This week, Harvard had the chance to hear from Aravind Srinivas, CEO and co-founder of Perplexity, and Johnny Ho, co-founder and SEAS alum, alongside Professor Jim Waldo. The conversation covered everything from future bets on open-source models to how to think about AGI, all while diving into what it really takes to build an AI-native company.
From pretraining to reasoning
Aravind Srinivas and Johnny Ho explained that the era of massive pretraining is plateauing. The future lies in post-training—teaching models to reason, solve tasks, and complete workflows. While pretraining helped models absorb general knowledge, post-training is about making them useful and capable of things like research, multi-step reasoning, and tool use. This is where all leading labs are now focusing their energy.
DeepSeek
They highlighted DeepSeek as a major moment for open-source. Beyond systems work like GPU efficiency and memory optimization, DeepSeek was the first to release an open-source reasoning model. Their BC-Zero technique demonstrated reinforcement learning without labeled data, a breakthrough in post-training. This proved that open-source isn't just catching up—it’s innovating at the frontier.
Bets
Perplexity’s growth came from making early, disciplined bets. Rather than burn cash training large models, they focused on product, waited for open-source to improve, and built data feedback loops. They assumed model quality would rise, serving costs would drop, and user data would become a flywheel—and they were right. This helped them scale efficiently and delay heavy infrastructure spend until it mattered.
Why they can compete with Google
Their training pipeline is powered by user feedback, human evaluations, AI model evaluations, and logs of source usefulness. To train classifiers at scale, they rely heavily on synthetic data—using large models to label queries for smaller models. The result is a tightly looped post-training system tuned for practical tasks like summarization, formatting, and source ranking.
Perplexity believes there are three structural reasons they can compete with incumbents. First, the infrastructure cost for LLM inference at Google-scale is enormous. Second, the reputational risk of hallucinations is higher for a brand like Google. Third, LLM answers threaten their ad model by reducing clicks, which advertisers rely on. These constraints don’t apply to smaller, focused players like Perplexity.
Not just another wrapper
They addressed critiques that Perplexity is "just a wrapper." The team sees this as a misunderstanding of strategy. They intentionally built on existing models while collecting data, building user trust, and waiting for the right moment to go deeper. Being AI-native means knowing when and how to build infrastructure—not doing it for show.
Unlike labs that scrape the entire internet for pretraining, Perplexity focuses only on post-training. They teach models specific skills like document navigation, button-clicking, or spreadsheet editing—tasks useful in real workflows. This doesn’t raise the same data concerns as pretraining, and it relies more on reinforcement learning than scraping.
They raised an open question: will AI agents call existing apps, or will those apps “wrap” the agents? It depends on the business model. Uber might benefit from agent integration, but platforms with ad revenue (like Amazon or Instacart) have reason to resist agents that bypass their monetization layer. This dynamic will shape how AI systems interact with the broader software world.
What AGI might look like
They offered a grounded view of AGI. Vertical “AGIs” (automated domain-specific agents) are already emerging, but a system that can debug infrastructure, write strategy roadmaps, or deploy code end-to-end is still out of reach. If an AI could do that, it would mark a true AGI milestone—and they’d happily pay millions to use it.
A future AI engineer
The problem they most want to solve: debugging. AI can autocomplete code, but it can’t yet fix bugs across complex systems. Today, debugging takes time, energy, and coordination. An AI that could reliably do this would be a breakthrough—one of the most valuable possible contributions AI could make to engineering work.