --> 00:26:59.340 I'd add one faster signal, the contracts. 00:26:59.920 --> 00:27:02.200 If providers tighten their terms on health use 00:27:02.200 --> 00:27:05.200 or startups start negotiating indemnity limits, 00:27:05.560 --> 00:27:08.140 the market is telling us who it expects to lose. 00:27:08.820 --> 00:27:10.200 Court rulings take years, 00:27:10.580 --> 00:27:12.820 and terms of service can change much faster. 00:27:14.740 --> 00:27:16.180 Here's Stephen's take. 00:27:16.700 --> 00:27:19.580 Let's get serious and talk about the provider liability. 00:27:20.260 --> 00:27:22.120 Note the provider is multifold. 00:27:22.680 --> 00:27:24.740 The reseller above the model has responsibility 00:27:24.740 --> 00:27:26.820 for their solution to the end user, 00:27:27.260 --> 00:27:28.740 and the model layer as well. 00:27:29.540 --> 00:27:31.300 SLAs are likely to be critical 00:27:31.300 --> 00:27:33.880 and are going to read much more real-world-like 00:27:33.880 --> 00:27:36.340 than the GI marketing hype, I believe. 00:27:38.320 --> 00:27:41.320 That's the deep dive and today's AI briefing. 00:27:42.060 --> 00:27:44.060 Every source is linked in the show notes, 00:27:44.060 --> 00:27:47.280 and the AI briefing webpage has the written edition, 00:27:47.740 --> 00:27:50.100 this deep dive and every past episode. 00:27:50.780 --> 00:27:52.480 The daily briefing is back tomorrow.