Interview, Fireside Chat
Aravind Srinivas:Will Foundation Models Commoditise & Diminishing Returns in Model Performance|E1161
- Future models will evolve through a self-improving cycle of output, reasoning, feedback elicitation, and iteration until convergence, a process termed "bootstrap reasoning" that may take one to four years to fully materialize.
- AI reasoning capabilities are projected to reach the level of a median college undergraduate on a pathway to eventually advising top-tier experts like Demis Hassabis, potentially breaking the $20/month subscription model to enable high-value single-session payments.
- Significant capital requirements for inference compute during self-improvement will advantage large players, with only four or five contenders likely pursuing this path, potentially consolidating the foundation model layer to one or a few winners such as OpenAI, Anthropic, or XAI.
- Long context windows will expand from 128k tokens to one or two million tokens, with paid throttling potentially introduced after specific limits, while instruction following improves to fix hallucinations.
- GPT-4 quality models will eventually become commoditized, though the next iteration may prioritize reliability, speed, and cost over a capability leap comparable to GPT-4's advancement over GPT-3.5, leaving only two or three viable alternatives initially.
- Large cloud providers are expected to acquire mid-tier models like Anthropic or Cohere within the next three to five years to complement core services, whereas OpenAI and Anthropic are unlikely to be acquired due to the difficulty of replicating their tacit team knowledge.
- Perplexity aims to build a business with 50 to 100 million paying users by 2034 to support an advertising model comparable to Netflix or YouTube, though this depends on cracking the relevance code, with consumer subscriptions or enterprise sales serving as the alternative primary engine.
- Perplexity's enterprise division will focus on compliance, security, data governance, and unified UIs, while the company expects to distinguish itself from ChatGPT within two years through superior browsing orchestration and UX innovation.
- The AI industry's impact is considered underhyped, with expectations of deep integration into familiar workflows and the emergence of an "AI native OS" organized around agent interaction rather than traditional file structures.
- Application layer companies are positioned as the biggest beneficiaries of model commoditization by packaging these tools into premium experiences, while OpenAI and Anthropic face a critical risk of losing leverage if they fail to deliver a new breakthrough model within approximately one year.
- Perplexity's primary internal risks include indecisive leadership, poor execution, lack of focus, and inefficient capital use, which could prevent it from achieving its status as an essential facts and knowledge assistant by 2034.