Home
May 24, 2026
View All

The AI Paradox, Perplexity's Mac Pivot, and What Autonomy Actually Costs

The Short Version#

Dan Shipper makes the case that AI doesn't reduce headcount — it raises the floor on what expertise means, which is good news for PMs and designers but demands a rethink of how we define value. Meanwhile, Perplexity quietly ships a Mac app overhaul that bets on local AI as a consumer differentiator, and Lenny's newsletter surfaces a sharp community read on where AI sits in design workflows right now.

Dan Shipper / Every — The AI Paradox: More Automation, More Humans, More Work#

Source: https://www.lennysnewsletter.com/p/the-ai-paradox-dan-shipper Credibility: High (first-person account from a founder who has actually automated his company at scale and tracked the results)

What happened: Dan Shipper published a piece — framed as an interview on Lenny's newsletter — laying out what he calls the AI paradox: Every has grown from 4 to 30 employees since GPT-3, despite aggressively automating with AI agents. His core argument is that making expert competence cheap doesn't reduce demand for expertise. It raises it. When anyone can access a good-enough version of something, the premium shifts to excellent, and excellent still requires humans who really know what they're doing. He also makes a specific prediction about workflow: most knowledge work will happen inside tools like Codex or Claude Code, the CLI era is ending, and every agent needs a human in the loop — which he sees as wildly bullish for PMs and designers specifically.

Key patterns:

  • AI makes the floor higher, not the ceiling lower — commoditized competence drives demand for genuine expertise
  • Every grew headcount 7.5x while automating extensively, suggesting AI investment and team growth are not in tension
  • The "every agent needs a human" argument positions PMs as the default decision-layer inside agentic workflows
  • Shipper's claim that most work will happen inside Codex/Claude Code (not traditional apps) is a UX surface shift PMs should take seriously
  • He explicitly calls out PMs and designers as the roles most positively impacted — not because they're safe from automation, but because they're positioned to direct it

Why it matters for PMs: If Shipper is right, the coming question isn't "will AI replace PMs" — it's "which PMs will be good enough at directing agents that they're worth paying for." That reframes how you think about career development and how you think about the value of your team. It also has direct product implications: if your users are increasingly working inside agentic toolchains, the interface layer for your product might not be where you think it is. The insight that CLI-era is ending and work moves into AI-native environments is worth watching for consumer and enterprise products alike.

Critical questions:

  • Every is a media company. Does the "automation + growth" pattern hold in product organizations, where headcount often scales with decisions rather than content output?
  • Is "every agent needs a human" a durable claim, or a transitional one that erodes as agents get better at self-correction?
  • If PM value rises because agents need directors, does that require a fundamentally different kind of PM than we're hiring today?
  • Shipper's optimism about PMs is notable but could be motivated reasoning — he needs PMs to want to use Every's tools. Worth pressure-testing.

Action you could take today: Map one workflow your team currently does in a traditional app (spec writing, user research synthesis, data pulls) and sketch out what it would look like if it ran inside Claude Code or Codex instead. What would the PM's role be in that version? What breaks?

Aravind Srinivas / Perplexity — New Mac App with Personal Computer Built In#

Source: https://x.com/AravSrinivas Credibility: Medium (announcement from the CEO on X — not a full changelog or blog post, but a direct first-party signal)

What happened: Perplexity's CEO announced today that they're deprecating the legacy Mac app and shipping a new version with "Personal Computer" built in. Personal Computer is Perplexity's local AI system for Mac — a feature that gives Perplexity access to local files, apps, and system context to answer questions or take actions on your machine. This isn't a minor refresh; deprecating the old app means they're making a hard bet on this direction rather than running both paths. The timing suggests they've seen enough signal from the Personal Computer beta to commit fully.

Key capabilities:

  • Local AI system for Mac — Perplexity can now read and reason over local files and context, not just the web
  • Persistent background agent running even when you close the app
  • Forces deprecation of legacy Mac app, meaning all Mac users move to the new experience

Why it matters for PMs: This is a consumer AI product making a deliberate bet on local context as a differentiator, at a time when most competitors are racing to add web access. The signal is about positioning: Perplexity is betting that the most valuable search isn't "what's on the internet" but "what's in my life" — calendar, docs, files. That's a different user need and a different competitive surface than Google or OpenAI's SearchGPT. For PMs thinking about AI assistant features, this shows the competitive frontier moving toward ambient, local context — not just retrieval. The deprecation decision also shows confidence: they'd rather break old users' habits than maintain two products.

Critical questions:

  • How much of Perplexity's user base is on Mac, and how disruptive is the forced migration? Are there retention risks baked into this move?
  • Local access to files and apps raises real privacy questions. How is Perplexity handling permissions and data handling transparency?
  • Does "works even when you close the app" mean a persistent background process? That's a meaningful system resource commitment — how does that land with users who are protective of their Mac's performance?
  • Who is this actually for? Power users who already use Perplexity heavily, or a new segment they're trying to reach with ambient AI?

Action you could take today: Download the new Perplexity Mac app when it's available and test it against a real workflow — try asking it something that requires both web context and local file context simultaneously. Note where it breaks or where the seams show. That's where the real PM learning is.

Lenny's Newsletter / Community Wisdom — The State of AI in Design in 2026#

Source: https://www.lennysnewsletter.com/p/community-wisdom-best-ai-cold-outreach Credibility: Medium (community aggregation — signal quality varies, but Lenny's audience skews toward working PMs and designers at real companies)

What happened: Lenny's community wisdom post from May 23 includes a thread on "the state of AI in design" as of 2026. The framing suggests the community was asked to weigh in on how AI is actually being used in design workflows — not theoretically but in practice right now. The excerpt highlights this as one of several questions answered, alongside QA relevance and AI cold outreach tools.

Key patterns (based on available context):

  • The question itself signals that AI's role in design is still contested enough to warrant a community temperature check — it's not settled
  • QA relevance appearing alongside "AI in design" suggests the community is actively wrestling with which professional roles are changing fastest
  • The pairing of "AI in design" with "advice for college students interested in PM" suggests practitioners are trying to figure out what skills actually matter now

Why it matters for PMs: PMs work alongside designers constantly. If AI is meaningfully changing how design work gets done — what tools designers use, how long things take, what they need from PMs — that changes the collaboration model. The fact that this is a "state of" question in mid-2026 (not "how to use AI in design") implies it's still a moving target. That's useful calibration: your design partners are probably figuring this out in real time, not running on a settled playbook.

Critical questions:

  • What's the actual community consensus — is AI helping designers go faster on the rote work, or is it creating new decision overhead?
  • Are PMs expected to have opinions on AI-generated design artifacts, or is that still purely in design's domain?
  • Does AI in design change the PM-designer relationship, or just the tools each side uses?

Action you could take today: Ask your design partners directly what AI tools they're using regularly (not experimentally) and what's changed about how they work with PMs because of it. Compare their answer to your assumptions.

Quick Hits#

  • Dan Shipper / Lenny's Newsletter: Interview on why PMs and designers are "wildly" well-positioned in the agentic era — every agent needs a human director, and that's the PM role (2026-05-24): https://www.lennysnewsletter.com/p/the-ai-paradox-dan-shipper

  • Aravind Srinivas / Perplexity: Deprecated legacy Mac app, shipping new version with Personal Computer (local AI system) built in — hard bet on local context as consumer differentiator (2026-05-24): https://x.com/AravSrinivas

  • Cursor: Automations now available in the Agents Window, multi-repo support added, new automation runs 50% off for 7 days — Cursor continues pushing toward ambient background agents (2026-05-20): https://cursor.com/changelog/05-20-26

  • Mistral AI: Emmi (industrial AI company) joins Mistral to accelerate AI-native industry adoption — first real signal of Mistral building a vertical enterprise play beyond model API (2026-05-22): https://mistral.ai/news/accelerate-ai-native-industry

  • Character.AI: Shipped "Smarter Memory for Smarter Chats" on May 21 — expanded memory across conversations, which is a direct retention play; smarter memory = stickier characters (2026-05-21): https://blog.character.ai/memory/

The Thread#

The autonomy dial is the new core PM decision. This week's signals keep returning to the same tension: how much should an AI system do without asking? Perplexity bets on "works even when you close the app." Cursor adds multi-repo automations that run in the background. Notion's custom agents handle entire workflows. Character.AI adds memory that persists across sessions. Each of these is a small turn of the dial toward more autonomy — and each one creates a new surface for user trust to break. The interesting PM question isn't "how autonomous should this be?" in the abstract. It's "what does the user need to feel in control even when they're not actively watching?"

Sit With This#

Dan Shipper's core claim is that Every grew from 4 to 30 employees while automating aggressively — and that AI raised demand for expertise rather than suppressing it. His explanation: when competence gets cheap, excellence gets more valuable.

For your product or team: Think about one AI feature you've shipped or are considering. If it makes a "good enough" version of something much cheaper or faster, who captures that value — users, the business, or neither? And is there a version of this feature that drives demand for something more excellent, rather than just replacing something adequate?