a16z on Protecting Little Tech: The Techno-Optimist AI Policy Agenda with Matt Perault
Summary
a16z’s policy call is to regulate harmful AI use while leaving research and development broadly open. Matt Perault insists this is not a deregulatory dodge: governments may need more investigators, technical expertise, enforcement resources, and targeted updates to existing law. The dividing line is that policymakers should punish discrimination, criminal conduct, and other harms without trying to “criminalize the math.”
The firm’s economics favor durable adoption, not a short-lived AI boom that collapses under accumulated harms. Perault contrasts a16z’s 10-year fund life with a four-year public-company vesting cycle: if enthusiasm explodes tomorrow but “the market cratered a year later,” the portfolio still loses. Little tech therefore needs both room to build and an ecosystem whose products people trust.
Threshold regulation could convert AI’s natural power laws into a state-reinforced oligopoly. Nathan Labenz notes that assessments of today’s frontier produce lists of only three to 10 companies, but Perault says rules designed for five, 10, or 15 developers could cement precisely that concentration. He prefers revenue thresholds where affordability is the objective; compute or $100 million training-cost thresholds measure development activity, not capacity to absorb compliance.
The unresolved disagreement is whether some AI harms would be irreversible before use-based enforcement could begin. Labenz raises lab leaks and the possibility that sufficiently capable AI development could get out of control without deployment; Perault uses a COVID hypothetical—if it came from a lab and 10-plus million people died globally—as a case where a later fine would be inadequate. Labenz later worries about closed-door AI-assisted ML research, while Perault answers that dangerous research may also produce defenses and breakthrough benefits and doubts publishing a safety plan has a strong “direct correlation” with actual product safety.
Transparency is the sharpest practical clash between frontier-risk monitoring and little-tech burden. Labenz cites the Claude 4 launch remark, “We want Claude to take over all the ML research so we can all go to the beach,” and worries the gap between laboratory and public capabilities will widen. Perault argues many proposed disclosures are speculative, unhelpful to consumers, and likely vulnerable under the First Amendment; a16z instead proposes concise “AI model facts,” including knowledge-cutoff dates.
California’s SB 813 is promising only if startups can realistically obtain its liability protection. Perault likes that the bill confronts tort liability through an opt-in, privately administered regime, but rejects the idea that any nominally voluntary system is automatically fair. His analogy: if immunity required paying $20 billion, incumbents would buy protection while startups retained the risk—“That’s not really voluntary.”
Temporary or adaptive rules still impose real competitive costs during a critical market-formation period. Erik Torenberg proposes three-to-five-year sunsets; Perault replies that even temporary stringency resembles making racers carry “a 20-pound backpack for the first mile.” He points to reported EU doubts about implementing the EU AI Act, Colorado’s governor supporting a federal moratorium affecting his state’s own law, and the Trump administration’s expected national AI action plan in June or July.
Deep dive
1. Abundance begins with access, not a technology forecast
Perault’s honest non-answer on predicting AI’s destination: “That’s something that I’ve never been particularly good at.” As a policy specialist rather than an engineer, he defines abundance operationally—build rules that let capable people create useful tools without making startup compliance costs exceed the benefits.
His most personal use case is mental-health access. With both parents and his sister working as mental-health professionals, Perault sees AI potentially reducing three barriers—cost, difficulty finding a provider, and stigma—provided deployment remains consistent with “a high quality level of care.”
Writing makes the opportunity concrete and uncomfortable. Perault once summarized his Facebook role as delivering “a bad first draft” to his boss, who replied that occasionally a good one would help; AI can already supply rough conclusions and may eventually produce very good first drafts, forcing writers to differentiate above baseline competence.
Torenberg calls his own posture “classically ambivalent”: Claude for coding and Operator for web tasks are genuinely thrilling, while deceptive or scheming behavior raises larger control questions. He worries less about jobs than gain-of-function-style research and describes himself as AI’s “Forrest Gump”—present in notable scenes, usually as an extra, trying to nudge the odds toward a radically positive outcome.
2. Ten-year venture economics make trust part of the return
Perault thinks everyday use has pulled attention from the “infinite future” toward marginal value creation: recipes, help with a child who will not sleep, family-vacation activities, structured ideas, or a draft conclusion. As AI literacy spreads through people comparing notes, the value proposition becomes closer and less frightening.
That familiarity changes the policy balance. Rules should help avoid harms, including possible long-term catastrophes, while leaving enough room for people to experience benefits; growing consumer contact with useful applications should affect which regulatory burdens look proportionate.
a16z’s 10-year fund life underpins this stance, Perault argues, versus the four-year vesting cycle he experienced at Facebook. The goal is not to “juice stock prices tomorrow”: an AI boom followed by a harm-driven crash one year later would frustrate the firm’s effort to build healthy, durable companies.
3. a16z wants harmful uses policed while development stays open
The core analogy is conventional software: regulators generally do not supervise its construction, but laws can punish software used harmfully. Perault wants the same separation for AI—leave development available to little tech, then deter unlawful outcomes rather than treating the research process itself as good or bad.
His Cambridge Analytica example separates legitimate access from misuse. Facebook’s initial data transfer to an academic researcher was permitted under its terms and reflected the generally desirable goal of research access; the problem came when data under the researcher’s control was abused.
Mapped onto frontier AI, Perault’s concern is that restricting development means making it harder for scientists to do science: “The development itself is just science. It’s math.” Research with dangerous potential can also generate defenses, medical access, better mental-health treatment, and tools that let more developers build valuable products.
“Regulate harmful use” requires active state capacity, not passivity. Drawing on a summer in the Justice Department’s Civil Rights Division, Perault stresses that cases arrive without tidy fact patterns; enforcing AI-related anti-discrimination law will require investigators who can identify the harm, connect it to a tool, understand the technology, and recognize where existing statutes need adjustment.
4. Catastrophic harms are where ex-post enforcement strains
The RAISE Act represents the compromise Torenberg wants examined: above a threshold intended to capture high-single-digit or low-double-digit numbers of companies, developers create and follow their own safety plans, disclose them to regulators or the public, and face some mix of audits, self-reporting, and incident disclosure—effectively “grading their own homework,” but under scrutiny.
Labenz’s pushback is that a sufficiently capable system—or a frontier research accident, by analogy to lab leaks—might create an irreversible harm without being publicly deployed. Perault later offers the COVID hypothetical: if it came from a lab and 10-plus million people died globally, a later fine would be inadequate; Torenberg adds that society might target narrow “choke points” rather than every researcher.
Perault accepts that safety plans might prevent some incidents but questions their efficacy: publishing protocols does not ensure safe performance and may have “no direct correlation” with it. If benefits are limited, the policy merely burdens development; he would instead invest in detecting and prosecuting harmful uses.
5. Thresholds can turn scaling economics into regulated oligopoly
For Perault, saying a rule reaches only five, 10, or 15 companies is not reassuring—it describes “the thing we fear most.” A healthy frontier should remain open to a founder with five or 10 engineers, rather than converting today’s leaders into the only organizations legally and financially able to advance capabilities.
Labenz gives the empirical countercase: his recurring “live players” exercise produces frontier lists as short as three and never longer than 10. Scaling laws, capital inputs, and power-law outcomes may already make concentration structural, so escalating thresholds could arguably follow the market rather than cause its shape.
If lawmakers truly mean to target companies able to afford compliance, Perault says revenue is the cleaner measure. Proposals may use $100 billion or other levels, but a company earning $500 billion annually plainly can devote some percentage to regulatory work; compute and training expenditures instead regulate the science.
Perault says an a16z technical colleague told him “every model costs $100 million to build,” undermining that figure as a little-tech carve-out. He also worries the RAISE Act could aggregate training costs, eventually capturing a startup building a large number of $25 billion models. Compute and data are already entry barriers; adding regulation risks a telecommunications-style “regulated monopoly” with weaker innovation.
6. Useful transparency is narrower than frontier-watchers want
Labenz fears today’s small gap between public and laboratory capabilities may widen as companies race for supremacy. Citing Daniel Cocotello’s AI 2027 scenario, he expects frontier developers could withhold systems that dominate competitors, leaving outsiders unable to assess what exists behind closed doors.
The Claude 4 launch supplied his memorable warning: someone from Anthropic said, “We want Claude to take over all the ML research so we can all go to the beach.” Labenz asks whether OpenAI, Anthropic, and Google should at least disclose the capabilities of systems in their labs pursuing recursive coding or ML-improvement loops.
Perault’s answer is substantially narrower. a16z believes many proposed disclosures—especially forecasts of long-run harm or speculative safety assessments—are neither factual nor useful to consumers and are probably unlawful; in his account, controversial, nonfactual, or burdensome compelled speech will struggle under First Amendment scrutiny.
The alternative is an “AI model facts” sheet containing practical information without excessive startup burden. A knowledge-cutoff date tells users why asking who won last night’s basketball game falls outside the training corpus—and why any confident answer may be a hallucination rather than information to trust.
7. SB 813 tests whether optional regulation can stay optional
Torenberg proposes regulatory reversibility for an unusually uncertain period: impose measures for three to five years, with a three-year sunset, or let companies enter and leave regimes. Under SB 813, a developer could accept an approved, privately administered framework in exchange for a liability shield, then potentially opt out later.
Perault likes policy experimentation but rejects the idea that temporary burdens are free: making an industry race with “a 20-pound backpack for the first mile” can alter the result permanently. He also notes—while qualifying that he does not know the report’s accuracy—that the EU was considering pausing EU AI Act implementation.
SB 813’s engagement with tort law is nonetheless promising because even mundane AI adoption may depend on manageable liability. The decisive question is whether little tech can comply: calling the framework voluntary misses the enormous value of immunity and the possibility that incumbents alone can afford it.
Perault’s $20 billion hypothetical makes the distributional problem explicit: large companies would purchase immunity while small ones bore legal exposure. He could move toward support if the regime becomes workable for startups. Beyond liability, he favors state and federal work on criminal misuse, workforce development, AI literacy, and model facts, while watching for the Trump administration’s national AI action plan in June or July.