Aravind Srinivas on Reasoning, Model Economics and Building Perplexity

Perplexity co-founder and CEO Aravind Srinivas tells Harry Stebbings that future models may improve by producing an answer, seeking feedback and reasoning again. He also believes application companies can benefit when foundation models become cheaper. Both claims are forecasts made in 2024 and reflect Perplexity’s commercial interests.

Srinivas entered machine learning almost accidentally. As an undergraduate needing an internship or prize money, he entered a prediction contest without knowing the field. He explored the scikit-learn library, tried approaches he did not yet understand and won, giving him confidence to study the subject.

He recalls advice from Sam Altman: a possible indicator of natural ability is something that feels easy to you but difficult to others. Srinivas studied machine learning more deeply, bought Christopher Bishop’s textbook second-hand and found a professor who introduced him to reinforcement learning.

Reinforcement learning gave him a framework of agents, environments, rewards and objectives. His early research project involved transfer learning between Atari games, inspired by DeepMind’s work. He remembers reading papers in the lab and borrowing gaming GPUs to train neural networks.

On diminishing returns, Srinivas rejects both a simple plateau story and the idea that more compute automatically wins. Larger models and more training tokens can still help, he says, but data curation, the mixture of languages, code and mathematics, and efficient architecture all matter.

He argues that a much larger model can perform worse than a smaller one if the details are handled poorly. His comparison of models is based on observations and claims current to the interview, rather than a timeless or independently established ranking.

Stebbings asks whether better data curation implies a future of specialized models. Srinivas is skeptical that domain specialization alone wins: he cites BloombergGPT and says GPT-4 performed better on finance benchmarks. Stebbings notes that one example cannot rule out a better specialist model.

Srinivas attributes much of large models’ usefulness to general capabilities emerging from training on diverse data, followed by instruction tuning. He cautions that researchers do not yet fully understand why models reason as well as they do or which kinds of data contribute most.

How good a model is at reasoning depends on the benchmark, he says. He compares then-current systems with different levels of human schooling, while saying they were not at the level of the strongest scientific thinkers. These are conversational judgments, not a controlled comparison.

If AI could give uniquely valuable advice or solve a difficult commercial problem, Srinivas argues that users might pay far more for a single result than a standard monthly subscription. He uses the example of identifying future winners for an investor, but presents such capability as hypothetical.

He says the method for a major reasoning breakthrough was not yet known. He mentions press reports about OpenAI research and published work on self-taught reasoning, but does not claim independent knowledge of a confidential project. The general idea is to use explanations, feedback and iterative training to improve outputs.

His hoped-for future system would produce an answer, check it, revise its reasoning and iterate. He does not know when that will be achieved. He also warns that inference and experimentation can consume substantial compute, potentially favoring well-funded organizations.

For memory, he distinguishes a practically large context window from an assistant that remembers every detail of a life indefinitely. He thinks expanding context and selecting relevant information can help, but says long inputs can confuse models or weaken instruction following. He does not know how to achieve truly unlimited memory.

As of the interview, Srinivas sees less capable models as becoming commoditized, while judging the strongest GPT-4-class alternatives still relatively scarce. He predicts those capabilities may become cheaper and more widely available, while leaving open whether a new frontier model will create another large gap.

He distinguishes a product from the model underneath it. Perplexity post-trains available models for its search product, he says, but does not train a frontier base model from scratch. He describes the economics of repeatedly funding large training runs and then being overtaken as especially difficult.

In his view, models below the frontier can become commodities while a few leading labs continue competing on capability. He speculates that a lab which first solves effective self-improving reasoning could gain a large lead, depending on how aggressively it invests. The mechanism and outcome remain uncertain.

He identifies OpenAI and Anthropic as plausible leaders, praising one for speed and capital and the other for research efficiency. These are his comparative opinions about companies and models at the time, not verified rankings or a prediction guaranteed to hold.

Stebbings predicts cloud providers will acquire smaller model companies. Srinivas disagrees about OpenAI and Anthropic, arguing their value includes the teams, tacit knowledge and compute infrastructure that repeatedly produce new models. If breakthroughs stop, however, he says their negotiating leverage could weaken.

Large incumbents have substantial cash, but Srinivas argues money alone does not easily assemble an entire research team: talented people also want to work with one another. He thinks a model provider ultimately needs either a sustainable business or continuing research advantages.

He says OpenAI had built meaningful revenue as well as models, citing an approximate contemporary figure, though not independently verifying it. Perplexity’s own consumer subscription, in his view, needed much greater scale and retention to become exceptionally attractive because model-serving costs limit margins.

Srinivas sees advertising as another potential revenue source. He says the team wanted to explore high-margin advertising while keeping sponsored content from influencing answers or citations. This is Perplexity’s proposed design objective, not proof that an advertising system cannot create conflicts.

He suggests sponsored placements could also appear in a discovery feed, as long as relevance is good. He explicitly says Perplexity had not yet solved ad relevance and that a large user base would be needed to make the system work well.

His aim is to avoid reliance on a single revenue source by combining subscriptions, advertising, APIs and enterprise services. He argues that Google’s advertising incentives can diverge from user interests, while hoping a diversified Perplexity can better align them. This comparison reflects his position as a competitor.

Perplexity Enterprise Pro began with a simple idea: people already search at work but may face restrictions on AI tools because of security and data-governance concerns. The company offered its consumer experience with enterprise controls and planned more features tailored to organizations.

Srinivas imagines a single interface for internal and external sources, multiple models and organized knowledge. He says Perplexity will try to build such an offering but does not claim it will succeed. The product vision and enterprise plans are specific to the time of the interview.

Stebbings questions whether building an enterprise sales motion will be difficult. Srinivas says organizations were still experimenting with AI tools and had relatively little lock-in, but accepts that a bigger brand and sales team have an advantage when products are not differentiated.

He argues an application is valuable when it combines models, data sources, search engineering and user experience in ways that are difficult to reproduce. He claims some users preferred Perplexity’s browsing to alternatives; the interview does not provide a controlled comparison.

If foundation models become commodities, application providers that own customer relationships can buy cheaper inputs and package them into useful products, he says. If models remain differentiated, Perplexity must still find a way to benefit. The outcome depends on execution and competition, not commoditization alone.

Fundraising was challenging despite the AI boom, Srinivas says. Investors asked how Perplexity would respond if OpenAI or Google added similar features, restricted access to models or outmatched its search and enterprise sales. He acknowledges that early answers are arguments about an uncertain future, supported by a track record of shipping.

He does not give an exact compute-spending percentage, but says most of the money Perplexity had spent went to compute, including GPUs, post-training and model APIs. He says it had not spent most of all capital raised. Avoiding frontier pretraining helped preserve resources for product and user acquisition.

He describes a strategy of offering useful features at attractive prices to build retention and a large user base. In five years, he predicts advertising could become the dominant revenue source if Perplexity reaches sufficient scale and makes ads work; otherwise subscriptions or enterprise may lead. This is a conditional forecast.

Srinivas distinguishes Perplexity’s aim of building a sustainable product business from a lab primarily pursuing artificial general intelligence. He expects the differences between Perplexity and ChatGPT to become clearer over time, but does not claim that outcome is already achieved.

Asked why he works on the company, he resists a polished origin story. He says he admired people such as Larry Page, then found interests and shaped his own path. Asking whether he would still choose the work if time were limited helps him examine his priorities.

In quick-fire questions, he says he has learned to take a longer-term view of people: someone who does not immediately excel may develop substantially with time. He also believes AI’s impact may be underestimated when it appears in familiar workflows rather than an unfamiliar chat interface.

He is skeptical about inserting search into messaging or social apps without considering why users open them. A new feature should fit the existing user intent and flow, he argues, rather than merely add another capability.

He thinks browsers could change more fundamentally when agents can complete tasks such as logging in and arranging a recording or purchase. Answering questions alone does not replace ordinary browsing. He also imagines an AI-native operating system while acknowledging that the systems he knew then could not yet deliver it.

For a startup CEO, Srinivas says, a difficult part of the job is repeatedly deciding between competing priorities: speed versus preparation, risk versus focus and experimentation versus improving what already works. Holding these tensions without an obvious answer is exhausting.

In a Perplexity pre-mortem, he names poor execution, indecision, lack of focus and inefficient capital use as potential causes of failure. He emphasizes management choices rather than treating a competitor’s action as the only risk. This is his own assessment, not a prediction of failure.

His ten-year ambition is for Perplexity to become an indispensable assistant for facts and knowledge. He believes people will continue to value accurate information even in a world with more AI agents. That is the company’s aspiration, not a guarantee of accuracy or future market position.