# Simon Gu

> Personal site of Simon Gu: essays on product development, technology and faith, his work history, and write-ups of things he has built with AI.

Exploring the intersection of technology, product development, and obedience to God. Every link below points to a Markdown version of the page; drop the `.md` for the web version. The full text of the site is at https://gufamily.org/llms-full.txt.

## Site

- [Home and about](https://gufamily.org/index.md): I'm a follower of Jesus, a devoted husband to Valentina, and a proud father to Arabelle and Barnabe. After those 4, I'm passionate about building products that matter and exploring how technology shapes our world.
- [Work history](https://gufamily.org/work.md): Simon has spent the past decade building workflow products in technology, founding two venture-backed startups with one exit, and now applies those lessons to building scalable AI products in NYC.
- [Writing](https://gufamily.org/writing.md): Thoughts on product development, technology, and faith.

## Writing: Product & Tech

- [All Product & Tech posts](https://gufamily.org/writing/product-tech.md): Insights on product development, technology, and building digital experiences.
- [What Can't Be Bundled](https://gufamily.org/writing/what-cant-be-bundled.md): A strategy guide for founders and incumbents on which AI products the labs will fold into their $20 bundles, using Amazon Basics, Microsoft Teams and the IBM PC as precedents, and which layers stay defensible, such as entitlements, systems of record, licenses, distribution and a company's own decision data. (2026-11-13)
- [A Bitter Lesson in Finding Needles](https://gufamily.org/writing/a-bitter-lesson-in-finding-needles.md): Applies Rich Sutton's Bitter Lesson to building AI products. Hand-written prompts, skills and routing rules age badly, while search primitives that agents call in loops get better with every model, and the post shows how that shapes global payment search for agents at J.P. Morgan. (2026-11-05)
- [Irreducible Context](https://gufamily.org/writing/irreducible-context.md): On AI-written docs and ‘meat proxies’ at work. Using compression and token surprise, it shows a model's output can't carry more than the context it was given, so specs still matter, a person's value is the context only they can add, and PM mentoring should move to the prompt. (2026-10-20)
- [Seeing Is Believing](https://gufamily.org/writing/seeing-is-believing.md): Why high-stakes AI agents should show their work instead of asking to be believed. Drawing on a glass-walled bank vault on Fifth Avenue and Claude Code's live browser pane, it describes a reconciliation layer that sends treasury users to the system-of-record page behind every agent answer. (2026-10-15)
- [Trust Is the Rail](https://gufamily.org/writing/trust-is-the-rail.md): Design principles for AI agents in regulated payments. Payments move at the speed of trust and every extra agent spreads that trust thinner, so the post makes the case for a central orchestrator as clearing house with thin agents, plus a heuristic for when reasoning belongs inside an agent. (2026-07-31)
- [The End of Software Scarcity](https://gufamily.org/writing/the-end-of-software-scarcity.md): What happens to software, products and tech jobs once AI makes code abundant. The post argues abundance is a net positive, that taste and point of view become the differentiator, and that guardrails have to move into infrastructure as policy-as-code. (2025-12-22)
- [Service-as-a-Software](https://gufamily.org/writing/service-as-a-software.md): On AI rollups that buy service businesses instead of selling them software, and the choice every AI startup faces about leaning into probabilistic models, with a case for building evals from day one and an example LlamaParse and LangSmith pipeline. (2025-12-13)
- [The Future of Product Development](https://gufamily.org/writing/future-of-product-development.md): How coding agents like Codex and Claude Code are changing product management, from forward-deployed PMs to context engineering with Markdown plans that act as the real PRD, and why PMs who want to stay relevant need to become more technical. (2025-11-15)

## Writing: Faith

- [All Faith posts](https://gufamily.org/writing/faith.md): Reflections on faith, obedience, and running the race well.
- [How Did You Treat This Person I Gave You?](https://gufamily.org/writing/how-did-you-treat-this-person-i-gave-you.md): A reflection on accountability to God for how we treat the people He places in our lives, starting with marriage. Prompted by Francis Chan's You and Me Forever, it uses the tesseract as a picture of extending God's love from self to spouse to neighbor. (2026-10-09)
- [Wrong Conceptions on Faith](https://gufamily.org/writing/wrong-conceptions-on-faith.md): Challenges the definition of faith as belief in the unseen, shared by believers and atheists alike, and argues from Luke 5:4–10 that faith is acting on God's word in the middle of doubt, the same everyday trust we already place in a chair. (2026-02-22)
- [God and Money](https://gufamily.org/writing/god-and-money.md): A founder's reflection on money, ego and startups through the book God and Money, covering Jonah, serving God over Mammon, building from an eternal mindset, and what he would do differently in a future company, like setting aside equity for God. (2025-12-25)
- [Every Good Endeavor](https://gufamily.org/writing/every-good-endeavor.md): A Christian response to 9-9-6 hustle culture at AI startups. Drawing on Timothy Keller's Every Good Endeavor, it treats work as co-creating with God and serving others, with Sabbath rest as the limit that keeps work from becoming an idol. (2025-10-10)

## AI write-ups

Standalone HTML write-ups at https://ai.gufamily.org.

- [A trainable decision head on a frozen Gemma 4](https://ai.gufamily.org/gemma4-decision-head): A classifier that reads a small local model’s hidden state to answer routing, scoring, and yes/no questions with a probability for every option. It works with no labels, and with 20 labeled examples per option it gained 2.5 to 18.9 points of accuracy across 14 intent benchmarks, including 7 held back until every setting was frozen. (2026-11-11)

## Optional

- [Full text of the site](https://gufamily.org/llms-full.txt): every post, plus the homepage and work history, in one file
- [RSS feed](https://gufamily.org/feed.xml): all posts