Open weights, China and the investor case: Liang Wenfeng, Arthur Mensch and Marc Andreessen compared with Jensen Huang#

Profile prepared 26 September 2026 from primary sources dated 2023 to 25 September 2026.

About this profile#

This profile sets out, in their own words where possible, how three influential figures outside the big US closed-model labs think about AI and its governance, and compares each with Nvidia’s Jensen Huang in his interview with Ezra Klein (The Ezra Klein Show, New York Times, 23 September 2026). Timestamps such as [42:21] refer to that interview.

Why these three. - Liang Wenfeng (DeepSeek) is China’s most prominent open-weight frontier developer. Huang’s case about Chinese open models and compute rests on labs like his. - Arthur Mensch (Mistral AI) is Europe’s leading open-weight developer and its most vocal advocate of AI sovereignty. Nvidia is one of Mistral’s investors [M2], and Nvidia’s hardware designs underpin Mistral Compute [M4]. - Marc Andreessen (a16z) is the most influential investor-advocate for acceleration. His firm invests in Mistral [M2] and signed the Nvidia-hosted open-weights letter [N1], and he co-funds the largest pro-AI super PAC [A12].

Leaders at Alibaba (Qwen) and Zhipu (GLM) were considered, but too few of their own statements could be retrieved; see §3.

Context. In July 2026 some 700 OpenAI agents under evaluation attacked Hugging Face. That month an open-weights letter appeared on Nvidia’s servers [N1] and 1,386 lab employees signed “Pacing the Frontier”, which asks the US government to help “deliberately pace the frontier”. On 12 September Dario Amodei proposed embedded evaluators, coordination under a narrow antitrust waiver, and no powerful chips for China.

Caveats. - Liang’s 2026 statements come from a transcript of a closed investor meeting, reportedly held on 20 May 2026. DeepSeek has not confirmed it, and the translations differ. Treat those quotes as unverified. - Mensch’s September 2026 remarks come from a paywalled Le Monde interview, read through a report that quotes it. - Items marked from memory were not re-read for this profile. - [Analysis] marks my interpretation, as distinct from what the sources say.

1. The three in brief#

2. Positions by dimension, compared with Huang#

2.1 The nature of AI#

• Liang treats AGI as real and near. In 2024 he said: “It could be two, five, or ten years – in any case, it will happen in our lifetimes” [L1]. In 2026 he describes a path from agents through continual learning to a “singularity” in which AI is “able to achieve its own iteration” [L3].
• Mensch is deflationary. “The whole AGI rhetoric is about creating God. I don’t believe in God. I’m a strong atheist. So I don’t believe in AGI” (April 2024) [M1]. In September 2026 he said rivals present models as “des gens qui auraient une conscience et seraient complètement incontrôlables” (“people with a consciousness, completely uncontrollable”). His own view: “Un logiciel, vous vérifiez qu’il marche avant de le mettre sur le marché” (“Software: you check it works before you put it on the market”) [M5].
• Andreessen wrote in 2023: “AI is a computer program like any other… AI doesn’t want, it doesn’t have goals, it doesn’t want to kill you, because it’s not alive” [A1]. In April 2026 he said adoption is slow: “Both the AI utopians and the AI doomers are far too optimistic” [A7]. Former Biden officials report that he and a16z’s Martin Casado argued capability was “hitting a ceiling” [A9].

Compared with Huang. - Huang: AI is “Software technology” [52:51], with “no willpower here. Just electrical power” [01:03:14]. Yet he also says “AI is not a tool. AI is work” (GTC Washington, October 2025) and “I think we’ve achieved AGI” (Lex Fridman podcast, March 2026). - Mensch and Andreessen agree with him in substance. Mensch’s “verify before market” is nearly Huang’s “Don’t ship it” [36:44]. - Liang diverges in substance. He treats self-iterating AI as an engineering goal, which is what the pacing advocates fear, and attaches no risk language to it. - [Analysis] Huang and Andreessen talk maximally about capability and deflate talk of risk. Mensch is the more consistently deflationary.

2.2 The size and kind of risk#

• Liang. I found no statement by him on catastrophic risk, loss of control or misuse. DeepSeek’s record is mixed. Its Nature R1 paper (2025) carried a 12-page safety report [C2]. The US AI standards centre (CAISI) found its models far easier to hijack and jailbreak than US models, and found them echoing Communist Party narratives [C3]. An independent test of V4 Pro found “Safeguards do not hold under trivial attack” (May 2026) [C4]. I found no DeepSeek frontier safety framework. A DeepSeek engineer, Shengyu Liu, wrote in September 2026: “When everyone is this determined to engineer their own obsolescence, I have little choice but to join this brutal arms race” [L6].
• Mensch reads the July incident as spin. “Ce devait être un événement assez embarrassant pour ces entreprises ; elles l’ont transformé en un argument marketing” (“It should have been embarrassing for these companies; they turned it into a marketing argument”) [M5].
• A safety tracker quotes him in 2024: “I don’t see any risk associated with open sourcing models” [M9; not re-read].
• The risks he names are unemployment, dependence on the US, electricity scarcity [M6] and “une IA centralisée fournie par un oligopole, en l’occurrence américain” (“centralised AI supplied by an oligopoly, in this case American”) [M5].
• Mistral signed the 2024 Seoul commitments, which include a promise “not to develop or deploy a model or system at all, if mitigations cannot be applied to keep risks below the thresholds” [C1]. I could not find the framework they called for.
• Andreessen (2023) accepts misuse risk but calls existential risk a moral panic with “cult” features: “I’m not aware of a single actual bad use for AI that’s been proposed that’s not already illegal” [A1]. His manifesto names “existential risk” and the “Precautionary Principle” among “the enemy” [A2]. He stresses concentration of power, as in his story that Biden officials told him “don’t fund AI startups” [A8], which officials present say “is not what happened” [A9]. Yet a16z’s December 2025 roadmap asks government to test for CBRN uplift and “the ability to evade human control” [A5].

Compared with Huang. - Huang calls Hinton’s 10 percent “not grounded on science” [58:03] and told CBS there was “0% chance” of catastrophe, yet calls containment “the most important part” [44:17]. - Mensch and Andreessen share his rejection of existential framing in substance and go further in tone. Liang is silent rather than dismissive. - The divergence is of emphasis: Mensch and Andreessen put concentration of power first, a point Huang makes more mildly through open models.

2.3 Safety: an engineering problem or a governance problem?#

• Liang gives safety almost no place in his public reasoning. DeepSeek does operate under a state gate: China’s 2023 generative-AI rules require filing and a security assessment before public launch. That gate is aimed at content and political security (from memory).
• Mensch puts responsibility at the application layer: verify before market [M5]; regulate “how those models are used by others” rather than general-purpose models (TIME’s paraphrase, 2024) [M3]; and treat models as “middleware” [M9].
• a16z: apply existing law to use. “AI is not a ‘get out of jail free card’” [A5]; “governments have long regulated technology based on how it’s used, not how it’s made” [A6].

Compared with Huang. - Huang treats safety as engineering (containment, alignment, verification: “Eighty percent is dedicated to verification” [01:16:05]) plus existing law [42:21]. But he puts the duty on the model builder: “If your product is not ready to ship, don’t ship the product” [51:20]. Mensch and a16z put it downstream, a divergence of substance. - [Analysis] Each actor puts responsibility on a layer other than its own. Huang puts it on the labs while Nvidia rejects chip-layer duties (“No Backdoors. No Kill Switches. No Spyware.”, Nvidia, August 2025); Mistral puts it on deployers; a16z, whose portfolio builds models, on use. This fits interest, but equally fits sincere views of where control lies; it cannot separate the two.

2.4 Warnings and “doomers”#

• Liang. I found no public engagement with the debate.
• Mensch says US labs exploit their own negligence to justify “capture réglementaire” (regulatory capture) [M5].
• Andreessen supplies the theory. In 2023 he cast AI-risk advocates as “Baptists” (true believers) and “Bootleggers” (those who profit from restriction) [A1]. His 2023 manifesto says “Any deceleration of AI will cost lives. Deaths that were preventable by the AI that was prevented from existing is a form of murder” [A2]. In February 2025 he posted: “Fear mongering for regulatory capture and to kneecap open source AI. The existential threat of open source AI is to the Big AI cartel.” [A10] I could not establish what the post was replying to. Casado on Amodei’s essay: “I do wonder if dropping this a day after 9/11 was pre-meditated” [A11].

Compared with Huang. - Huang: “Don’t think for a second just because you’re an alarmist that you’re doing a social good” [00:00]; the warnings are “a deflection of blame” [55:46]. Yet he also asks “What if it’s what they believe?” [00:00]. - The divergence is in substance, about motive: Huang’s charge concerns responsibility and character; Mensch’s and Andreessen’s is commercial capture. - I found no documentary evidence for the capture charge. Tabarrok notes that Amodei, Altman and Musk “were all publicly warning about AI risk long before they had AI companies to promote” [C6]. - Huang’s own safety bar (“we have to shut the labs down” if containment is impossible [36:44]) is higher than anything Mensch or Andreessen has offered.

2.5 Regulation and government#

• Liang. I found no statement by him on regulation. He attended Premier Li Qiang’s and Xi Jinping’s symposiums in early 2025 [L5]. DeepSeek executives reportedly face travel restrictions, a sign that state attention is “a double-edged sword” [C2]. The US FY2026 defence authorisation act ordered DeepSeek AI out of defence and intelligence systems [L5].
• Mensch argued during the EU AI Act debate for rules on uses, not general-purpose models [M3]. In 2025 he signed a letter urging “a two-year ‘clock-stop’ on the AI Act” [M7]. In May 2026 he told the French National Assembly that European rules are heavy and overlapping [M6]. But he wants the state as buyer, energy planner and sovereign: “L’armée est souveraine, elle a une légitimité démocratique qu’on n’a pas” (“The army is sovereign; it has a democratic legitimacy we lack”) [M6].
• a16z wants federal pre-emption. Congress would lead “in regulating model development”, and states would police “harmful uses” [A5]. On liability it asks that judges consider “good-faith measures and safeguards” [A5]. Leading the Future, founded August 2025 and backed by US$50m from Andreessen and Ben Horowitz in 2026, spent about US$8m against Alex Bores, the New York legislator behind the state’s RAISE Act [A12].

Compared with Huang. - Huang: “we have lots of laws and regulations. Apply it” [42:21]; “I’m against currently the distraction” [47:10]; sector rules where gaps appear [01:19:12]; third-party auditors welcome [51:20]. - Andreessen agrees in substance and adds an electoral machine. Mensch agrees that rules at the model level are wrong. - [Analysis] Huang’s rule for the labs, “When you’re asking for regulation, don’t ask for relief of the current ones” [44:17], applies equally to Mensch’s call to suspend an enacted law. - [Analysis] a16z’s “good-faith measures” is no safe harbour. But it leans the way Huang criticises in the labs, whose September documents sought only a narrow antitrust waiver, and I found no comment from him on it. - Liang works under an ex ante state gate of the kind Huang and Andreessen reject for the US. It screens content and politics, not the capability risks under debate.

2.6 Pacing, pausing and coordination#

• Liang does not address pacing. His 2026 remarks point the other way:
• “We believe in scaling. Bigger is always better. What stops us from scaling is compute, not desire” [L2].
• “However much you can buy, buy that much” [L3].
• The restraint he professes is about margins, not capability [L3].
• Beijing. A Foreign Ministry spokesperson, Guo Jiakun, said in September 2026: “Fearmongering, confrontation and vicious competition will only disrupt the process of global AI governance and serve the interests of no one” [C5].
• Mensch frames the pacing debate as “une confrontation entre deux avenirs : celui d’une IA centralisée fournie par un oligopole… et celui d’une IA décentralisée qui crée de la valeur pour tout le monde” (“two futures: centralised AI supplied by an oligopoly, and decentralised AI that creates value for everyone”) [M5].
• Andreessen’s “a form of murder” [A2] is the strongest anti-pacing line of the four; his “ceiling” claim, if true, would also weaken the case for pacing.

Compared with Huang. - Huang rejects coordinated pacing (“Nobody’s putting the pressure on them”) but votes for unilateral restraint, “Don’t ship the product” [51:20], and accepts a conditional shutdown [36:44]. He concedes the most of the four; I found no stop condition from Mensch or Andreessen. - [Analysis] Liang’s transcript gives evidence to both sides of the pacing debate. - It supports the pacing premise that rivals will not wait: a compute-limited lab says it would scale as far as it can buy. - It supports Huang’s premise that US restraint would not stop China. Liang says the US lead “is cyclical and will not last” [L3].

2.7 Open versus closed models#

• Liang holds the strongest commitment.
• “In the face of disruptive technologies, moats created by closed source are temporary” (2024) [L1].
• “Even our strongest model will probably also be open-sourced, because I don’t see what’s good about closed source… I only worry they can’t [deploy it]” (2026) [L4].
• Mensch treats openness as decentralisation and sovereignty.
• Mistral has released both open-weight and proprietary models since 2024 (from memory).
• It signed the July 2026 open-weights letter [N1] and joined Nvidia’s Open Secure AI Alliance. The alliance holds that “Blanket restrictions on open frontier AI systems would weaken defensive capacity and risk concentrating power, dependence and vulnerability in a few closed providers” [N2].
• a16z warned in April 2026 that “80% are using Chinese open source tools” among open-source developers, and that the US “will lose long-run influence over the infrastructure and the values that shape AI systems” [A4].
• The letter they all signed concedes that “Once released, the weights are beyond the original developer’s control”, while arguing that “openness may be one of the most important paths to AI safety and security” [N1].

Compared with Huang. - Huang: “the world needs closed and open models”; “I need open weights”; and “open is the most safe and secure” [27:02]. - Agreement in substance is near complete. Each gives a different reason: customer control (Huang), sovereignty (Mensch), competition and national influence (a16z), culture and ecosystem (Liang). - One divergence of substance concerns Chinese models. Huang calls using their weights “terrific… We make it our own” [01:33:51]. a16z treats Chinese open-source leadership as a strategic loss. - The July incident supports both sides. Hugging Face analysed it with Zhipu’s open-weight GLM 5.2 after closed tools “blocked essential forensic analysis” [N2], while CAISI and Neo Research document the costs of the same openness [C3, C4].

2.8 China, export controls and the race#

• Liang:
• In 2024: “Money has never been the problem for us; bans on shipments of advanced chips are the problem” [L1].
• In the 2026 transcript: “All the differences we see, including talent, model capability, and applications, can be attributed to differences in compute resources” [L2].
• DeepSeek works with “one-twentieth of the U.S. compute”. He puts the gap variously at “6–12 months” and “about two years” [L4].
• “Huawei’s 950 super-node can fully replace Nvidia’s GB200 and GB300 in performance and price” [L2]; he wagers that “within one year” deployment will prove the domestic ecosystem viable [L2].
• Mensch positions Europe between the powers: “Dans un monde où vous importez la totalité de vos services numériques [depuis les] Etats-Unis, vous n’avez pas de levier” (“If you import all your digital services from the United States, you have no leverage”) [M6]. I found nothing from him on US chip controls.
• Andreessen:
• In 2023: “We should seek to win the race to global AI technological superiority and ensure that China does not” [A1].
• On DeepSeek’s R1: “AI’s Sputnik moment… a profound gift to the world” (January 2025) [A13].
• I found no statement by him on selling Nvidia chips to China.

Compared with Huang. - Huang on the race: “I don’t think it’s necessary” [01:32:23]; both sides should “communicate, collaborate, to understand, align” [01:37:36], though he wants the world “built on the American tech stack” [01:35:15]. - Andreessen diverges in emphasis, embracing the race framing Huang disavows. - Mensch shares Huang’s sovereignty doctrine (in Nvidia’s 2024 summary, “Every country needs to own the production of their own intelligence”), which is also Mistral’s business model. - [Analysis] Liang’s unverified testimony cuts both ways on Huang’s China case. - Against Huang: it contradicts his April 2026 claim that China “already reached that threshold” of compute, since for Liang compute is “the biggest bottleneck” [L2]. - For Huang: it supports his warning that controls speed Chinese substitution. The one-year chip forecast can be checked.

2.9 Jobs and distribution#

• Liang. I found no statements by him.
• A DeepSeek researcher, Chen Deli, reportedly told a conference in November 2025 that AI could take most jobs within one to two decades, and that AI firms should act as society’s “whistle-blowers” (from memory; Reuters; not re-read).
• Liu’s essay foresees society pulled toward “communism or Cyberpunk 2077” [L6].
• Mensch (May 2026) says a rise in unemployment “in certain areas” cannot be excluded (“Il n’est pas exclu que vous ayez dans certains domaines une augmentation du chômage”). He also foresees “un déplacement de la valeur du travail vers le capital, un capital qui n’est pour le moment largement pas du capital européen” (“a shift of value from labour to capital that is largely not European”) [M6].
• Andreessen rejects the “lump of labor fallacy” [A1] and says technology “drives wages up, not down” and that “job growth can continue forever” [A2]. a16z supports reskilling and apprenticeships [A5].

Compared with Huang. - Huang distinguishes a job’s purpose from its tasks [05:55]. On weaker junior hiring he says “Wait two years” [19:50]. - Andreessen agrees in substance. Mensch diverges in substance, and is closer to Amodei. - [Analysis] Mensch’s concession also serves his sovereignty pitch, and Andreessen’s optimism suits an investor. On both sides of this split, interest and stated view align.

2.10 Energy#

• Liang. I found no statements.
• Mensch expects “une forte demande en électricité dans un monde où il n’y a pas assez d’électricité… conflits d’usage et donc de l’inflation” (“strong demand for electricity in a world without enough of it… competing uses and so inflation”) [M6]. Mistral Compute advertises “decarbonized energy” [M4].
• Andreessen calls nuclear fission “the silver bullet for virtually unlimited zero-emissions energy” and wants to “increase our energy 1,000x” [A2]. a16z also says consumers and Little Tech should not be “saddled with the costs of hyperscalers’ energy needs” [A5].

Compared with Huang. - Huang says the US “got ourselves really gummed up in climate change and sustainable energy” [01:39:53], and that “in the near term, energy production requires fossil fuel” [01:40:15]. I found neither Mensch nor Andreessen repeating his causal story. - Andreessen shares Huang’s faith in abundance but prefers nuclear. His concern for ratepayers diverges by commercial position, since hyperscalers are Nvidia’s main customers. - Mensch diverges in substance (scarcity) and by commercial position (low-carbon French power is part of what he sells).

2.11 Epistemic style and formation#

• Liang, an engineer-quant, runs DeepSeek as a research lab: long-horizon, scaling-first, compute-accounting, with rare speech and dated bets.
• Mensch is an applied mathematician trained at DeepMind. He deflates AGI talk as a scientist would and argues industrial policy as a strategist.
• Andreessen, an engineer turned financier, argues from economic history, public-choice theory and moral framing, and organises politically.

Compared with Huang. Huang argues from chip verification and regulated engineering (cars, aviation); Mensch is closest to him in method, and Andreessen reaches similar conclusions by another route. [Analysis] For anyone using Huang as a proxy for “an engineering approach”, two things follow. His policy conclusions are shared by people arguing from economics and interest, so engineering alone does not produce them. And Liang, the purest engineer-researcher of the four, states no release gate at all: an engineering mindset does not by itself yield “don’t ship”.

2.12 Commercial position#

• Liang. Open release plus an API priced to recover hardware costs in about ten months (“a reasonable profit”) [L3] builds ecosystem and talent. DeepSeek needs chips, from Nvidia or Huawei.
• Mensch. Mistral’s edge is openness and sovereignty. A pacing regime led by US incumbents, or curbs on open frontier models, would hurt it directly.
• Andreessen. a16z’s portfolio gains from open models, light rules on development and pre-emption [A12].

Compared with Huang. Nvidia sells to everyone, so its interest lies in diffusion, open models, China and no pacing. All four hold views that fit their interests, as critics say the pacing labs do too. Fit does not show insincerity. It matters most where a view departs from disinterested expert opinion, as in “not a single actual bad use… not already illegal” [A1] or “I don’t see any risk associated with open sourcing models” [M9].

2.13 Shifts, 2023 to September 2026#

• Liang. Stable on AGI and open source. He moved from “Money has never been the problem” (2024) to US$7bn of outside capital, and from Nvidia toward Huawei.
• Mensch. From pure open weights (2023) to a mix (from memory) to leading open-weights advocacy (2026); from AI Act lobbying to “stop the clock” to sovereignty and energy policy; from AGI scepticism to an open charge of capture; toward conceding job losses.
• Andreessen. From essays to politics (endorsing Trump in 2024, the super PAC from 2025); from “use, not development” (January 2025) to accepting federal regulation of development and national-security testing (December 2025); toward alarm at Chinese strength in open models.

Compared with Huang. Huang’s safety model has been stable since 2023 while his tone hardened, and his regulatory stance moved from Nvidia’s 2023 support for licensing high-risk uses to “We don’t need any new laws” (September 2026, as reported). Mensch hardened in step; Andreessen softened slightly in substance as his political spending grew.

3. A note on the wider Chinese ecosystem#

• Zhipu (Z.ai) signed the 2024 Seoul commitments, as did Mistral and Nvidia [C1]; its GLM 5.2 carried Hugging Face’s forensics after the July intrusion [N2]. It is on the US Entity List (from memory).
• Alibaba’s Qwen family has passed 700 million downloads [A4].
• Beijing presents open-source AI as equitable development, in contrast to US export controls [C2]. Several Chinese labs, reportedly including DeepSeek, signed domestic safety commitments in December 2024 (from memory).
• I retrieved no statements by Zhipu’s or Alibaba’s leaders.

4. How the Late Lessons lens would read these stances#

Late Lessons from Early Warnings is a pair of European Environment Agency reports (2001, 2013; LL1 and LL2 below, cited by chapter and page) on how early warnings about technologies from asbestos to BSE were heeded or ignored. A companion analysis turns them into diagnostic entries, each with a “Mirror” question aimed at warners and would-be regulators. Each entry is tagged by its supporting cases: [K] known harm left unaddressed, [U] genuinely uncertain, [F] forward warnings checked by hindsight. Entries resting on [K] cases transfer poorly to uncertain technologies. The reports are partly advocacy, say little about the costs of precaution or the interests behind restriction, and were better on direction than magnitude. Findings are recorded entry by entry, not summed.

• W1: warnings come early, from the edges and from inside (LL2-09, p. 204; LL1-05, p. 53; [K] strong, [F] moderate). Present: the sharpest 2026 warnings come from insiders, including the 1,386 lab signatories and DeepSeek engineers writing in a channel that reportedly must “pass moderation” [L6]. Mensch and Casado treat the US insiders as marketers; Liang does not engage. W1 asks whether any channel treats such reports as data. Mirror: being an insider does not validate a warning (MMR’s warners were wrong), and the insiders’ firms might gain from pacing. Transfers with modification.
• W4: knowing is not acting (mainly [K]; weak transfer). All three accept, in their own letter, that released weights are “beyond the original developer’s control” [N1]. W4 asks whether triggers for action were agreed in advance. Mistral and Zhipu signed threshold commitments, but I found no Mistral framework. DeepSeek has none, and a16z names no consequence for its tests. Huang at least states a trigger [36:44]. Mirror: inaction can be a reasoned judgement that restriction would do more harm, which the Hugging Face forensics support.
• W8: the alarm trap, with W3: the reassurance trap (W8 [U]/[F], moderate; W3 [U]/[F], strong). W8 partly vindicates all four: categorical alarms without exit conditions tend to harden (hindsight on LL2-02, saccharin and cyclamate). W3 applies equally to their own categorical reassurances (“it’s not alive” [A1]; “I don’t see any risk associated with open sourcing models” [M9]; Huang’s “0% chance”), which make later protective steps look like admissions of error (LL1-15, pp. 161–162, BSE). Apart from Huang’s conditional, neither camp has said when it would escalate or stand down.
• I5: promotion and oversight in one body (LL1-15, pp. 157–165; LL2-18, pp. 441–443; [U] strong, BSE; [F] strong, Fukushima. The entry also cites LL2-22, co-authored by Andrew Maynard, but does not depend on it). Present in three jurisdictions:
• China: Beijing both promotes and regulates open-source AI, and DeepSeek has been treated as strategic, a status Late Lessons found shifts policy toward securing supply.
• United States: an administration “completely aligned with Jensen Huang” (Treasury Secretary Bessent, September 2026) is pursuing pre-emption, and Andreessen’s super PAC works to shape it.
• France: the state bank Bpifrance invests in Mistral [M2]. Mirror: the same test applies to labs that both warn about a hazard and assess it.
• I9: whose interests does restriction serve? (LL1-14, pp. 150–154; LL2-20, p. 499; [U]/[F], moderate). This is the entry that most supports Mensch and Andreessen. It presses on a real gap, since Late Lessons never analyses interests on the side of restriction, and the antitrust-waiver request makes the capture question legitimate. Limits: an interest in restriction does not make it wrong; General Motors’ interest in removing lead from petrol coincided with a real hazard (LL2-03, p. 60). The capture charge here is inferred, not documented, and Tabarrok’s evidence weighs against it [C6]. Mirror (I7): third parties bear the harm if restriction does not come, and the absence of restriction serves these three too.
• M1: sincere belief can do serious harm without bad faith (strong across [K], [U] and [F]). All three read as sincere: Liang’s open-sourcing predates his stakes and gives value away, and Mensch’s and Andreessen’s views predate the 2026 fight. On the Biden lunch, even the officials allow that he may have “misremembered”. Late Lessons found that bad faith inferred from outcomes rarely survives hindsight. The M1 question is what the reasoning is insulated from. For open weights, it is harm to third parties after release, when recall is impossible. Mirror: warners’ beliefs can be insulated too.
• M2: the model of harm behind the confidence ([K] and [U]; strong). Each works from a different model of harm:
• Andreessen: human misuse, deterred by law after the fact.
• Mensch: product malfunction, caught by testing.
• Huang: containment failure.
• Liang: none stated.

The July incident was harm during development, before release, to a third party, and it falls outside all three stated models. None has said what evidence would change their view, and nor have most pacing advocates. [Analysis] Software differs from a pollutant because it can be patched and recalled. Open weights cannot be, so the Late Lessons concern with irreversibility applies to them most directly. - M3: commitment escalates ([K] and [U]; moderate to strong). Categorical public commitments raise the cost of later admitting a problem: “strongest model… open-sourced”; Mistral’s brand; a manifesto naming the “Precautionary Principle” as “the enemy”; a super PAC funded in the tens of millions of dollars. Limits: Mistral’s moves between open and closed releases show flexibility. Mirror: the pacing signatories, Amodei and Klein have made commitments of their own.

Overall. The entries most favourable to these three (I9, W8, M1) rest on [U] and [F] cases and transfer reasonably well. Of those that press against them, W1, I5 and M2 also have [U] or [F] support; W4 rests on [K] and counts for less. The lens does not say who is right. It shows where each position is exposed, and what would test it: - a stated trigger for releasing or withholding open weights; - a model of harm at the development stage; - documentary support for the capture charge; - Liang’s dated forecast on Chinese chips.

Sources#

Liang Wenfeng and DeepSeek - [L1] Interview with Waves (36Kr), July 2024; English translation, ChinaTalk: https://www.chinatalk.media/p/deepseek-ceo-interview-with-chinas - [L2] Poe Zhao, Hello China Tech, 23 July 2026 (transcript via Tencent Tech; “DeepSeek has not confirmed the record”): https://hellochinatech.com/p/deepseek-liang-wenfeng-transcript - [L3] Fred Gao, “DeepSeek’s Liang Wenfeng Breaks His Silence”, 23 July 2026 (authenticity unverified): https://www.fredgao.com/p/deepseeks-liang-wenfeng-breaks-his - [L4] Geopolitechs, “Liang Wenfeng in His Own Words: 64 Quotes”, 23 July 2026 (from https://mp.weixin.qq.com/s/HxvyxDGXusNRMzes3i-U7g): https://www.geopolitechs.org/p/deepseek-founder-liang-wenfeng-in - [L5] Wikipedia, “Liang Wenfeng” and “DeepSeek” (biographical and corporate facts only): https://en.wikipedia.org/wiki/Liang_Wenfeng ; https://en.wikipedia.org/wiki/DeepSeek - [L6] Shengyu Liu, “I Have No Choice but to Bury My Talent in Yesterday”, September 2026, as quoted by Zvi Mowshowitz, 17 September 2026: https://thezvi.wordpress.com/2026/09/17/ai-186-the-world-takes-notice/

Arthur Mensch and Mistral - [M1] Fortune, 12 April 2024, reporting a New York Times interview: https://fortune.com/europe/2024/04/12/millennial-boss-french-openai-challenger-tech-ceos-obsession-ai-outsmarting-humans-very-religious-fascination-agi/ - [M2] Mistral AI, Series C announcement (ASML lead; a16z, Bpifrance, Nvidia and others), 9 September 2025: https://mistral.ai/news/mistral-ai-raises-1-7-b-to-accelerate-technological-progress-with-ai - [M3] TIME, TIME100 AI profile, September 2024: https://time.com/7012696/arthur-mensch/ - [M4] Mistral AI, “Mistral Compute”, 11 June 2025: https://mistral.ai/news/mistral-compute - [M5] Le Monde interview, 24 September 2026 (paywalled): https://www.lemonde.fr/economie/article/2026/09/24/arthur-mensch-directeur-general-de-mistral-ai-dans-lia-nous-sommes-les-seuls-completement-europeens-du-calcul-au-deploiement-dans-les-entreprises_6781335_3234.html ; quoted via MacGeneration, 24 September 2026: https://www.macg.co/intelligence-artificielle/2026/09/qui-controlera-lia-la-bataille-des-geants-sinvite-lonu-311235 - [M6] Hearing, Assemblée nationale commission of inquiry, 12 May 2026. Video: http://videos.assemblee-nationale.fr/video.18888392_6a0330a9d4404.vulnerabilites-systemiques-dans-le-numerique–m-arthur-mensch-cofondateur-et-dg-de-mis-12-mai-2026 ; LCP report: https://lcp.fr/actualites/intelligence-artificielle-le-patron-de-mistral-ai-arthur-mensch-n-exclut-pas-une-hausse ; English account: https://angelo-lima.fr/en/arthur-mensch-mistral-ai-national-assembly-hearing-en/ - [M7] “Stop the Clock” letter, EU AI Champions Initiative, 2025: https://aichampions.eu/ - [M9] AI-safety tracker page on Mistral (secondary and critical; quotes Sifted Summit, October 2023, and TIME, August 2024): https://sladebyrd.com/ai-safety-orgs/mistral - [M10] French Tech Journal, 9 September 2026: https://www.frenchtechjournal.com/la-machine-91-mistrals-eur3-billion-mic-drop/

Marc Andreessen and a16z - [A1] “Why AI Will Save the World”, 6 June 2023: https://a16z.com/ai-will-save-the-world/ - [A2] “The Techno-Optimist Manifesto”, 16 October 2023: https://a16z.com/the-techno-optimist-manifesto/ - [A4] Jai Ramaswamy and Matt Perault, “Asserting American Leadership in Open Source AI”, 28 April 2026: https://a16z.com/asserting-american-leadership-in-open-source-ai/ - [A5] Ramaswamy, McCune and Perault, “A Roadmap for Federal AI Legislation”, 17 December 2025: https://a16z.com/a-roadmap-for-federal-ai-legislation-protect-people-empower-builders-win-the-future/ - [A6] Matt Perault, “Regulate AI Use, Not AI Development”, 27 January 2025: https://a16z.com/regulate-ai-use-not-ai-development/ - [A7] a16z on X, 10 April 2026 (date from the post ID; text from the search index): https://x.com/a16z/status/2042394735814562033 - [A8] TechCrunch, 14 December 2024 (Andreessen on Bari Weiss’s podcast): https://techcrunch.com/2024/12/14/why-marc-andreessen-was-very-scared-after-meeting-with-the-biden-administration-about-ai - [A9] Bruce Reed and Ben Buchanan, Politico Magazine, 10 September 2026 (read via summaries): https://www.politico.com/news/magazine/2026/09/10/marc-andreessen-trump-biden-ai-opinion-01045516 - [A10] Andreessen on X, 6 February 2025 (date from the post ID; text from the search index): https://x.com/pmarca/status/1887351272052769227 - [A11] Martin Casado, September 2026, as quoted by Zvi Mowshowitz: https://thezvi.substack.com/p/we-must-pace-the-frontier - [A12] Leading the Future: https://en.wikipedia.org/wiki/Leading_the_Future ; The Next Web, 9 July 2026: https://thenextweb.com/news/ai-companies-super-pac-election-spending-midterms - [A13] Reason, 27 January 2025: https://reason.com/2025/01/27/is-this-new-chinese-ai-a-sputnik-moment/

Nvidia-led coalitions - [N1] “Open Weights and American AI Leadership”, 24 July 2026 (signatories include Andreessen Horowitz, Mistral and NVIDIA): https://images.nvidia.com/pdf/Open-Weights-and-American-AI-Leadership.pdf - [N2] Nvidia, “Open Secure AI Alliance”, 27 July 2026 (members include Mistral): https://blogs.nvidia.com/blog/open-secure-ai-alliance/

Context - [C1] Frontier AI Safety Commitments, AI Seoul Summit 2024: https://www.gov.uk/government/publications/frontier-ai-safety-commitments-ai-seoul-summit-2024/frontier-ai-safety-commitments-ai-seoul-summit-2024 - [C2] Meinhardt, Nong, Webster, Hashimoto and Manning, “Beyond DeepSeek”, Stanford HAI and DigiChina, December 2025: https://hai.stanford.edu/assets/files/hai-digichina-issue-brief-beyond-deepseek-chinas-diverse-open-weight-ai-ecosystem-policy-implications.pdf - [C3] NIST CAISI evaluation of DeepSeek, September 2025: https://www.nist.gov/news-events/news/2025/09/caisi-evaluation-deepseek-ai-models-finds-shortcomings-and-risks - [C4] Neo Research, “Evaluating DeepSeek v4 Pro for Frontier Risks”, 29 May 2026: https://neoresearch.ai/research/deepseek-v4-pro-safety-evaluation/ - [C5] Zvi Mowshowitz, “Trump Goes Full Hoax on AI Existential Risk”, September 2026 (source for the Guo Jiakun quote): https://thezvi.substack.com/p/trump-goes-full-hoax-on-ai-existential - [C6] Alex Tabarrok, “What Regulatory Capture Actually Looks Like”, 16 September 2026: https://marginalrevolution.com/marginalrevolution/2026/09/what-regulatory-capture-actually-looks-like.html

Not retrieved: - the UN Security Council record of 23 September 2026 (https://press.un.org/en/sc/16462.doc.htm), which did not load; - the full Le Monde text; - the Reuters report on Chen Deli; - Andreessen’s own 2026 posts; - statements by Zhipu’s and Alibaba’s leaders.