Why Fewer Senior PMs Might Beat Any Ratio, and Perplexity Goes Multiplayer

Perplexity — Computer Becomes a Multiplayer Agentic OS with Projects#

Source: https://x.com/AravSrinivas (August 2, 2026) Credibility: High (first-party announcement from CEO Aravind Srinivas)

What happened: Aravind Srinivas announced the launch of Projects on Perplexity Computer, framing it as turning Computer into "a multiplayer agentic operating system for work." Projects give Computer persistent memory, file storage, and sessions scoped across hubs and users. This is a significant step beyond individual AI tool usage: teams can now operate within shared agentic contexts rather than each user starting fresh every session.

In a separate announcement the same day, Srinivas also confirmed a partnership with Intel to bring Perplexity Personal Computer with local models and hybrid inference to Intel Ultra Series 3 laptops, extending Perplexity's reach into native hardware integrations.

Key capabilities:

• Projects create persistent, scoped sessions across users and hubs
• Files and memory persist within a Project context, not just per-conversation
• Multiplayer support means multiple users can share agentic workflows, not just view results
• Intel Ultra Series 3 integration adds local model inference alongside cloud, enabling hybrid compute paths
• Hardware partnership signals a play for on-device AI that reduces latency and data egress concerns

Why it matters for PMs: This is the classic individual-to-team expansion play, and it's arriving faster in AI tooling than most people expected. Perplexity's Computer started as a personal agentic browser. Projects turn it into shared infrastructure for team workflows. If your team uses Computer or is evaluating it, this changes the adoption calculus: it's no longer a per-seat productivity tool, it's potentially a shared workspace. The Intel hardware deal also matters for enterprise buyers who need local inference for compliance or latency reasons. Hybrid inference (local + cloud) is an architecture pattern that will show up increasingly in procurement conversations.

Critical questions:

• What does "multiplayer" actually mean here operationally: real-time co-presence, async handoffs, or shared state between sequential sessions? The announcement doesn't specify.
• How does Perplexity handle access control and permissions within Projects? Shared agentic sessions with file access create real security surface area.
• Does the Intel integration require specific hardware SKUs, or is it a software layer that runs on any Intel Ultra Series 3 chip?
• Is this a premium tier feature, or included in existing subscriptions? Pricing clarity will determine enterprise adoption speed.

Action you could take today: If your team has been using Perplexity Computer individually, map one workflow where shared persistent memory would eliminate duplicated context-setting. That's your proof-of-concept pitch for trying Projects with a small team.

Lenny's Newsletter / Tom Verrilli — The CPO Who Regrets That Product Management Exists#

Source: https://www.lennysnewsletter.com/p/this-cpo-regrets-that-product-management (August 2, 2026) Credibility: High (Lenny's Newsletter, primary interview with Tom Verrilli, CPO of Whatnot)

What happened: Tom Verrilli, CPO of Whatnot, went on Lenny's podcast and made an argument that's going to circulate in PM circles for a while: he regrets that product management exists as a distinct organizational function. His core claim is that the traditional PM model optimizes for coordination and coverage rather than outcomes, and that fewer, more senior PMs doing real individual contributor work outperform any team structured around a headcount-to-engineer ratio.

The framing "regrets that product management exists" is deliberately provocative, but the underlying argument is about org design. Verrilli's position is that PM proliferation creates a coordination layer that slows decision-making, and that the better model is senior PMs who stay close to the actual product work rather than managing up and managing across.

Key patterns:

• The "fewer, more senior" argument rejects ratio-driven org design (e.g., 1 PM per N engineers) in favor of sparse but high-leverage IC-style PMs
• Verrilli's critique is that PM as a function has evolved to optimize for organizational coverage, not product quality
• The implication is that AI-assisted execution reduces the need for PMs as coordinators, shifting the value of a PM toward judgment and taste rather than throughput
• This is consistent with a pattern appearing across AI-era product orgs: PM roles narrowing from generalist coordinators to decision-makers embedded in smaller, faster teams

Why it matters for PMs: This is one of the sharper articulations of a tension that's been building since coding tools started reducing engineering cycle time. If AI handles more of the execution, the PM-as-coordinator role becomes redundant faster than expected. Verrilli is essentially arguing that the valuable part of PM was always the judgment and product taste, not the meeting-running and backlog-management. That's a defensible position, and it maps directly onto what's happening in teams that are shipping faster with smaller headcounts using AI tooling.

Critical questions:

• "More senior, fewer PMs" sounds clean in a CPO's framing, but what does Whatnot's actual ratio look like and has it changed? Data would make this argument more than an opinion.
• Does this model hold for complex multi-product orgs, or does it only work at Whatnot's stage and surface area?
• If the PM function shrinks, where does the user research and discovery work go? Senior ICs still need customer feedback infrastructure.
• Is this an AI-era argument or a Whatnot-specific argument? The case would be stronger if Verrilli connected it explicitly to how AI tooling has changed what execution requires.

Action you could take today: Audit your last two weeks of calendar. What percentage of time was judgment and taste work versus coordination and documentation? That ratio is the honest answer to how much of your PM role is already AI-replaceable.

Vercel AI Gateway — Logs Page and Cost Observability#

Source: https://vercel.com/changelog/ai-gateway-logs (August 1, 2026) Credibility: High (first-party changelog)

What happened: Vercel's AI Gateway now has a dedicated logs page that surfaces per-request cost, token counts, duration, and fallback paths across providers, with shareable URL-based filters. This arrived alongside a separate update to spend budgets and alerts. Taken together, these two changes mean that AI Gateway has moved from a routing and proxy layer into a cost management and observability tool.

Key technical details:

• Per-request breakdown: cost, tokens in/out, latency, and which fallback path was taken
• Filters are shareable via URL, which is meaningful for team workflows (link a filtered view to a Slack thread or Jira ticket)
• Spend budgets and alerts were added simultaneously, enabling teams to cap spend by team or project
• Provider fallback paths are now visible, which helps debug routing behavior when primary models are unavailable

Why it matters for PMs: Cost visibility on AI features has been one of the most consistently under-built parts of AI product tooling. Most teams running AI workloads have a rough sense of model costs at the API invoice level, but very little visibility into which features or user flows are responsible for the spend. Vercel's logs page is a step toward feature-level cost attribution. For PMs, that means being able to walk into a roadmap conversation with actual cost-per-use data rather than estimates. It also changes the conversation around fallback models: if you can see when requests fell back to a cheaper model and whether quality held, that's real data for a build-vs-buy or model selection decision.

Critical questions:

• Does this work across all providers on AI Gateway, or only the natively integrated ones? The excerpt mentions "across providers" but it's unclear if third-party or self-hosted endpoints are included.
• Is there a retention period on logs, or do they roll off? For cost auditing purposes, 30-day retention may not be enough.
• Can costs be attributed to specific routes or feature surfaces, or is it per-request only? Feature-level attribution would be significantly more useful.

Action you could take today: If your team is using Vercel AI Gateway, pull the logs for your highest-traffic AI feature and see what the actual cost-per-request distribution looks like. Most teams will find a long tail of unexpectedly expensive requests worth investigating.