WEBVTT 00:00:00.000 --> 00:00:07.300 On today's AI briefing, we think the TSMC disclosure is less about anthropic than about the sector. 00:00:07.960 --> 00:00:11.440 Every frontier lab carries the same Taiwan exposure, 00:00:11.940 --> 00:00:17.460 which turns compute access into a geopolitical position rather than a procurement line. 00:00:19.060 --> 00:00:23.020 This is the AI briefing for Friday, October 30. 00:00:23.340 --> 00:00:23.920 I'm Michael. 00:00:24.400 --> 00:00:28.940 Like everything here, this briefing is written and read by AI from public sources. 00:00:29.640 --> 00:00:33.860 Today's stories share a theme, who controls compute and who gets funded. 00:00:34.540 --> 00:00:41.140 Anthropic has filed to go public, and the filing names a heavy dependence on TSMC chips made in Taiwan. 00:00:41.940 --> 00:00:45.420 Cohere is raising a series E of up to CAD $7 billion. 00:00:46.920 --> 00:00:53.900 Ottawa is putting in $1.6 billion, released only as jobs and revenue milestones are met. 00:00:54.620 --> 00:00:58.600 Mistral has released open weights for a trillion-parameter coding model, 00:00:58.940 --> 00:01:01.540 which could change what security teams pay for. 00:01:01.980 --> 00:01:08.020 And new funding counts show only 16% of 2024 seed companies have made it past seed. 00:01:08.640 --> 00:01:11.840 We'll take the main stories first, then a few quick hits. 00:01:13.860 --> 00:01:15.560 Stevens' take on the top of the day. 00:01:16.220 --> 00:01:20.520 Cohere has been an amazing story, a company that just keeps finding a way, 00:01:20.520 --> 00:01:23.940 a grip on the mountain to cling to, and a wedge to expand. 00:01:23.940 --> 00:01:31.820 It's easy to be critical of their situation, but I have been a first mover in AI before and rapidly eclipsed by Valley mentality 00:01:31.820 --> 00:01:36.940 Fundraising as a weapon against faster smarter incumbents and in my case 00:01:36.940 --> 00:01:41.660 Naivete about how to play the game as our friend AA puts it 00:01:41.660 --> 00:01:44.220 I like being Canadian 00:01:44.220 --> 00:01:51.120 Non-aligned with corporate malfeasance and market affecting behaviors and I am proud to see them moving forward 00:01:51.840 --> 00:01:57.520 Hopefully, this customer strategic investment by countries gets them to break out velocity. 00:01:59.440 --> 00:02:03.780 First story, and it's an update on one we've been following for several days. 00:02:04.300 --> 00:02:07.700 Anthropics IPO what's new is that the S-1 is filed. 00:02:08.320 --> 00:02:11.820 Tis V. Moshowitz noted the filing and don't worry about the vase. 00:02:12.400 --> 00:02:17.240 TechMeme reports it discloses heavy reliance on TSMC chips made in Taiwan, 00:02:17.740 --> 00:02:20.440 citing geopolitical risk from tensions with China. 00:02:21.140 --> 00:02:23.520 Share price and other financials are not stated. 00:02:24.100 --> 00:02:28.120 The chip claims are unverified and the full filing hasn't been read. 00:02:28.700 --> 00:02:29.660 Our read holds. 00:02:30.420 --> 00:02:34.020 We think this says more about the sector than about Anthropic. 00:02:34.700 --> 00:02:41.420 Every frontier lab carries the same Taiwan exposure, so compute access becomes a geopolitical position. 00:02:42.140 --> 00:02:47.500 For apps built on one lab, a fair question is how they would fail over to a second provider. 00:02:47.500 --> 00:02:52.500 The market reading is that a shared risk won't set the IPO apart. 00:02:53.220 --> 00:02:56.380 Revenue growth and compute commitments will drive the price. 00:02:58.200 --> 00:02:59.720 Here's Stephen's take. 00:03:00.320 --> 00:03:05.180 Anthropic might get there and end up driving value like research breakthroughs themselves, 00:03:05.560 --> 00:03:10.680 but better code and system models are in whole unlikely to get them to the promised land. 00:03:11.160 --> 00:03:14.760 But better and better models are def getting you closer to big, 00:03:14.760 --> 00:03:21.320 valuable breakthroughs in alliance with scientists, research labs with testing facilities, and 00:03:21.320 --> 00:03:24.800 more, maybe they become the world's biggest pharma company. 00:03:26.440 --> 00:03:28.840 Our second story comes from Canada. 00:03:29.500 --> 00:03:34.380 Cohere has recently opened a Series E, according to BetaKit and Cohere's own announcement. 00:03:34.380 --> 00:03:41.060 The company is targeting up to $7 billion CAD at a post-money valuation of roughly $36 00:03:41.060 --> 00:03:41.480 billion. 00:03:42.400 --> 00:03:44.280 Schwarz Group is leading the round. 00:03:44.760 --> 00:03:47.960 Inovia Capital and Canada's CDIC are participating. 00:03:48.620 --> 00:03:53.540 Alongside that, Canada commits $1.6 billion through CDIC. 00:03:54.140 --> 00:03:57.240 That money is staged on jobs and revenue milestones. 00:03:58.620 --> 00:04:04.880 In return, Cohere pledges $1.7 billion in Canadian compute by 2030. 00:04:06.240 --> 00:04:07.540 How well does it hold up? 00:04:07.540 --> 00:04:09.320 The core reporting holds. 00:04:09.820 --> 00:04:12.180 But the $7 billion is a target. 00:04:12.180 --> 00:04:15.180 The size actually closed has not been disclosed. 00:04:16.340 --> 00:04:19.720 The government money flows only as milestones are met. 00:04:20.160 --> 00:04:24.600 And nothing in the reporting supports the claims about Cohere's competitiveness. 00:04:25.380 --> 00:04:29.040 Our view is that the staging matters more than the headline. 00:04:29.740 --> 00:04:32.480 Ottawa is buying jobs and domestic compute. 00:04:32.960 --> 00:04:36.180 It is not underwriting a $36 billion valuation. 00:04:36.180 --> 00:04:42.360 valuation. What would justify that price is revenue from private enterprises, outside 00:04:42.360 --> 00:04:47.120 government. That looks like the number worth watching. 00:04:47.120 --> 00:04:53.740 Read through a policy lens. The same facts look different. Public money staged on jobs, 00:04:53.740 --> 00:05:00.960 revenue and $1.7 billion of domestic compute is how a country builds sovereign capacity. 00:05:00.960 --> 00:05:07.260 On that reading, judging the deal on venture returns alone misses its purpose. 00:05:07.260 --> 00:05:11.500 For early-stage companies, the compute pledge could matter. 00:05:11.500 --> 00:05:16.660 Canadian startups that need domestic hosting might benefit, but whether outside companies 00:05:16.660 --> 00:05:20.420 can actually buy that capacity is not stated. 00:05:20.420 --> 00:05:26.560 That may be a question worth asking before anyone plans around it. 00:05:26.560 --> 00:05:28.380 Here's Stephen's take. 00:05:28.380 --> 00:05:34.020 access to sovereign capability is not unreal. If there was an ally that would not be capable of 00:05:34.020 --> 00:05:38.760 using compute as a weapon, they would have less runway but in reality world affairs, 00:05:39.060 --> 00:05:44.660 and the third scaling Law Plus access to unique content and experimental capability and 00:05:44.660 --> 00:05:49.400 multi-billion dollar national contracts could take them where people like Helsing have gone, 00:05:49.400 --> 00:05:53.940 and others like the entire Silicon Valley emerged in just this way. 00:05:54.760 --> 00:05:56.420 Read Steve Blank for example. 00:05:58.380 --> 00:06:06.120 For our third story, Mistral has released open weights for large-for-its-trillion-parameter model, nicknamed LeChonk. 00:06:06.580 --> 00:06:11.220 The news comes from Mistral's own announcement, with a report from TechCrunch. 00:06:12.500 --> 00:06:15.440 Mistral claims it is the best open-weight coding model. 00:06:15.440 --> 00:06:23.440 Its benchmarks show 19.7% on Frontier Code 2.0 and 41% on Terminal Bench 4.0. 00:06:23.440 --> 00:06:32.400 Mistral also says it ties GLM 5.3 at 48% while using 47% fewer tokens. 00:06:33.120 --> 00:06:39.440 TechCrunch reports the model is aimed at cybersecurity work, with more control than closed models offer. 00:06:40.660 --> 00:06:44.380 How well does it hold up? So far, the picture is mixed. 00:06:44.940 --> 00:06:48.460 Those benchmark leads are self-reported and unverified. 00:06:48.460 --> 00:06:52.480 The claim of being the best goes further than the evidence shows, 00:06:52.480 --> 00:06:55.820 and the cybersecurity angle comes with little detail. 00:06:57.220 --> 00:07:00.360 We read this as a pricing event before a capability one. 00:07:01.000 --> 00:07:04.380 A one trillion coding model that a company can host itself 00:07:04.380 --> 00:07:08.160 opens room for security and regulated sector startups 00:07:08.160 --> 00:07:11.180 that cannot send their code to a closed API. 00:07:12.280 --> 00:07:15.820 For early stage companies, that cuts two ways. 00:07:16.440 --> 00:07:21.500 Startups whose edge is wrapping closed coding models for security work could lose ground. 00:07:21.500 --> 00:07:28.500 Startups with fine-tuning and on-prem deployment skills may gain from a credible open-weight base. 00:07:28.500 --> 00:07:31.500 A few questions seem worth asking. 00:07:31.500 --> 00:07:36.500 Do these results hold when someone outside Miestral runs the benchmarks? 00:07:36.500 --> 00:07:40.500 What does the cybersecurity use case look like in practice? 00:07:40.500 --> 00:07:44.500 And for a startup building security tools on closed models, 00:07:44.500 --> 00:07:49.500 how much of its edge survives if a self-hosted open model is good enough? 00:07:51.500 --> 00:07:56.680 Here's Steven's take. Impressive. I guess we can all distill how they built this. 00:07:58.320 --> 00:08:03.800 The last of the main stories turns from models to money and to the companies trying to raise it. 00:08:04.320 --> 00:08:09.520 In a piece published yesterday, Crunchbase News looks at what large Series A rounds are hiding. 00:08:10.180 --> 00:08:16.940 In the third quarter, U.S. Series A rounds fell about 28% year-over-year to 272. 00:08:17.680 --> 00:08:21.760 Seed rounds dipped to 984 down from 1,059. 00:08:22.180 --> 00:08:26.360 The median U.S. seed round in 2025 was about $3 million. 00:08:27.680 --> 00:08:30.260 The striking figures are on graduation. 00:08:30.960 --> 00:08:37.980 Only 24% of 2023 companies that raised a seed round of $1 million or more have moved past seed. 00:08:38.460 --> 00:08:40.780 For 2024, it's 16%. 00:08:40.780 --> 00:08:44.500 The piece is a guest column by Mark Schroeder of MGV. 00:08:44.960 --> 00:08:46.900 On the evidence, it's a mixed picture. 00:08:46.940 --> 00:08:50.860 The crunch-based counts do show a real Series A squeeze. 00:08:51.440 --> 00:08:56.620 The prescription that founders need to clear a revenue bar is one investor's view. 00:08:57.500 --> 00:09:01.000 It also doubles as a pitch for MGV's own approach. 00:09:01.440 --> 00:09:05.000 And a data company has its own reasons to highlight market stress. 00:09:05.460 --> 00:09:06.540 Here's where we come out. 00:09:06.900 --> 00:09:11.700 We think the 16% graduation rate for 2024 seeds is the number to carry. 00:09:12.160 --> 00:09:15.820 The Series A bar now looks like revenue quality, not traction. 00:09:15.820 --> 00:09:22.600 The data supports that much. The prescription is MGVs from a fund pitching its own method. 00:09:23.200 --> 00:09:25.900 For seed investors, the implication is practical. 00:09:27.220 --> 00:09:33.860 With so few companies graduating, seed funds may need larger reserves for extensions and bridge rounds. 00:09:34.540 --> 00:09:40.580 A fair question for any fund might be how its reserves plan looks against those graduation rates. 00:09:40.580 --> 00:09:55.320 Here's Stephen's take. I don't know what a Series A is. I just know that this is the gold rush day zero, and people better start laying claims and digging holes fast and supporting those that do fast. 00:09:55.320 --> 00:09:59.960 Now two quick hits, both about scale. 00:10:00.540 --> 00:10:05.860 First, NVIDIA is in talks to give N-Scale billions of dollars in credit guarantees. 00:10:06.580 --> 00:10:11.700 The debt would fund a 460-megawatt data center that Anthropic plans to use. 00:10:12.140 --> 00:10:16.600 It's another sign that chip makers are backstopping the financing of their own demand. 00:10:16.920 --> 00:10:24.500 Second, Microsoft will offer a voluntary exit to about 60,000 of its 223,000 staff. 00:10:25.320 --> 00:10:42.260 Azure argues this reflects productivity gains, with revenue per employee up two-thirds since 2022 to $1.5 million. He calls AI the cause in a not-yet sense. Our evidence check finds that argument overstated. 00:10:43.900 --> 00:10:52.340 On Simon Willison on Spawn's AI-driven game development, here's Stephen's take. To be honest, this is sort of easy now. 00:10:52.340 --> 00:10:56.880 Stephen's take on the quick hits. 00:10:56.880 --> 00:11:02.920 NVIDIA credit guarantees is interesting and important for emerging meta-hyperscalers being 00:11:02.920 --> 00:11:08.980 held back by access to compute by debt risk and the potential to use this as yet another 00:11:08.980 --> 00:11:12.940 lever against competition. 00:11:12.940 --> 00:11:18.340 And Stephen's take on the day, or doubles, and what is your plan to separate yourself 00:11:18.340 --> 00:11:24.420 from OpenAI disaster, lack of compute as you need it at a predictable price, being able 00:11:24.420 --> 00:11:30.660 to offload inference to delayed or training to background. Basically, disaster and change 00:11:30.660 --> 00:11:31.140 planning. 00:11:32.780 --> 00:11:38.260 To close, a diligence question that might be worth asking this week. What does your gross 00:11:38.260 --> 00:11:43.740 margin look like if your model providers price or supply changes by half? And one question 00:11:43.740 --> 00:11:50.740 to watch, how much of Cohere's series E actually closed, and what releases Canada's $1.6 billion. 00:11:51.420 --> 00:11:57.980 A closing announcement, or published CDIC terms with a milestone schedule, would settle it. 00:11:59.680 --> 00:12:05.860 That's the briefing. Stay with me. Right after this, Bella takes you through today's deep dive 00:12:05.860 --> 00:12:12.280 on Anthropics leaked AI bill and the fight over preemption. You can also read it or listen to it 00:12:12.280 --> 00:12:18.300 on its own on the AI Briefing webpage, where you'll find the written edition and every past episode. 00:12:18.740 --> 00:12:20.480 The link is in the show notes. 00:12:23.520 --> 00:12:31.060 Thanks, Michael. This is the AI Briefing Deep Dive for Friday, October 30, Anthropics Leaked AI Bill 00:12:31.060 --> 00:12:38.120 and the Fight Over Preemption. I'm Bella. Like the briefing, it is written and read by AI from public sources. 00:12:38.120 --> 00:12:45.900 On October 16, Politico reported that Anthropic had been quietly showing lawmakers on Capitol 00:12:45.900 --> 00:12:51.800 Hill a draft federal bill to regulate frontier AI. Politico had the confidential draft. 00:12:52.440 --> 00:12:56.440 Anthropic has not published it. The reaction has kept building since then. 00:12:57.060 --> 00:13:01.640 This week, Tizvi Moshowitz, who writes the newsletter Don't Worry About the Vase, 00:13:01.640 --> 00:13:04.480 called the leaked draft unacceptable and troubling. 00:13:04.480 --> 00:13:10.420 California State Senator Scott Weiner said it falls far short of protecting public health and 00:13:10.420 --> 00:13:16.520 safety. Anthropic has partly retreated. Its head of federal affairs, Anthony Cimino, said the 00:13:16.520 --> 00:13:21.560 preemption wording was not precise enough and would be updated. He also said the proposal is 00:13:21.560 --> 00:13:26.600 meant to be the strongest on the table. So this is the moment to ask a structural question. 00:13:27.220 --> 00:13:32.300 Suppose a frontier lab helps write the federal rules, and those rules override the states. 00:13:32.300 --> 00:13:35.900 Who does that lock in and who gets locked out? 00:13:37.580 --> 00:13:41.300 Here is the draft as Politico and other outlets describe it. 00:13:41.740 --> 00:13:44.700 The Commerce Department would register independent evaluators. 00:13:45.340 --> 00:13:49.440 Those evaluators would periodically assess frontier developers and their models. 00:13:50.160 --> 00:13:54.980 Developers could pay their own evaluators, as long as the payment is not tied to the result. 00:13:55.740 --> 00:14:00.640 Evaluators would have to tell Commerce within 72 hours if they find an imminent risk. 00:14:01.240 --> 00:14:04.800 The bill also requires risk reports and whistleblower protections. 00:14:06.320 --> 00:14:08.760 ZVI calls its security requirements strong. 00:14:09.320 --> 00:14:10.920 The fight is over two other parts. 00:14:11.400 --> 00:14:13.200 The first is the bar for stopping a model. 00:14:13.720 --> 00:14:19.640 The government could step in only when a model poses a foreseeable and imminent risk of more than $1 billion in damage, 00:14:20.060 --> 00:14:23.040 or of killing or seriously injuring more than 50 people. 00:14:23.700 --> 00:14:26.480 The risk would also have to come from a narrow set of scenarios. 00:14:27.100 --> 00:14:32.260 Bioweapons, a massive cyberattack, a model evading its developers' safety controls, 00:14:32.620 --> 00:14:34.500 or research that speeds those up. 00:14:35.320 --> 00:14:38.540 Only a federal judge in Washington could order a suspension. 00:14:39.240 --> 00:14:40.540 The second part is preemption. 00:14:41.080 --> 00:14:43.520 That means a federal law displacing state law. 00:14:44.020 --> 00:14:49.920 The draft would override California's SB53 New York's Rays Act and Illinois SB315. 00:14:50.740 --> 00:14:52.560 One way to picture it is building codes. 00:14:53.140 --> 00:14:57.480 Today, states write their own fire codes for one specific kind of tall building. 00:14:58.140 --> 00:15:00.960 Under this draft, Washington writes the only code. 00:15:01.520 --> 00:15:03.060 Only Congress can change it. 00:15:04.340 --> 00:15:08.280 Inspectors are hired and paid by the builder, and no one licenses them. 00:15:08.820 --> 00:15:11.280 A city that sees smoke cannot close the building. 00:15:11.780 --> 00:15:16.140 It has to ask a federal judge, and the judge must weigh what closing would cost the owner. 00:15:16.800 --> 00:15:19.860 That picture also shows what the draft does not appear to cover. 00:15:19.860 --> 00:15:22.120 It is aimed at frontier developers. 00:15:22.760 --> 00:15:25.060 The reporting does not say who counts as one. 00:15:25.700 --> 00:15:30.080 A seed stage company building on top of a model is probably not the regulated party. 00:15:31.120 --> 00:15:35.880 It may still feel the effects, through its suppliers and through the market for audits. 00:15:37.620 --> 00:15:39.260 Start with what is established. 00:15:39.960 --> 00:15:42.560 Politico reported the draft from a confidential copy 00:15:42.560 --> 00:15:46.360 and from three people familiar with Anthropic's legislative work. 00:15:47.000 --> 00:15:48.440 Anthropic has not denied it. 00:15:49.020 --> 00:15:54.120 Cimino publicly said the company circulated the ideas to get early feedback from policymakers. 00:15:55.000 --> 00:15:59.400 On October 18, he posted that Anthropic would update its preemption language 00:15:59.400 --> 00:16:02.000 to better match its own published framework. 00:16:02.840 --> 00:16:06.500 He also said the company would address the standard for blocking models, 00:16:06.960 --> 00:16:09.460 risk thresholds, and requirements for evaluators. 00:16:10.040 --> 00:16:11.180 So the draft is real. 00:16:11.800 --> 00:16:14.840 Anthropic stands behind its direction, but not its wording. 00:16:15.660 --> 00:16:18.360 Several provisions are confirmed by more than one outlet. 00:16:19.000 --> 00:16:22.880 These include the $1 billion and 50 casualty thresholds, 00:16:23.200 --> 00:16:24.660 the federal court in Washington, 00:16:25.200 --> 00:16:27.080 the evaluators paid by labs, 00:16:27.400 --> 00:16:29.360 and the list of state laws it would preempt. 00:16:30.700 --> 00:16:34.440 Cimino says the harm thresholds were pulled directly from California law. 00:16:34.960 --> 00:16:37.700 That holds up, with a catch that Weiner points out. 00:16:38.320 --> 00:16:42.000 SB 53 uses those thresholds to trigger disclosure and assessment. 00:16:42.240 --> 00:16:45.620 The draft uses them to decide when a model can be stopped. 00:16:46.280 --> 00:16:47.600 Next come the legal readings, 00:16:47.600 --> 00:16:49.940 which are analysis, not fact. 00:16:50.560 --> 00:16:55.160 Charlie Bullock of the Institute for Law and AI says courts must weigh the harm a blocking 00:16:55.160 --> 00:16:56.520 order does to the developer. 00:16:57.520 --> 00:17:00.940 In his reading, they are not told to weigh the harm of not blocking. 00:17:01.680 --> 00:17:05.320 Writing in Transformer, he warns of a race to the bottom among auditors. 00:17:06.000 --> 00:17:09.480 In this reading, they would compete to offer the least bothersome audits. 00:17:10.140 --> 00:17:15.060 Nathan Calvin, as ZVI quotes him, argues that passing the bill would leave us less 00:17:15.060 --> 00:17:16.800 protected than existing law does. 00:17:17.460 --> 00:17:20.120 These are serious readings by people who know the field. 00:17:20.620 --> 00:17:22.280 They are also from safety advocates, 00:17:22.700 --> 00:17:25.700 and they are interpretations of a text the public cannot read. 00:17:26.640 --> 00:17:28.740 Then there is what nobody has shown. 00:17:29.240 --> 00:17:30.820 The full text is not public. 00:17:31.300 --> 00:17:34.500 Who counts as covered is not stated in any coverage we found. 00:17:35.120 --> 00:17:37.420 Neither is any compute or revenue cutoff. 00:17:37.420 --> 00:17:41.960 The revised draft and the promised framework for evaluators have not appeared. 00:17:42.620 --> 00:17:46.780 Any claim about how this affects small builders is, for now, inference. 00:17:48.480 --> 00:17:50.920 The same facts support several readings. 00:17:51.560 --> 00:17:55.080 The safety reading, held by ZVI, Calvin, and Wiener, 00:17:55.500 --> 00:17:58.160 comes from people who take catastrophic risk seriously. 00:17:58.900 --> 00:18:02.200 In this view, the draft is weaker than the bipartisan Frontier Act 00:18:02.200 --> 00:18:04.460 from Representatives Obernault and Trahan. 00:18:05.060 --> 00:18:07.980 They argue it removes state backstops that work today 00:18:07.980 --> 00:18:11.620 and freezes weak federal rules that only Congress can tighten. 00:18:11.960 --> 00:18:17.320 Zevi's sharpest line is that Anthropic seems more afraid that the government 00:18:17.320 --> 00:18:21.220 might enforce rules against Anthropic than worried about the government's power 00:18:21.220 --> 00:18:22.980 to enforce rules against everyone. 00:18:23.640 --> 00:18:26.100 The market reading is the strongest case for the draft. 00:18:26.660 --> 00:18:29.620 A single federal standard is cheaper to comply with than 50. 00:18:30.340 --> 00:18:34.300 Clear thresholds and a single court cut legal uncertainty for every developer. 00:18:34.660 --> 00:18:38.580 One outlet reports that child safety, consumer protection, 00:18:38.780 --> 00:18:41.140 and civil rights laws are carved out of preemption. 00:18:41.960 --> 00:18:45.260 If so, states keep their role on everyday harms. 00:18:45.760 --> 00:18:47.400 The realist reading is more skeptical. 00:18:48.080 --> 00:18:51.920 The scenarios that would justify a shutdown are narrow, and none has happened yet. 00:18:52.460 --> 00:18:56.860 On that view, how blocking works matters less than who writes the standards for evaluators 00:18:56.860 --> 00:18:58.000 and who pays for them. 00:18:58.460 --> 00:19:01.140 That is where an incumbent's advantage would build up over time. 00:19:01.680 --> 00:19:04.100 The policy reading points to a shift in roles. 00:19:04.760 --> 00:19:08.900 Earlier, Anthropic endorsed SB 53 and similar state laws. 00:19:09.340 --> 00:19:11.140 It is now proposing to replace them. 00:19:11.960 --> 00:19:13.920 That doesn't prove bad faith. 00:19:14.380 --> 00:19:18.400 It does mean a lab's policy position is something we should track over time, 00:19:18.700 --> 00:19:19.840 not take as fixed. 00:19:21.580 --> 00:19:24.100 Our read is that the label of incumbent capture 00:19:24.100 --> 00:19:27.040 fits the evaluation market more than it fits the founders. 00:19:27.820 --> 00:19:31.520 Frontier-level obligations will almost certainly fall on a few large labs. 00:19:32.100 --> 00:19:35.300 Most seed and series A companies will not be covered directly. 00:19:35.900 --> 00:19:37.740 The indirect effects are where it gets real. 00:19:38.260 --> 00:19:41.540 One is evaluators who are paid by labs and not licensed. 00:19:41.960 --> 00:19:47.560 another is standards that only congress can update a third is state regimes that disappear 00:19:48.760 --> 00:19:53.000 together they decide what assurance means for every model your product sits on 00:19:53.000 --> 00:19:58.760 for startups selling evaluation red teaming or ai assurance this is close to an existential 00:19:58.760 --> 00:20:04.360 question a weak federal floor could create a big market for cheap compliance but not for rigorous 00:20:04.360 --> 00:20:10.760 testing for application companies the question is quieter their enterprise customers may now point to 00:20:10.760 --> 00:20:16.040 SB53 disclosures from their model providers. Those could vanish or change shape. 00:20:16.800 --> 00:20:22.420 We hold all of this loosely because the text is not public and a revision has been promised. 00:20:23.140 --> 00:20:28.260 One question worth exploring with an assurance startup is this. If the federal standard were 00:20:28.260 --> 00:20:33.780 the floor in this draft, would the main buyers still pay for your product? Or does your revenue 00:20:33.780 --> 00:20:40.700 depend on state law that could be preempted? Three things would settle this. First is 00:20:40.700 --> 00:20:42.820 Anthropic's revised preemption language. 00:20:43.360 --> 00:20:45.640 Simino promised it, but it has not appeared. 00:20:46.240 --> 00:20:48.020 When it does, the test is simple. 00:20:48.500 --> 00:20:52.860 Does it still stop states from blocking a model in an emergency, or only from setting 00:20:52.860 --> 00:20:54.460 conflicting disclosure rules? 00:20:55.040 --> 00:20:56.920 Second is the evaluator framework. 00:20:57.540 --> 00:21:02.320 Licensing, conflict of interest rules, and a say for commerce over standards would answer 00:21:02.320 --> 00:21:04.000 most of the race-to-the-bottom worry. 00:21:04.980 --> 00:21:10.680 Third is whether any member of Congress introduces the draft, or folds it into the Frontier Act. 00:21:10.700 --> 00:21:15.540 A bill text with a sponsor is the moment this becomes law in the making, 00:21:15.940 --> 00:21:17.460 rather than a lobbying document. 00:21:18.220 --> 00:21:20.100 None of these has a public date yet. 00:21:21.980 --> 00:21:23.380 Here's Stephen's take. 00:21:23.980 --> 00:21:27.260 Companies becoming governments is a big part of sci-fi lore, 00:21:27.600 --> 00:21:29.120 and here we are living in the future. 00:21:31.040 --> 00:21:34.040 That's the deep dive and today's AI briefing. 00:21:34.780 --> 00:21:36.780 Every source is linked in the show notes, 00:21:36.780 --> 00:21:40.000 and the AI briefing webpage has the written edition, 00:21:40.000 --> 00:21:42.820 this deep dive and every past episode. 00:21:43.500 --> 00:21:45.200 The Daily Briefing is back tomorrow.