Baptists and Bootleggers: The AI Safety Debate

Transcript

SPEAKER_1: Alright, so last time we established that Big Tech is actively lobbying for regulations it knows will crush smaller rivals — that regulatory capture framing was sharp. And now the 20VC episode from March 29th takes that a step further with this 'Baptists and Bootleggers' idea. Where does that even come from? SPEAKER_2: It's a theorem from economist Bruce Yandle, originally about US Prohibition. Baptists wanted alcohol banned for moral reasons. Bootleggers wanted it banned because it eliminated legal competition. Same regulation, completely different motives — and crucially, they never had to coordinate. The Baptists gave the Bootleggers political cover they couldn't buy. SPEAKER_1: And the episode maps that directly onto AI regulation? SPEAKER_2: Precisely. The Baptists in AI are genuine safety advocates — people like Eliezer Yudkowsky warning that misaligned superintelligent AI could lead to human extinction. That's a sincere belief. The Bootleggers are companies like Google, Anthropic, OpenAI — firms that signed the March 2023 open letter calling for a pause on giant AI experiments, while simultaneously racing to build giant AI experiments. SPEAKER_1: So what our listener might be wondering is — how do you tell them apart? Both groups say the same things publicly. SPEAKER_2: Follow the incentives. Anthropic closed a four-billion-dollar deal with Amazon in September 2024. OpenAI converted to for-profit in May 2024. These are not the moves of organizations prioritizing safety over scale. And then look at what regulations they actually lobby for — the EU AI Act imposes compliance costs that a startup with twelve engineers simply cannot absorb, but Google's legal team handles in a quarter. SPEAKER_1: Right, and we saw that in the last lecture — seed funding in regulated sectors collapsed 45% after the EU AI Act. So the mechanism is compliance cost as a barrier to entry. SPEAKER_2: Exactly. And it compounds. Top ten firms now hold 80% of AI compute as of January 2026, per Epoch AI data. When you layer safety mandates on top of that concentration, you're not leveling the playing field — you're cementing it. The US Executive Order on AI from October 2023 mandated safety testing that burdens smaller players disproportionately. California's SB-1047, vetoed in September 2024, would have done the same thing under a transparency banner. SPEAKER_1: The watermarking angle is one I hadn't considered. Google lobbying for watermarking mandates through the C2PA updates in February 2026 — how does that become a Bootlegger move? SPEAKER_2: If you control the watermarking standard, you control the content authentication market. It sounds like consumer protection. It functions like a toll booth that only you built. That's the episode's core critique of the five-hundred-billion-dollar AI market projection for 2026 — the fight isn't just over who builds the best model, it's over who writes the rules that determine which models are allowed to exist. SPEAKER_1: So why does Andreessen keep coming back to open-source as the answer here? SPEAKER_2: Because open-source breaks the Bootlegger coalition. Meta's Llama 3.1, released July 2024, is a direct challenge — anyone can run it, modify it, deploy it without asking permission. The episode notes that safety-based crackdowns on open-weight models intensified almost immediately after Llama's release. That's not coincidence. Open-source distributes the compute and the capability, which is exactly what the Bootleggers are trying to prevent. SPEAKER_1: There was something striking in the episode about an internal OpenAI memo from February 2026 linking safety language to pricing power. What was that about? SPEAKER_2: The memo, as disclosed by a guest, showed internal framing where safety requirements were explicitly connected to justifying premium pricing tiers. So 'safety' becomes a product differentiator and a regulatory moat simultaneously. That's the cynical version of the Baptist-Bootlegger alliance playing out inside a single organization. SPEAKER_1: And the UN compute governance proposals from November 2025 — the Baptists pushing those are presumably sincere? SPEAKER_2: Many are. The Future of Life Institute types genuinely believe unchecked compute leads to rogue AI development. But the episode's point is that sincerity doesn't neutralize the effect. When Baptists push compute governance and Bootleggers benefit from it, the outcome is the same regardless of intent — power concentrates, competition shrinks. SPEAKER_1: So the question becomes... is a regulated monopoly in AI actually safer than the risks of AI itself? SPEAKER_2: That's the sharpest question in the episode. The argument is no — a small number of companies controlling AI development with regulatory protection is a single point of failure, both for innovation and for safety. Distributed development, including open-source, creates redundancy and accountability through competition. The episode predicted the regulation coalition fractures by Q3 2026, specifically over open-weight models, because the Bootleggers will eventually turn on each other. SPEAKER_1: That fracture prediction is interesting. What drives it? SPEAKER_2: Diverging interests. Google wants watermarking standards. OpenAI wants compute licensing. Anthropic wants safety certification regimes. These aren't the same regulation — they just temporarily aligned against open-source. Once open-source is sufficiently constrained, the Bootleggers compete against each other, and the Baptist cover evaporates. SPEAKER_1: So for Sergey and everyone following this course, what's the frame they should carry into the next lecture? SPEAKER_2: The push for AI regulation is rarely just about safety and rarely just about greed — it's usually both, running in parallel. The Baptist-Bootlegger lens is the tool for separating them. When someone hears a safety argument for a new AI rule, the first question should be: who benefits competitively if this passes? That question cuts through the rhetoric faster than any technical analysis. The real risk isn't unaligned AI — it's aligned incumbents using safety language to align regulation with their own market position.