AI companies are pivoting from creating gods to building products. Good.

🌈 Abstract

The article discusses the challenges and limitations that AI companies face in commercializing generative AI models, despite the significant investments being made in hardware and data centers. It examines the mistakes made by AI companies, the barriers they need to overcome, and the progress they are making to address these issues.

🙋 Q&A

[01] Mistakes Made by AI Companies

1. What were the two opposing but equally flawed approaches to commercializing LLMs (Large Language Models)?

• OpenAI and Anthropic focused on building models and not worrying about products, leading to a delay in releasing user-friendly apps.
• Google and Microsoft rushed to integrate AI into everything without thinking about which products would actually benefit from AI and how they should be integrated.

2. How did these approaches contribute to a poor public perception of the technology?

• The DIY approach of OpenAI and Anthropic meant that early adopters were more likely to be bad actors, as they were more invested in figuring out how to adapt the new technologies for their purposes.
• The AI-in-your-face approach by Microsoft and Google led to features that were occasionally useful but more often annoying, and caused unforced errors due to inadequate testing.

[02] Barriers to Commercializing Generative AI

1. What are the five limitations of LLMs that developers need to tackle to make compelling AI-based consumer products?

• Cost: Even in simple applications, cost concerns dictate how much history a bot can keep track of, as processing the entire history for every response quickly becomes prohibitively expensive.
• Reliability: AI systems need to perform tasks correctly 100% of the time, like traditional software, but achieving perfect accuracy is intrinsically difficult with statistical learning-based systems.
• Privacy: There are concerns around training AI assistants on sensitive user data, such as emails and documents, as well as concerns about the potential for misuse of personal data.
• Safety and security: Accidental failures, misuses of AI, and hacks that can leak user data or cause harm are significant challenges.
• User supervision: In many applications, the unreliability of LLMs means that there must be a way for the user to intervene if the bot goes off track, which can be challenging to implement, especially in natural language interfaces.

2. Why are the sociotechnical challenges discussed in the article likely to take a decade or more to solve, rather than a year or two?

• The challenges are not purely technical, but also involve integrating AI into existing products and workflows, and training people to use it productively while avoiding its pitfalls.
• Even if AI capability improves rapidly, developers still need to solve the challenges discussed, which are sociotechnical in nature, not just technical.