Build something real: Former product lead at Meta AI on why physical AI is valuable but demands unglamorous repetition

AI made everyone an idea person – it didn’t make anyone a builder. Here, Avinav Pashine breaks down why the gap between a flashy demo and a resilient product is widening, how agents earn human trust and why physical AI refuses to behave like a benchmark.

AI has simplified the initial prototyping quest. It hasn’t made the execution any easier.

As software becomes dramatically faster to build, putting together a slick demo is no longer the hard part. The real challenge becomes deciding what actually deserves to be built, whether people genuinely need it, and whether it survives contact with the messy reality of daily use.

That’s where former Meta Reality Labs product lead Avinav comes in. We recently sat down with him for a chat ahead of his upcoming talk at How to Web Conference 2026 to unpack what happens when the demo ends and real-world execution begins:

“Building has undoubtedly become faster and nearly free, but here’s the fun part: you still have to find out what the user actually does with your product on a Tuesday afternoon when nobody’s watching. Challenging yourself to solve that before you start building is your real foundation.” 

A feature proves feasibility, a product earns users

Having spent nearly 18 years across engineering, consulting and product – from building search and personalization engines at Zalando to spending eight years at Meta Reality Labs working on mixed reality, computer vision, spatial AI and agentic wearables – Avinav has operated at the absolute frontier of technology. Today, he is building a new company focused on human movement intelligence, combining real-world physical signals with AI to understand how people recover, adapt and perform.

When he steps onto the stage next week, he’ll be challenging our core question head-on: while anyone can generate an idea, the distance between a weekend hack and a lasting company is getting wider, not narrower.

“Speaking of that foundation, the edge cases you didn’t solve will still keep you up at night, and without distribution, the most innovative product on earth gets buried in the founder’s den. A feature proves something is technically feasible. A product still has to earn its users – and a company still has to earn its commercial stamp.”

Trying to build purely for short-term hype is a trap. The real job of a builder is grounding technology in real commercial markets, domain mastery and technical integrity.

99.99% reliability doesn’t matter if it fails when it counts

Inside a web browser, inputs are reasonably well-behaved. Out in the real world, the light changes, someone steps between the camera and the subject, the hardware overheats and the cellular connection drops at the worst possible moment.

Building physical AI at Meta taught Avinav a harsh truth: what looks like an edge case in a controlled lab environment is just a normal day in a user’s life.

“A model can be technically impressive and still feel completely broken if it fails in the one moment the user actually needs it. People forget the 99.99% of the time it worked. An impressive demo is the easy part. The real job is working out the conditions under which the product can be trusted, and being honest about where those conditions end. Outside them, fail gracefully and keep pushing to widen the line.”

The path from demo to global scale relies on designing around hardware, battery, thermal and environmental limits to deliver consistency and quality across unpredictable human behavior – piling on extra surface-level features won’t fix a product that breaks in the wild.

From understanding information to understanding situations

For Avinav, the bigger shift happening in frontier tech is moving away from AI systems that merely process information to systems that understand situations: what a person is doing, what is happening around them, what they are trying to accomplish and whether now is the right moment to intervene.

This situational awareness is where physical AI and agentic systems intersect. As AI is going through a makeover, becoming a context-aware assistant, users judge an agent by its discretion and discernment.

Avinav compares deploying agentic systems to onboarding an intern: it watches first, then suggests, then asks, and only after being helpful on low-risk tasks does it earn the right to handle consequential ones.

“Say you’re driving to the grocery store and your assistant reminds you to pick up milk. Alone in the car, that feels helpful. With a friend in the passenger seat mid-conversation, the same reminder feels like an interruption. Same information, same words, one extra person in the car. Users will judge agentic AI by how well it understands what it shouldn’t do.”

This is why augmentation, starting with the human and asking what’s getting in their way, is vastly more resilient than pure automation. People adopt technology far more readily when it gives them agency rather than asking them to hand over total control on day one.

“The next generation of AI products will win because they can exercise judgment about what deserves attention, memory and action,” Avinav added.

Move fast without being careless – commitment over compliments, always

Validating product demand inside a tech giant comes with subtle traps. Built-in distribution, research teams and thousands of internal testers can easily make a mediocre value proposition look far healthier than it is.

Moving back to zero-to-one building as a founder strips away those cushions. From scratch, demand is much less polite. It requires someone to give up their time, share sensitive data, alter their workflow or hand over real money.

Instead of chasing compliments, founders must look for true commitment: Will the user return? Will the organization navigate internal friction to run a pilot? Will they tolerate the inconvenience of an unfinished product because the underlying problem matters that much?

To maintain technical integrity while moving at founder speed, Avinav advises categorizing operational progress strictly into three distinct buckets: what you’ve demonstrated, what you currently believe and what you haven’t validated yet.

“AI lowered the cost of starting, not the standard for creating”

How to Web Conference’s core tagline, Keep Building, speaks directly to the resilience required to ship technology that lasts. Descending on Bucharest to speak to founders, product leaders and engineers, Avinav sees a clear imperative for the next generation of products.

“AI has lowered the cost of starting. It has not lowered the standard for creating something that matters. Choose a real problem. Stay close to the people experiencing it. Build with integrity. And then keep building.”

Meet Avinav Pashine next week at How to Web Conference 2026

Secure your spot today to learn directly from global leaders in physical AI, product strategy and spatial computing.

Get your tickets for How to Web Conference 2026.