The Workforce Bifurcation, and What Perplexity's Failures Tell Us

Lenny Rachitsky / Noam Segal - Annual AI Sentiment Survey: The Great Tech Bifurcation#

Source: https://www.lennysnewsletter.com/p/how-tech-workers-actually-feel-about Credibility: High (first-party survey data, published with methodology via Lenny's Newsletter)

What happened: The 2026 annual AI sentiment survey — authored by Noam Segal and featured on Lenny's Newsletter — lands with a stark finding: the tech workforce is splitting in two. Half the workforce is thriving, half is struggling, and burnout just hit a record high. The framing is "the great tech bifurcation," and the implication is that AI is accelerating the divergence rather than leveling the playing field.

Key patterns:

• The split is correlating with AI adoption: workers who have integrated AI tools into their workflows are on the thriving side; those who haven't are on the struggling side
• Burnout is at a record high — which is a notable data point given that AI was supposed to reduce cognitive load and repetitive work
• The bifurcation framing suggests the adoption gap is widening, not closing, which has real implications for team-level AI rollouts

Why it matters for PMs: This is the best leading indicator we have for where enterprise AI adoption actually stands. If you're a PM shipping AI features to internal teams or to B2B customers, this data is directly relevant to your adoption strategy. The finding that AI is accelerating divergence — rather than democratizing capability — should make you think hard about onboarding and floor-raising features, not just ceiling-raising ones. The workers who are thriving aren't all more talented; many are just earlier adopters who got help getting started. That's a product problem.

The burnout data is worth sitting with separately. If AI is increasing output but also increasing the pace of work and expectations, you may be building tools that help individual contributors sprint faster toward the same wall. That's a retention problem in disguise.

Critical questions:

• Is the bifurcation driven by access differences (some workers don't have good tools) or adoption differences (some workers have tools but aren't using them effectively)?
• If burnout is rising alongside AI adoption, what's the mechanism — are AI-augmented workers taking on more work rather than doing the same work with less effort?
• What does this mean for team-level AI rollouts where some people adopt fast and others don't — does the gap create friction or resentment within teams?
• How do you design onboarding for AI tools that actually closes the gap rather than rewarding people who were already high-adopters?

Action you could take today: Look at your team's AI tool usage data and identify who's in the thriving half vs. the struggling half. If you don't have that data, that's the first problem. If you do, the second step is figuring out whether the struggling half needs different tools, different training, or a different expectation-setting conversation with leadership.

Ben Tossell - Perplexity's Real-World Failures as a Product Quality Case Study#

Source: https://x.com/bentossell/status/1915392814390694184 Credibility: Medium (practitioner critique, primary observation from a builder who uses the product — not a controlled study, but specific and concrete)

What happened: Ben Tossell posted a pointed critique of Perplexity AI, using specific product failures to argue that despite raising approximately $900M, the product doesn't deliver on its core claims. The examples he cites: reservation booking fails across different locations, reminders are set incorrectly, and location-specific features don't work reliably. The critique is framed around the gap between what Perplexity claims to do and what it actually does in everyday use.

Key patterns:

• Location-aware features are failing — a foundational capability for any assistant claiming to handle real-world tasks
• Reminders and scheduling features are unreliable — exactly the kind of high-stakes feature where a single failure destroys trust
• The product is raising at a valuation that implies category leadership, while shipping features that don't survive basic QA in real-world conditions
• This is a specific instance of a broader pattern: AI products demo well in controlled conditions and fail in edge cases that real users hit constantly

Why it matters for PMs: Tossell's critique is a useful template for stress-testing your own AI feature claims. The question isn't "does this work in our demo?" but "does this work for a user in a city we didn't test, on a Tuesday afternoon, for a task we didn't anticipate?" Location-aware features and time-based features are particularly brutal here because they're highly variable and contextual. If you're shipping anything in that category — reminders, reservations, local recommendations — the failure modes Tossell describes should be on your pre-launch checklist.

There's also a trust curve implication. Perplexity has been building brand equity on the claim that it can replace search for real tasks. Every failure like this trades trust for novelty. At $900M raised, the cost of a broken reservation booking is higher than the feature is worth if it's unreliable.

Critical questions:

• At what reliability threshold does an AI feature earn the right to be surfaced to users vs. kept in beta or behind a flag?
• Is Perplexity's failure here a model problem, a product design problem, or a QA problem? The fix is different depending on which one it is.
• How do you build user trust back after an AI feature fails on a high-stakes task like a reservation?
• Is there a version of this feature that should have launched with explicit scope limits ("works in these 10 cities") rather than appearing to promise broad coverage?

Action you could take today: Take one AI feature you're currently shipping or planning to ship and write out five failure scenarios that wouldn't show up in internal testing — different locations, different times of day, different user contexts. If you can't answer how the feature behaves in each one, you have a gap.

Aravind Srinivas / Perplexity - Local Models on Intel Laptops via Hybrid Inference#

Source: https://x.com/AravSrinivas Credibility: Medium (first-party post from Perplexity CEO — announcement without detailed product specs)

What happened: Perplexity CEO Aravind Srinivas announced a partnership with Intel to bring local models and hybrid inference to Intel Ultra Series 3 laptops. The framing is "Personal Computer with local models" — meaning Perplexity is making a move toward on-device inference as part of its product strategy, not just relying on cloud-based search.

Key patterns:

• Hybrid inference means the product can route queries between local and cloud models depending on task complexity, connectivity, or privacy requirements
• Intel Ultra Series 3 is the hardware target — positioning this as a near-term user-facing capability, not a research initiative
• This is a direct response to the privacy and latency concerns that hold enterprise and power users back from fully committing to cloud AI tools
• The "Personal Computer" framing is deliberate — it echoes the original PC revolution and positions local AI as a return to user-owned compute

Why it matters for PMs: Two things to watch here. First, hybrid inference as a product pattern: if Perplexity pulls this off, it creates a meaningful differentiation vector — a product that can do private, fast, local processing for sensitive queries while still accessing cloud-scale retrieval for broad search. That's a genuine build-vs-buy signal for anyone thinking about AI features in enterprise contexts where data residency matters. Second, the Intel partnership reveals that Perplexity is betting on hardware-integrated distribution as a growth channel — getting into the OS or device layer is a classic way to acquire users without relying on search or app stores.

Critical questions:

• What's the actual performance gap between local and cloud inference for Perplexity's core use cases? If the local model is noticeably weaker, does hybrid inference mask that or make it worse?
• How does Perplexity handle the handoff between local and cloud — is it transparent to users, or does it happen invisibly?
• Does this move change Perplexity's privacy story in a meaningful way, or is it primarily a latency/distribution play?

Action you could take today: If you're building anything with a privacy-sensitive use case, pull up the Intel Ultra Series 3 spec and look at the NPU capabilities. The on-device AI hardware story is moving fast enough that what was "not viable" six months ago may now be in range.