---
title: building a research vs. a product company
description: "what changed when we stopped solving today's problems and started betting on 2–5 years out"
author: Matthew Woo
publishedAt: 2026-08-21
publication:
  kind: syndicated
  originalUrl: https://matthewedanwoo.substack.com/p/building-a-research-vs-a-product
  originalPublisher: "Matthew Woo's Newsletter"
---

I didn’t intend to build a research company. I don’t have a PhD — no one on our team has one either. For the last decade I’ve been a product leader who wants to root a team in concrete, hair-on-fire, painkiller problems. But in the past 2 months of building we’ve pivoted further and further away from concrete problems (esp in software) toward more ambiguous ones that require a leap of faith.

## Why the shift

- concrete problems (esp in software) are saturated at significant levels both from buyer and supplier — there are [~30–40K SaaS companies globally](https://colorlib.com/wp/saas-statistics/) (~17K in the US alone) and growth in SaaS revenue is slowing ([12% in 2026 vs. 30% in 2021](https://www.saas-capital.com/blog-posts/four-early-2026-saas-trends/)) as buyers are [consolidating spend](https://www.bettercloud.com/monitor/saas-statistics/) vs. adopting new software

- you can still find niche spaces to start, but I found that the moment we extended beyond that initial audience we immediately hit against [several companies doing the same thing](https://techcrunch.com/2024/11/22/y-combinator-often-backs-startups-that-duplicate-other-yc-companies-data-shows-its-not-just-ai-code-editors/)

- that’s not a bad thing if you’ve found a space where you can build a tailored solution at a fraction of what it used to cost — but the outcomes feel like they’re in the [M not B range](https://aventis-advisors.com/saas-valuation-multiples/) for many ideas in the software space

## The questions we’re following instead

- so in the last 2 months the team & I have followed our intuition on where the world will be 2–5 years from now, by asking questions like:

  - what will be the dominant way people interface with AI — and how does that differ by situation?

    - devices → phone, desktop, smart speakers, AR glasses, VR

    - modality → text, voice, BCI

    - surfaces → chat apps, software, vehicles

  - based on those interfaces, how do the behaviors and expectations people have with AI change?

  - and based on those behaviors and expectations, how will software and hardware have to adapt?

- we’re still early in navigating these questions (will share more of our opinions later), but some observations on what’s different operating like a ‘research’ company vs. a ‘product’ one:

## Experiments vs. milestones

- we spend more time thinking through what experiments to build and run to answer different questions — more of a ‘breadth’ search vs. the ‘depth’ search of milestones that sequentially build on each other

## Custom instrumentation vs. out-of-the-box solutions

- when I read Kuhn’s work on scientific revolutions, one thing that stood out was his [argument about measurement](https://www.journals.uchicago.edu/doi/abs/10.1086/349468): it’s the *quantitative* anomalies — the gaps that only become visible once your ability to measure a phenomenon improves — that crack the current belief system and usher in a new one. Qualitative anomalies just get patched over

- not a perfect analogy, but I think some of that applies in ‘research’ mode — if you’re doing something at the frontier, your measurements need to be custom to your experiment, because out-of-the-box tools won’t give you the fidelity to generate new insights

## Research papers vs. best practices

- spend at least a day reading research papers to understand where the technology is going and what happens once it’s commercialized

- for example: the initial exploration of full-duplex voice models started appearing in late 2024 with Kyutai’s [Moshi paper](https://arxiv.org/abs/2410.00037), [accelerated through 2025](https://arxiv.org/pdf/2502.13472), and by mid-2026 showed up in products — Thinking Machines’ [interaction models in May](https://techcrunch.com/2026/05/11/thinking-machines-wants-to-build-an-ai-that-actually-listens-while-it-talks/), and OpenAI’s [GPT-Live](https://openai.com/index/introducing-gpt-live/) replacing Advanced Voice Mode in ChatGPT in July. Reading the papers gave you an ~18-month preview

To be fair — at our core we’re still not a research company (at least not yet), and I think there’s real value in bringing product principles to this stage. Will share more thoughts in a future post.