At some point on Friday night, I discovered that OpenAI was running Build Week. The challenge had opened on Monday, more than 40,000 people had already registered, and submissions were due Tuesday. I would need to build something with Codex, write up the project, provide the code and make a YouTube demo. The sensible response would have been to wait for the next one. Naturally, I signed up. I treated it like an insane challenge: okay, let’s see what I can do.
I chose the easy idea
I started with a practical question. What could I build quickly that might actually help me in clinic? I did not want something that depended on a new clinical trial, a controversial medical idea or an elaborate set of numbers. I wanted a problem where the medicine was already reasonably well known and the main barriers were simplicity and availability. Treatment side effects seemed perfect. This was, in retrospect, my first optimistic assumption.
The idea became Ariad, a patient-facing tool that would present treatment side effects in simple language. It would break a cancer regimen into its component drugs, make those drugs searchable through their side effects and help translate the way a patient describes a problem into a medical symptom the system could recognize. Patients do not necessarily say “peripheral neuropathy.” They say their fingertips feel buzzy and they keep dropping small objects. Ariad was meant to connect those descriptions without pretending to diagnose what was happening.
Codex with GPT-5.6 Sol helped me build the application, and the building was fast. Very fast. Inside Ariad, GPT-5.6 Luna used the API for the narrower job of mapping a patient’s description to a medical symptom concept. That part felt like exactly what AI should be good at: helping a person find the right information without requiring them to speak like a medical textbook.
Codex became extremely concerned about diarrhea
The conversations with Codex were a different matter. As a clinician builder, I move constantly between medicine and product design. Is this a clinical decision? Is it wording? Navigation? A safety boundary? Often it is several of those at once, and Codex did not always appreciate the distinction.
At one point, it wanted Ariad to say that the symptoms a patient was experiencing could not be causally linked to the drug they were taking. Strictly speaking, Ariad cannot determine what caused an individual patient’s symptom, so that part was fair. But there is a difference between diagnosing causation and acknowledging a well-established treatment toxicity.
Severe diarrhea in a patient taking capecitabine is not an exotic theoretical connection. The FDA prescribing information specifically warns about diarrhea, and Cancer Care Ontario’s patient information tells patients that capecitabine can cause it. Fever during chemotherapy creates a similar problem. Ariad cannot diagnose why someone has a fever, but a tool that becomes too nervous to communicate its importance is not being helpful. BC Cancer advises patients with a fever during chemotherapy to go directly to an emergency department.
So I found myself having slightly surreal arguments with an AI that could help build a sophisticated patient-facing application in a few days but became alarmed by the idea that capecitabine and diarrhea might be related. I had deliberately chosen the non-controversial project. Apparently, that was optimistic too.
The interface outran the medicine
I was using information from the FDA, Health Canada, BC Cancer and Cancer Care Ontario. These were not random internet searches or medical advice generated from scratch. But credible sources do not magically become a complete clinical product. The information still has to be extracted, organized and reconciled. Regimens have to be connected to their component drugs. Side effects need consistent definitions. Important warnings cannot disappear during translation. Most importantly, the knowledge has to be sufficiently complete that a patient does not receive an excellent answer in one pathway and an empty space in the next.
I knew I was not going to finish that medical foundation before the deadline. I needed to keep working on Ariad while constantly reminding myself that the knowledge underneath it would not be ready in time. That is a strange feature of building with AI: the visible product can take shape much faster than the least visible—and most important—part of it.
At 8:00 p.m. on Tuesday, the last possible moment, I hit submit. Then I started running through cases as an oncologist rather than as someone trying to finish a Build Week entry. Would this give my patients the help and support they need? The honest answer was no. Not yet.
Ariad could demonstrate the experience I wanted to create. It could show side effects in simpler language, connect regimens with their drugs and map a patient’s description to a symptom. But the medical knowledge was not present to the depth and completeness required for clinical use. This was a Build Week submission, not something I could deploy for patients.
Nyx first
That was when I understood how much I wished Nyx had been involved. To be clear, it was not involved in building Ariad at all. But if Nyx were fully operational, the underlying medical knowledge would exist independently of Ariad. The relationships between diseases, regimens, drugs, side effects and supporting sources could be established, reviewed and maintained before a product tried to present them.
I could have spent Build Week on Ariad’s actual job: transforming reliable medical knowledge into language and navigation that work for patients. Instead, I was trying to build the product and its medical foundation at the same time.
Once I saw that, it became difficult not to see it everywhere. The same medical knowledge might eventually be needed in a clinician-facing product, by a pharmaceutical company entering a new disease area, or by government and health technology assessment bodies evaluating new submissions. The context and decisions will differ, but the underlying medicine should not have to be rebuilt from scratch every time—or change depending on who happened to build it.
That is what I mean when I call Nyx an operating system. The medical knowledge and the decisions around it need to exist independently of any one application, so different products can use that foundation without quietly reinventing it.
Ariad is now being judged as a Build Week project, and it is not available yet. Once judging is complete and I am able to work on it again, I intend to return to it. But Ariad will only be built to clinical grade after we have the Nyx foundation it needs.
I signed up on Friday night to see what I could build. The answer was: quite a lot, remarkably quickly. I also learned what I should not try to build ad hoc, made a YouTube video at the last possible minute and added a substantially larger project to my list.
Would I do it again? Obviously.