Field Notes
OpenAI Build Week: What Would You Ask Sam?
OpenAI emailed me about Build Week with four days left, which meant choosing between several questionable Codex projects and figuring out how to make a meaningful addition to SamNav—an application that was already built, deployed, and working. The answer became “What would you ask Sam?”: a question-first way to navigate Sam Altman’s public writing without pretending to impersonate him. The new system uses SamNav’s existing taxonomy, semantic dossiers, and GPT-5.6 to interpret nuanced questions and plot a grounded route through things he has actually published.
I received an email about OpenAI Build Week on July 17th. Submissions were due July 21st.
The premise was straightforward: build something with Codex and GPT-5.6, document the process, and submit it for judging. Conveniently, I have built a distressing number of things with Codex, nearly all beginning with a sentence that should have caused a responsible adult to close the laptop and go to bed.
There were options. There was the sandwich-betting app, which remains a serious intervention into the social consequences of casual wagering. There was also the new thing I had just started working on, which we will not discuss because it is secret and currently enjoying the brief period when it has no bugs.
SamNav was the obvious choice.
First, it was built with Codex. Second, it was finished. More finished than most of my software ever gets, honestly. It was live, functional, indexed, tested, deployed, and quietly doing the peculiar job it was created to do: making Sam Altman's public blog navigable.
The original blog contains more than a decade of writing and almost no meaningful way to move through it. SamNav adds the missing furniture. It organizes posts by year and topic, provides search, connects related entries, assigns a theatrical Navigation Severity Score, and offers curated reading paths through recurring ideas. It does not republish the posts. It stores directions to them.
Also, Sam has not been posting much lately. I do not know why. Perhaps he has been busy. Someone should check on him.
The contest presented one significant problem: pre-existing projects were eligible, but only if they were meaningfully extended during the submission period. That is reasonable. It is also sometimes harder than starting something new.
An unfinished application is full of obvious invitations. A finished one has logic. Every additional feature has to justify the disruption. SamNav already had a clear thesis and a complete information architecture. Adding an AI chat window because this was an AI contest would have made the site larger and the idea smaller.
The addition needed to feel like the next version of SamNav, not a small machine bolted to its side.
The question finally arrived in almost embarrassingly literal form: What would you ask Sam?
The constraint is the product: he can only answer through his published works.
This is not a Sam Altman simulator. It does not speculate about what he might say, adopt a synthetic persona, or pour his archive into a chatbot and hope the autocomplete behaves. A readers asks a question, and SamNav constructs a route through the existing writing that addresses it. The answer is not generated in Sam's voice. The answer is the arrangement of the source material.
That required a different kind of navigation than the original site performed.
SamNav's original navigation layer was static. Each post has a date, summary, reading time, related entries, Navigation Severity, and a set of topic labels. Those labels work well when the reader already knows the territory. Click AI and receive the AI shelf. Follow the startup path and receive an editorially ordered route through startup writing. Search for fusion and receive records containing fusion.
Human questions are less cooperative.
"How did his thinking about AI change over time?" is not one topic lookup. "Where do energy, abundance, and intelligence intersect?" crosses several shelves. "What does the archive suggest about technological optimism?" may not use the phrase technological optimism anywhere. The reader supplies an intention, not a tag.
The new system has to interpret that intention without inventing evidence. SamNav creates structured semantic dossiers for the existing posts, then combines embeddings with lexical and relational signals to retrieve a bounded shortlist. GPT-5.6 selects and sequences a small set of relevant records from that shortlist. It returns post IDs, route roles, and navigational notes—not authoritative metadata. SamNav validates every ID against its own catalog and supplies the canonical titles, dates, links, and classifications locally.
If the model invents a destination, the route is rejected. If the archive does not contain a defensible answer, the site says so. "No clear route" is better than confident fiction.
The same caution extends below the waterline. The live endpoint accepts bounded questions, screens them with moderation, applies rate limits and timeouts, and fails closed when shared safeguards are unavailable. Full article bodies are used only during the non-public indexing process; live route requests contain a bounded shortlist of public-safe dossiers, and OpenAI requests use store: false. Raw questions are excluded from ordinary logs, analytics, and session replay. Structured responses are validated before anything is rendered, so model output cannot quietly improvise its way around the catalog. None of this makes public software invulnerable. It does mean recklessness was not accepted as a feature.
The result looks like a plotted expedition rather than a conversation: where the idea begins, how it develops, where it becomes complicated, and the latest available evidence. Every stop leads back to the original post. The model interprets the question and draws the route, but it does not get to manufacture the terrain.
That is the meaningful addition. SamNav.ai began as a map for a blog that needed one. Now the reader can describe where they want to go.
We have 3 days to make it presentable.