Claude: While Li raises some valid concerns about potential impacts on innovation, her article appears to mischaracterize several aspects of the bill [proposed AI regulation SB-1047].

Asking AI services: Please read the letter from Gary Marcus to Fei-Fei Li and compare it to Fei-Fei Li’s article in Fortune and answer this really simple question: Who is right and why? Don’t be a diplomat and argue they both make interesting points. Make a clear choice and provide evidence for your position.

Claude’s answer:
Based on the information provided in both documents, I believe Gary Marcus makes a stronger and more accurate case regarding California's proposed AI regulation SB-1047. Here's why:

1. Accuracy of claims: Marcus points out several inaccuracies in Li's Fortune article:

   • Li claims SB-1047 holds developers liable for any misuse, but Marcus clarifies it only focuses on specific "critical harms" like mass casualties and large-scale cyberattacks.

   • Li argues the bill could harm open-source AI development due to "kill switch" requirements, but Marcus notes this doesn't apply to open-source models once out of the original developer's control.

   • Li states the bill will hurt "little tech," but Marcus highlights that many requirements only apply to models with $100 million+ training runs, which aren't "little tech."

2. Addressing concerns: Marcus directly addresses Li's concerns and provides context that Li's article lacks. For example, he explains that the bill only requires "reasonable safeguards" against severe risks, not full mitigation.

3. Industry context: Marcus points out that asking for standards and care is common across many industries, like pharmaceuticals and aviation. Li's article doesn't acknowledge this broader context.

4. Concrete alternatives: Marcus notes that Li doesn't offer any concrete suggestions for addressing the serious risks the bill aims to mitigate, whereas SB-1047 provides a specific approach.

5. Balanced view: While Marcus disagrees with Li, he acknowledges her good faith and intentions. In contrast, Li's article presents a more one-sided view without addressing potential benefits of the regulation.

6. Specificity: Marcus's letter provides more specific details about the bill's contents and requirements, while Li's article uses broader, more generalized statements.

While Li raises some valid concerns about potential impacts on innovation, her article appears to mischaracterize several aspects of the bill. Marcus's response provides a more accurate and nuanced understanding of SB-1047's actual contents and potential impacts. This makes his argument more convincing and well-supported by the evidence presented in both documents.

ChatGPT-4o’s answer:
Between Gary Marcus and Fei-Fei Li, Gary Marcus is more persuasive and aligned with the realities of AI governance, and here's why:

1. Accuracy in Representing SB-1047: Fei-Fei Li argues that SB-1047 would "unduly punish developers" and stifle innovation by holding developers liable for all possible misuses of their models. However, Gary Marcus clarifies that the bill specifically targets critical harms, such as mass casualties or large-scale cyberattacks, and requires only "reasonable safeguards" from developers. His explanation suggests that the bill is more narrowly tailored than Fei-Fei Li portrays, focusing on severe risks that should be mitigated.

2. Misrepresentation of the "Kill Switch" Requirement: Li expresses concern that the bill's "kill switch" mandate would stifle open-source development. Marcus counters this by explaining that the kill switch requirement does not apply to open-source models once they are outside the original developer's control. This clarification indicates that Li's concerns may be based on an incomplete understanding of the bill’s provisions.

3. Impact on "Little Tech" and Academia: Fei-Fei Li argues that SB-1047 would disadvantage smaller tech companies and academic research, putting them at a further disadvantage to tech giants. Marcus challenges this by noting that the bill applies primarily to models with substantial training costs ($100 million+), which are typically within the domain of large, well-funded companies. This implies that the bill is less likely to impact the smaller players that Li is concerned about.

4. Lack of Alternatives in Li’s Argument: While Fei-Fei Li critiques SB-1047, she does not offer concrete alternative solutions for mitigating the severe risks that AI poses, such as those Marcus mentions. In contrast, Marcus acknowledges the need for regulation in areas with potential catastrophic risks and supports SB-1047 as a step in the right direction, albeit open to alternative suggestions.

In summary, Gary Marcus provides a more precise and well-argued position, addressing Fei-Fei Li's concerns with clarifications that suggest her critique may be based on misunderstandings of the bill's scope and intent. His defense of SB-1047 as a necessary measure to prevent severe AI-related harms is more compelling, especially in the absence of viable alternative solutions from Li.


Perplexity’s answer:
In comparing Gary Marcus's letter to Fei-Fei Li's article, Marcus's argument appears to be more compelling based on the details provided.

Fei-Fei Li argues that California's proposed AI regulation, SB-1047, will stifle innovation, harm open-source development, and disadvantage academia and small tech companies. She contends that the bill's requirements, such as the "kill switch," will negatively impact developers and that the legislation does not adequately address the potential harms of AI, such as bias and deepfakes.

Gary Marcus, on the other hand, counters these points by clarifying that SB-1047 focuses on preventing severe risks like mass casualties and cyberattacks, requiring only "reasonable safeguards" rather than full mitigation. He also notes that the "kill switch" requirement does not apply to open-source models once they are beyond the original developer's control. Marcus argues that the bill's requirements are aimed at large-scale models with significant resources, not small tech companies, and he criticizes Li for not offering alternative solutions to the risks posed by AI.

Marcus's argument is more persuasive because he addresses specific misconceptions about the bill and emphasizes the importance of having some form of regulation to prevent catastrophic risks associated with AI. He also highlights that many companies have already committed to similar safeguards voluntarily, suggesting that the bill's requirements are not overly burdensome. Additionally, Marcus points out the lack of concrete alternative proposals from Li, which weakens her position against the bill.

Source: https://www.ft.com/content/0b210299-4659-4055-8d81-5a493e85432f