Sam Schillace on Innovation and the Future of AI Products

Guest: Sam Schillace — Corporate VP and Deputy CTO at Microsoft; inventor of Google Docs (Writely, acquired by Google 2006); previously SVP Engineering at Box (through IPO); founded six startups; VC at Google Ventures.
Host: Lenny Rachitsky
Source: Lenny’s Podcast. Recorded ~2023–24.

Overview

A wide-ranging conversation spanning disruptive innovation frameworks, the origin of Google Docs, Microsoft’s AI trajectory under Kevin Scott, and career philosophy. The AI content is more philosophical than technical — Schillace is an innovation optimist applying what-if thinking to the generative AI moment, with concrete observations about multi-agent memory and the future of documents.

Key ideas

What-if vs. why-not thinking. Why-not questions block disruption: “what if you’re on an airplane?” “What if it hallucinates?” What-if questions enable it: “What if we can build software around a reasoning engine? What if this is the second Industrial Revolution?” Most engineers are structurally pessimistic; Schillace argues there is not much prize for being pessimistic and right.
“AI isn’t a feature of your product; your product is a feature of AI.” The platform shift is total. In the same way the internet swallowed every business, AI will be the platform on which products are built, not an enhancement bolted onto existing products.
Pixels will become free. The internet made distribution free; AI will make pixel (interface) production free. Just as free distribution changed what was worth building, free pixel production changes what interfaces are worth building. Corollary: if you are an interface-production business, your moat is eroding.
Multi-agent whiteboard memory. In Microsoft’s experiments with multi-agent systems, giving agents shared working memory (a “whiteboard” visible to all) made them measurably smarter — unexpectedly so. Explicit shared state outperformed just more context or more compute.
Future of documents: bots or docs. Schillace’s frame: every document is a conversation-in-waiting. The future is dynamic, intentional, semantic, fluid, personalised documents that know what they are and respond accordingly. The question “is this a bot or a document?” will dissolve.

What-if vs. why-not

The Google Docs origin story is a live demonstration of the framework. Early objections: browsers are not ready; people will not trust cloud storage; no connectivity on aeroplanes. All why-not questions. None killed the product.

The pattern: a genuinely disruptive idea polarises responses. If people mostly mildly like or mildly dislike something, it is incremental. If there is a strong love–hate bifurcation — as with Google Docs and with ChatGPT — that is a signal the disruption is real. The haters do not determine outcome; what matters is whether you run out of the people who love it.

Corollary: “toy” is a tell. When critics describe something as just a toy, they often mean they cannot find a better objection. Toy status frequently signals real disruption.

Google Docs origin

Writely (2005): three file servers in a Texas data centre running C#, with a sysadmin in the Philippines. Migrated to Google infrastructure (Java + Bigtable) with eight hours of maintenance-mode on a Sunday. Sergei Brin asked three days later when migration would happen — it had already happened without anyone noticing.

The first Google Doc (October 2005) still exists and still opens. Backend rewritten twice, frontend rewritten twice. Schillace is ambivalent about whether it is “the same document.” It contains a line asking a co-founder if collaboration is working and a pasted picture of Edna from The Incredibles.

AI as cognitive surplus

Schillace’s framing for the generative AI moment: the first Industrial Revolution gave humanity surplus physical energy beyond human bodies (water wheels, steam). Generative AI is the second Industrial Revolution — surplus cognitive energy beyond human brains.

This framing makes the why-not questions structurally similar to the objections to industrialisation. The relevant question is not “does it hallucinate?” but “what can you build if you assume the reliability curve bends toward good-enough?”

Multi-agent systems

Microsoft’s experiments with multi-agent systems (Infinite Chatbot, Semantic Kernel) produced an unexpected finding: shared whiteboard working memory made agents measurably smarter. Putting a debugger agent and a self-documenting agent in the same loop, with a shared writable scratch space visible to all, improved output more than adding context or compute to individual agents.

Implication: explicit, structured shared state is a meaningful lever in multi-agent architectures — not just a convenience for coordination, but a capability amplifier.

Pixels will become free

The internet made distribution free, which destroyed the value of being a distribution business (retail, media, physical software) and unlocked everything that depends on free distribution (streaming, SaaS, open source). AI will make pixel production free — generating interfaces, layouts, copy, personalised visual experiences at marginal cost.

Businesses whose moat is interface production (design agencies, template businesses, certain SaaS UI layers) face structural threat. Businesses whose moat is knowledge, data, or trust are better positioned.

Future of documents

Schillace’s prediction: the distinction between “bot” and “document” dissolves. A document that knows what it is, responds to questions about its own content, routes readers to the right section, and updates itself is not a document in the traditional sense. The end state is intentional, semantic, personalised artefacts — documents that behave like specialised conversational agents.

Career philosophy

“Do what you feel guilty to get paid for.” If something is easy and fun, that is a signal to pursue it harder, not a reason to feel its value is low.
Virtue from error (personal motto): mistakes are the main input to learning; treat error as a resource, not a failure.
Optimism beats pessimism as a prior: “I’ve missed out more by being pessimistic than by being too optimistic too early.”