Playbook
Tasteful Skills
“Tasteful Skills” argues that the best agent skills are not documentation or best-practice lists.
Back to Podcast Digest
AI Agents: Mirage Or Real Revolution? — With Dmitry Shevelenko
TL;DR
Perplexity says the agent push is about revenue, not a consumer AI stall-out — Dmitry Shevelenko dismisses flat DAUs as a vanity metric and points instead to ARR growing from under $250 million at the start of 2026 to over $500 million a month before the interview.
The core bet is that AI users really wanted work leverage all along — even when Perplexity looked like a search product, Shevelenko says people were using it as a “secret weapon” for work, and now agent products like Perplexity Computer simply meet that demand more directly.
Shevelenko’s big analogy: everyone just got ‘100 employees’ — his thesis is that AI shifts people from doing tasks to acting like executives, delegating work to agents for marketing, analysis, coding, taxes, and meeting prep.
Perplexity thinks its edge is multimodel orchestration, not one giant frontier model — Shevelenko says a single task might use Opus for planning, GPT for writing, Gemini for audio, Grok for fast research, and Sonnet for Python, which closed ecosystems like OpenAI or Anthropic can’t fully match.
Trust is the real adoption bottleneck, so Perplexity is pushing guardrails and ‘Final Pass’ — after discussing scary permissions like Gmail and Calendar access, Shevelenko argues humans still need to set objectives, spot-check outputs, and apply taste, while products like Final Pass fact-check decks, spreadsheets, and PDFs for external and internal inconsistencies.
On pricing, Perplexity argues the future looks more like Costco than unlimited SaaS — Shevelenko says subscriptions get users in the door, but agentic work increasingly needs usage-based computer credits because one task may cost $0.05 while another, like long-horizon video generation, can cost $50.
The Breakdown
From “wrapper company” to independent agent player
Alex Kantrowitz opens by joking that he expected Perplexity to be an Apple subsidiary by mid-2026, teeing up the old criticism that Perplexity was just a “wrapper.” Shevelenko leans into the company’s independence, says Apple is excited about Perplexity Computer’s use of Mac minis, and reframes the whole category with a memorable line: what started as a wrapper is now a “harness” for multimodel orchestration.
Flat consumer AI usage vs. a revenue story that looks very different
Alex brings the hard numbers: consumer AI app growth has flattened, ChatGPT traffic growth is modest, and Perplexity’s share in AI search has leveled off. Shevelenko’s answer is blunt — the metric he checks every morning is revenue, not MAUs, and he says Perplexity went from under $250 million ARR at the start of the year to over $500 million recently, arguing that paid usage is a truer signal of value than hypey top-line user counts.
Why AI growth cooled — and why he thinks curiosity is now the bottleneck
On the broader slowdown, both agree that voice, image generation, and moments like Studio Ghibli-style avatars created novelty spikes that didn’t always turn into habits. Shevelenko’s diagnosis is that model capabilities have run ahead of user behavior — people are still using AI in a “Web 1.0” way for weather, sports scores, and basic retrieval, while the real constraint is now human curiosity and willingness to explore deeper workflows.
The real pitch for agents: not novelty, but payroll replacement
The conversation gets concrete when Alex describes using Perplexity Computer to generate a daily digest from Gmail and Calendar. Shevelenko says that unlike image-generation crazes, Computer usage keeps rising week over week, because users think of it less as software spend and more as payroll — a way to avoid hiring agencies, analysts, or support staff, with people effectively managing teams of digital workers.
Taxes, fact-checking, and the three jobs humans still keep
Asked whether people should trust AI with sensitive work like taxes, Shevelenko flips the frame: one strong use case is using AI to catch human mistakes, including accountants’ errors. He highlights Perplexity’s upcoming “Final Pass” workflow for checking PDFs, presentations, and spreadsheets, then lays out his three enduring human roles in an AI world: setting objectives, validating outputs, and exercising taste.
The scary permissions question — and why the Mac mini matters
Alex reads out the long list of Gmail, Calendar, and Workspace permissions he granted Perplexity Computer, making the trust issue feel very real. Shevelenko argues users can start with zero connectors or read-only access, and says the Mac mini setup is less about hiding the agent away than about unlocking more capability — access to local files, iMessage, and 24/7 long-running tasks — while predicting a hybrid future where some inference runs locally and some in the cloud.
Why Perplexity thinks it can survive OpenAI and Anthropic
Pressed on competition from Codex and Claude Code, Shevelenko makes Perplexity’s case in three parts: multimodel orchestration, accuracy, and usability. His best example is personal — building mini podcasts for his kids with one task that can use Opus for planning, GPT for scriptwriting, Gemini for audio, Grok for research, and Sonnet for stitching code together, all while Perplexity’s search and data flywheel improve grounding and accuracy.
Models, China, pricing, and staying fast with a 300-person team
In the final stretch, Shevelenko says model labs actively want Perplexity using their APIs, so he’s less worried about being cut off than about maintaining execution speed. He distinguishes using Chinese-developed open models like Kimi K2 from hosting Chinese APIs, says Perplexity runs and post-trains such models in US data centers, compares future AI pricing to Costco memberships plus usage-based credits, and closes with a company-building stat that explains the pace: Perplexity reached roughly $500 million ARR with about 300 people, growing headcount only 34% while revenue rose 5x.
Share
Keep Reading
Two Ways to Follow
The Weekly Echo. The inbox-shaped summary of what mattered.
New editorials announced here.
