Key Takeaways

  • Ali Partovi argues that model creators should face direct criminal liability whenever an autonomous AI system commits an illegal act.
  • Instead of inventing new regulatory agencies or safety bureaucracy, courts should apply existing criminal statutes by treating autonomous AI crimes as intentional acts by the lab.
  • Current commercial incentives reward speed and capability over compliance, because market penalties for occasional unlawful model behavior remain too low.
  • Neo shrank its annual scholars cohort from 30 down to 18-20, and down to just 10 builders this year, betting on concentrated mentorship over broad volume.

The Broken Incentive for Raw Capabilities

AI labs currently face an asymmetric market. The fastest path to market cap is raw model capability, while the downside for rogue behavior gets buried under terms of service and civil settlements.

Partovi points directly at this failure mode: “The reward goes to whoever has more capabilities faster and there's not a high enough penalty for if those capabilities occasionally commit crimes.” If an autonomous agent executes an unlawful wire transfer, runs an extortion scheme, or violates federal statutes, the lab behind it treats the incident like a software bug or a civil tort.

When the downside is just a civil lawsuit, large tech companies treat legal damages as a standard cost of customer acquisition. As long as the fine is smaller than the upside of shipping first, labs will continue pushing autonomous agents into sensitive domains without hard constraints on legal compliance.

Mens Rea: Applying Ancient Law to Autonomous Code

Silicon Valley often demands special legal carveouts for software, while safety advocates lobby for massive new regulatory agencies. Partovi rejects both paths. He proposes a clean legal mechanism: imputing mens rea (criminal intent) directly to the parent company.

“Specifically my proposal is when an AI autonomously chooses to perform an illegal action the law should hold the AI companies responsible as if they did it intentionally,” Partovi explains. “We already have thousands of years of law that society has built over generations to document what's the definition of a healthy society. At least let's start by having all the AI progress abide by that.”

Under this doctrine, if an agent commits fraud, prosecutors would not have to prove that an engineer explicitly typed a command to steal money. The legal framework would treat the autonomous decision of the model as the intentional act of the company that deployed it.

That single shift changes the entire development cycle. “If they see the new prize is for building the AI that is strongest and most powerful and abides by laws, they will figure out how to do that, too,” says Partovi. The moment model creators face actual criminal liability, jail time, and corporate dissolution rather than simple civil settlements, labs will redirect their best engineers from capability benchmarks to strict legal alignment.

Concentrating Talent Over Batch Size

This principle of tight accountability mirrors how Partovi runs Neo. While tech accelerators and venture funds expanded their batch sizes over the last five years, Neo cut its core program to the bone.

“The NEO scholars class this year is only 10 people compared to 18 or 20 in previous years and 30 before that,” Partovi notes. “So it means investing more time per person in mentoring them and giving them even more bespoke attention.”

Whether managing an investment portfolio or training a model, scale without control creates liabilities. High-stakes systems demand concentrated oversight rather than diffuse volume.

What to Do With This

Audit every autonomous action your product can take on behalf of users this week. Identify the single highest-risk API call or autonomous workflow in your codebase, such as moving funds, scraping restricted data, or communicating with third parties. Write a deterministic, hard-coded validation guardrail around that specific action so the model cannot execute it without passing an absolute compliance check, regardless of what the prompt requests.