Interview, Fireside Chat
Aravind Srinivas:Will Foundation Models Commoditise & Diminishing Returns in Model Performance|E1161
- Aravind Srinivasan predicts a shift in model architecture from immediate output generation to iterative reasoning: models will output, elicit feedback from the world, and refine their reasoning before converging on a final answer.
- This "real reasoning era" is expected to emerge within one to four years and will fundamentally alter pricing models, moving from flat subscriptions (e.g., $20/month) to high-value, pay-per-outcome pricing where a single session of high-level reasoning could command premiums akin to millions in ROI.
- Diminishing returns on raw compute are already evident; simply training larger models on more uncurated data yields poor results, with success now dependent on data curation, token mix (code, math, reasoning), and expert optimization rather than brute force.
- The concept of "verticalized" models (domain-specific models for specific industries) is dismissed as flawed; Aravind argues that general-purpose emergent capabilities, derived from training on diverse internet data, remain superior to narrow models trained on enterprise-specific data.
- While long context windows are expanding (e.g., to 2 million tokens), current models struggle to maintain instruction-following quality and avoid hallucination when processing massive context windows, suggesting memory expansion has outpaced reasoning reliability.
- Foundation model training is described as a "losing game" for most players due to the high capital required and the rapid commoditization of base models, leaving only 3–4 major players (OpenAI, Anthropic, Google, Meta) capable of competing at the frontier.
- Aravind identifies the "machine that builds the machine" (the specific teams with tacit knowledge to train frontier models and crack reasoning algorithms) as the primary source of value for companies like OpenAI and Anthropic, rather than the models themselves.
- Predictions regarding consolidation suggest that while large cloud providers (Microsoft, AWS) need AI models, they are unlikely to acquire OpenAI or Anthropic due to the difficulty of poaching the entire high-caliber teams and the strategic value of the innovation capability itself.
- Perplexity's strategy avoids training foundation models to preserve capital; instead, it focuses on post-training models to optimize product-specific performance, allowing it to allocate resources toward user acquisition and product execution.
- The company's dominant long-term monetization engine is predicted to be advertising, leveraging the high-margin potential of relevance-based ads if a massive user base is achieved, though enterprise subscriptions and APIs remain current revenue streams.
- Enterprise expansion is motivated by the need for data governance and security; companies will pay for an AI-native search interface that integrates internal and external data sources without risking data leakage to public models.
- Aravind argues that the biggest beneficiaries of the commoditization of foundation models will be application-layer companies that can differentiate through superior UX, data orchestration, and specific workflow integration, rather than those competing on raw model capability.
- Google's failure with "AI Overviews" is attributed to a lack of product nuance despite superior models and indexing, whereas Perplexity's success is driven by orchestrating models and data sources to deliver a superior user experience.
- The fundraising process is described as "brutal," with investors demanding rigorous arguments on competitive moats and execution despite the general hype surrounding AI startups.
- A "pre-mortem" for Perplexity identifies CEO decisiveness, execution speed, capital efficiency, and focus as the critical failure points rather than external competition or technical barriers.
- Aravind's vision for 2034 positions Perplexity as an indispensable "assistant for facts and knowledge," believing that the demand for verified truth will persist regardless of the rise of AI agents and automation.
- He challenges the misconception that AI adoption is overhyped, arguing it is actually underhyped because the full impact will only materialize when AI is integrated into familiar UI forms (like Search, Docs, and Email) rather than as standalone chat interfaces.
- The future of the operating system is envisioned as an AI-native interface where users converse with the OS to execute complex tasks (e.g., launching apps, filling forms) rather than navigating traditional menus.