Insights
Clinical decisions, evidence, and where technology belongs.
Writing from a practising oncologist building systems for the decisions that remain difficult.
All Articles
I Gave Codex a Chalk Talk
On a walk, I taught Codex how to make a presentation and learned from the exchange. A study of developers asks what stays with us after AI helps finish the work.
How Much Benefit Would Make This Worth It?
At a case conference, chemotherapy seemed obvious to many colleagues. I thought the patient had a meaningful choice. What was inside that disagreement?
Ten Billion Tokens Later
I switched to OpenAI on July 13 and reached 10.1 billion tokens 48 days later. The work behind that number changed how I lead agents and build medical knowledge.
Where Is the Clinician Baseline Table?
A dermatology study measured decisions before and after advice from artificial intelligence. It raises a question that job titles do not answer: what can the intended user do without the system?
Why Can’t You Make Your Stuff Better?
A Canadian can pay full price for a machine and still be blocked from making it better. That belongs in Canada’s trade fight.
The Machine Got the Right Answer Without Doing All the Work
A Nyx run reached the clinical conclusion we kept and still failed. Missing support links show why a sound answer may still be unfinished work.
The Endpoint We Had Already Chosen
PFS formalizes how patients, clinicians, and HTA models understand cancer control, even though it does not reliably predict overall survival. Its clinical meaning holds only when the trial endpoint obeys the same model.
Overall Survival Is the Weaker Endpoint
Overall survival measures an entire treatment pathway. PFS may be the stronger endpoint for the treatment decision a trial was designed to test.
The Manager Was Set to Medium. The Meeting Took 13 Million Tokens
The specialists needed XHigh reasoning. Their orchestrators ran best at Medium. Another workflow used 13.5 million tokens on alignment alone.
The Human in the Loop Is a Regulatory Fiction
In oncology, a clinician cannot inspect, verify, or control a generative model. Physician review cannot become a liability shield for the companies that put it into clinical use.
The Claim Stops Where the Evidence Stops
A 9,691-patient trial shows why better documentation, better decisions, and better outcomes require separate evidence.
I Found Out About Build Week on Friday Night
I found OpenAI Build Week on Friday night, built Ariad by Tuesday, and learned why fast clinical products still need reliable medical knowledge.
The Treatment He Could Not Get Still Belonged in the Conversation
An Ontario oncologist discussed a treatment the system could not yet provide. That conversation matters for patients, clinical records, and funding decisions.
The Hardest Decision in Oncology Is the One You Never Get to Check
Why the survival benefit number from an adjuvant trial is never the number a patient is asking for — and where decision support can close that gap.
Canada Is Building the Foundation. The Value Is One Layer Up.
Canada is building sovereign AI infrastructure for health. The durable value will be made in the layer that turns it into better decisions at the bedside.
The Doctors Who Use AI Were Never Worried About It
AI researchers are walking back doomsday predictions. In oncology the reversal is unsurprising — these tools help with what clinicians never had time to finish.
We Turn Your Survival Curves Into Yes or No
Pharma invests years in continuous trial data. By the time it reaches clinic, most of that precision is gone. Here's where the value actually goes.
What Oncologists Actually Want from Clinical Decision Support
Most CDS tools assume doctors need help remembering guidelines. That's not where oncologists get stuck. Here's what the harder conversations actually look like.
Stay in the loop
New writing on clinical decisions, evidence, and where technology belongs. Updates are occasional and you can opt out at any time.
Your email is handled under our Privacy Policy.