The Cognitive Revolution · 2025-11-17 · 123 min

Using AI to navigate son's cancer diagnosis

Erik TorenbergNathan Labenz

YouTube ↗

TL;DR

AI changed the trajectory of Nathan Labenz’s response to his six-year-old son’s rapidly advancing Burkitt leukemia. Human clinicians twice discounted an LDH reading around 1,500 because the samples showed hemolysis and said they had “ruled out anything life-threatening”; GPT-5 Pro, Claude 4.1 Opus, Gemini 2.5 and Grok 4 instead treated it as a red flag requiring action within 24–48 hours. Labenz’s strongest practical conclusion is categorical: in a serious medical crisis, “use both human and AI doctors aggressively.”
The outcome so far is close to the best case available for an otherwise terrifying disease. Burkitt leukemia can double in as little as 24 hours, and Ernie’s disease included a large abdominal mass, widespread lesions, bone-marrow involvement and possibly ambiguous spinal-fluid involvement. Roughly 10 days after treatment began, the main mass had shrunk about 80%, many lesions had disappeared, and the latest marrow and spinal-fluid tests were clear—making cure, rather than chronic management, the overwhelmingly likely outcome.
The product lesson is that model choice, tone and context presentation can alter real-world decisions even when the underlying data are identical. Claude 4.1 Opus was so alarmist that it helped trigger a panic response, while GPT-5 Pro delivered essentially the same warning in a longer, more clinical voice that Labenz preferred; yet he concedes Claude’s urgency “might have been what I needed.” Regular GPT-5 also repeated the doctors’ error when portal screenshots foregrounded the hemolysis warning, showing why users should seek “n opinions,” vary the prompt and inspect the totality rather than trust a single output.
At bedside, GPT-5 Pro functioned less like a diagnosis app than a continuously available clinical analyst. Labenz fed it nearly 20 rounds of labs and updates over nine or 10 days, using it to validate the treatment protocol, interpret changing vitals and question mineral replacement, blood-pressure medication and additional testing. It could not examine Ernie, prescribe drugs or perform procedures, but it gave the family enough understanding to advocate intelligently and enough agreement with clinicians to trust that the local team was delivering state-of-the-art care.
AI’s most differentiated contribution may be personalized synthesis across specialties and beyond standard workflows. When a basement flood created mold concerns during immunosuppressive chemotherapy, GPT-5 Pro combined the child’s oncology context with air-test and remediation data—an analysis that would otherwise require arranging a meeting between “your oncologist and your mold guy.” It supported a concrete plan involving remediation precautions and eight to 10 HEPA filters, reducing both health anxiety and the perceived need to leave the house.
Labenz intends to depart from standard follow-up by pursuing circulating tumor DNA, or ctDNA, to detect a relapse earlier than symptoms or scans might. The proposed workflow fingerprints stored tumor tissue, then looks for that signature in blood at sensitivity he believes may reach five cancer cells per million; his oncologists regard it as unproven or nonstandard. His reasoning is that a 24-hour doubling cancer should be attacked at the smallest detectable burden, both to reduce treatment-linked cytokine storm and to give the cancer fewer “at bats” to evolve drug resistance: “If ctDNA comes back positive on a Tuesday, you want to be infusing one of these next-generation drugs on a Thursday.”
The revealed willingness to pay makes healthcare one of the clearest demonstrations of AI consumer surplus, while exposing the regulatory bottleneck ahead. Labenz pays $200 a month for GPT-5 Pro against an estimated $500,000–$1.5 million six-month treatment bill and says he would willingly pay $10,000 a month in this crisis; in his phrase, the economics are both “GDP destroying” and “absolutely insane.” He expects AI-supported patients to challenge standard-of-care rules, clinical-trial access, liability doctrine and poor data portability—while still supporting a ban on uncontrolled superintelligence because faster medical progress and existential-risk governance are not, in his view, contradictory. The experience also left him feeling there are “no decels in the pediatric oncology unit”: delays in AI progress have real human costs.

Digest · the substance, structured for research

1. A terrifying diagnosis quickly produced a best-case response

Nathan Labenz opens with the fact that reordered everything else: his previously healthy six-year-old son, Ernie, has Burkitt leukemia. Doctors initially called it Burkitt lymphoma, but bone-marrow involvement changed the classification as the family understood it.
The disease’s scale was frightening—a large abdominal mass, lesions throughout the abdomen and above the diaphragm, marrow involvement and possible, still-ambiguous spinal-fluid involvement. Burkitt can have a doubling time “as fast as 24 hours,” making delays unusually consequential.
The counterweight is equally important: this is among the cancers most responsive to chemotherapy. After the initial treatment wave, the central mass was down roughly 80%, many colonies had vanished, and the latest marrow and spinal-fluid tests were clear.
Labenz waited for that milestone before speaking publicly. With the response “about as well as one could hope,” the overwhelming expectation is cure: the cancer disappears, never returns and Ernie lives a long, normal life.

2. Cancer made the abstraction of exponential growth brutally concrete

Labenz’s analogy joins the subject of his podcast to the family crisis: both AI capability and cancer can look “flat behind you and vertical in front of you.” One malignant B cell began dividing out of control until seemingly minor symptoms crossed thresholds into a crisis.
Looking backward, the family cannot know whether a week of stomach pain in early September was the first “little murmur” of the disease. A doctor’s constipation explanation appeared to fit, prune juice seemed to help, and the complaint disappeared.
Knee pain after a trampoline-and-bounce-house weekend also had an obvious benign explanation. Amy worried sooner and even considered leukemia, partly through her own AI use; Nathan’s baseline response was closer to “Yeah, he’ll be fine,” as it had always been before.
By the October 22 vacation, fatigue, knee pain and the earlier stomach complaint still each had plausible explanations. Within days, however, Ernie was restless at night, moaning, intermittently screaming in agony and reporting pain in his tooth, head and multiple other places.

3. Episodic illness repeatedly disappeared inside the clinic

On October 24, after participating fairly normally in a Louisiana food festival, Ernie screamed in pain during the drive home and asked to go to the hospital—then chose to wait until morning. An October 25 urgent-care knee X-ray was normal, and staff could not complete a blood draw.
The mismatch mattered: clinicians saw a calm child with a tablet, not the child his parents had watched screaming. Labenz’s practical advice is to film the worst episodes because “there was a disconnect between what we had seen and the way he was presenting.”
He also recommends producing a written symptom chronology once a confusing pattern becomes concerning. Giving a clinician the entire history “in black and white” forces a survey-level view before any one symptom pulls the visit down another benign rabbit hole.
Labenz adds a third retrospective recommendation: record major appointments and use AI to transcribe them. Preliminary interpretations often arrived through short conversations without a formal report, leaving the parents with less precise information than the care team possessed.

4. One discounted lab value became the hinge of the case

At Children’s Hospital New Orleans, nearly every blood count and chemistry result looked normal except lactate dehydrogenase, or LDH, at roughly 1,500—around three times the cited reference range. LDH reflects cell turnover and can rise when cancer cells are proliferating and dying.
Both samples carried a hemolysis warning because cells had begun breaking down in the specimen. Clinicians treated the elevated LDH as an artifact and told Amy they had “ruled out anything life-threatening,” while recommending routine pediatric follow-up after the family returned home.
Back in Detroit, Labenz manually copied every portal result and assembled the parents’ full chronology. He trusted himself more than an agent for this one-off, high-stakes extraction: “This is N of one,” and accuracy mattered more than automating a repetitive task.
GPT-5 Pro, Claude 4.1 Opus, Gemini 2.5 and Grok 4 varied in tone but agreed on the substance: hemolysis might raise LDH, probably not that much, and the family should obtain a proper workup within 24–48 hours. Leukemia was not necessarily most likely, but it urgently needed exclusion.

5. The same evidence produced materially different AI behavior

Claude 4.1 Opus responded as though leukemia were the likely explanation and used language Labenz found emotionally overwhelming. It precipitated something like a mild panic attack and pushed him toward GPT-5 Pro’s longer, more clinical and less emotional style afterward.
His judgment remains deliberately unresolved: “Claude might have it right.” For a father inclined to assume things will pass, the model’s alarm may have been precisely what moved him out of denial quickly enough for a cancer capable of doubling daily.
Amy’s regular GPT-5 made the opposite error when she supplied screenshots containing the portal’s hemolysis disclaimer. It discounted the LDH much as the physicians had, whereas Nathan’s text prompt presented the same caveat in the parents’ narrative and elicited an urgent warning.
The resulting method is not “ask AI once.” Labenz recommends the best available models, Pro for anything that truly matters, multiple formulations of the evidence and comparison across outputs: AI makes “n opinions” practical, but presentation can still steer an individual answer badly.

6. Parent advocacy and a physical exam finally exposed the mass

Their pediatrician still considered a self-resolving childhood complaint most likely, but called Children’s Hospital of Michigan himself and secured an oncology appointment the next morning. The oncologist initially favored an autoimmune explanation over cancer.
The decisive moment came when she palpated Ernie’s lower-left abdomen and he recoiled. Amy and Nathan immediately reported that the pediatrician had elicited the same reaction the prior day, turning a detail that might have been forgotten into a repeated physical finding.
Labenz’s hindsight is that AI should have interviewed the parents for missing information and guided them through a parent-observed physical check. Models cannot replace an exam, but they could have helped identify where touch produced a uniquely abnormal response and made that evidence legible to clinicians.
Amy then pressed to obtain an ultrasound immediately rather than accept a later appointment. The October 31 scan found a roughly 5 cm by 6 cm by 4 cm abdominal mass, with possible liver, spleen and kidney involvement; once the report reached the portal, AI made the seriousness difficult to deny.

7. Most-likely thinking underweighted the fastest plausible disease

An MRI was scheduled for Monday under general anesthesia. Over the weekend, clinicians and AI both leaned toward neuroblastoma, described as slower-moving on a weeks-to-months scale, and advised that waiting a couple of controlled days was reasonable if pain and fever stayed manageable.
Labenz now thinks the system anchored too heavily on that leading hypothesis. AI estimates at the time put neuroblastoma around 30% and lymphoma around 25%; the smaller probability was not remotely small enough to ignore when its time sensitivity was radically greater.
Monday’s MRI led to a Tuesday biopsy, liver sampling, marrow aspiration and other procedures, followed by direct admission. The scheduling was impressively agile, yet Ernie’s decline showed how a system can move quickly by normal standards and still nearly lose a race against a 24-hour exponential.
The broader lesson is not that every uncertain case demands maximum intervention. It is that high uncertainty should be tested against the worst plausible trajectory, especially when the consequences of waiting are asymmetric and weekends materially slow the medical system.

8. Clinicians broke protocol when the trajectory outran pathology

By November 5, Ernie was swollen, profoundly weak and barely leaving bed while the team waited for definitive pathology. The doctors’ refrain—“We need to get to the diagnosis”—reflected the legitimate need to match treatment to cancer subtype.
Nathan’s pushback was temporal: if results slipped through the weekend, Monday was five days away, and “five days from now seems like a really long time.” When Ernie then said he could not move his legs, the bedside oncologist’s concern visibly jumped.
Surgical appearance had already made neuroblastoma unlikely and lymphoma increasingly probable. The oncologist returned ready to “break the rules a little bit”: candidate lymphoma regimens were similar enough to start treatment under imperfect information and refine it when the reports arrived.
A steroid began Wednesday, followed by conventional chemotherapy Thursday; the working diagnosis reported Friday, November 7, was Burkitt lymphoma. The family’s subsequent understanding was Burkitt leukemia because of marrow involvement, while further genetic and subtype testing remained pending. Labenz says he does not know the counterfactual timeline, though it seemed plausible that a few more days could have brought organ failure or death.

9. Fast growth created both the danger and the therapeutic opportunity

The pathology reframed the outlook: “super fast growing, super aggressive, but also super treatable.” Chemotherapy targets cells actively growing and dividing, so Burkitt’s defining liability also makes it unusually responsive compared with slower cancers that can be harder to kill.
Treatment still begins cautiously because rapid tumor destruction can itself overwhelm the body through tumor lysis syndrome. The prephase, or debulking phase, was intentionally less intense and nevertheless reduced Ernie’s cancer volume by about 80%.
The planned six-month roadmap then moves through two hard-hitting phases, two moderate “mop-up” rounds and two milder consolidation rounds. The goal is every last cell: residual disease can restart the exponential, mutate under selection and return in a much harder-to-treat form.
Labenz is equally careful not to generalize this pathway to all cancer. For this particular disease, diagnosis converted panic into a managed schedule because a highly effective, protocolized treatment exists; many slower-growing, rarer or surgery-dependent cancers pose entirely different problems.

10. GPT-5 Pro became a continuous bedside second opinion

The first parental lever after diagnosis was nonstop verification of the team’s work. Blood was drawn as often as every six hours, medications and vitals changed rapidly, and fluid treatment once caused a nearly 50-pound child to urinate close to a gallon within several hours.
Labenz maintained a long-running GPT-5 Pro thread containing symptoms, medications, reports and successive portal PDFs—close to 20 uploads over nine or 10 days. Each update asked what the results meant, what clinicians might be missing and whether the proposed response made sense.
The model repeatedly confirmed that the main protocol was state of the art and highly standardized. That helped answer whether the family should relocate to a more prestigious center: for this known disease and established regimen, the analysis said they were already receiving the care a competent leading center would provide.
Its value was therefore partly adversarial and partly reassuring. Rather than constant Googling and free-floating anxiety, Labenz gained enough mechanistic understanding to advocate when necessary and enough independent confirmation to believe the local team’s judgment was sound.

11. Productive disagreement improved questions without displacing doctors

GPT-5 Pro was somewhat more eager than clinicians to replenish calcium and magnesium when fluid dilution pushed them low. The family’s advocacy may have moved supplementation earlier on the margin, though Labenz characterizes the difference as small rather than a discovered medical error.
When Ernie’s blood pressure barely crossed the treatment threshold, GPT-5 Pro favored remeasurement and watchful waiting. The physicians preferred medication; armed with the model’s explanation, Labenz questioned them, found their reasoning compelling and said, “I trust your judgment. Let’s go for it.”
The model also proposed extra tests for abnormal liver enzymes that doctors considered predictable after a liver biopsy. Again, their disagreement was marginal: AI wanted more information, while clinicians could place the number inside a familiar pattern and avoid unnecessary investigation.
Labenz’s boundary is clear: AI cannot perform procedures or exams, order tests or prescribe. Its role was to make the parents competent participants in a human care system—and, when human explanations survived informed questioning, to strengthen rather than weaken trust.

12. Long context turned fragmented records into a usable case model

The recurring prompt grew toward 100 pages: parental history, MRI, CT and PET reports, spinal-fluid and marrow analyses, pathology, cellular-marker percentages and other PDFs. Labenz’s experience is that current models handle such long context well enough that old fears about “overwhelming” them no longer fit.
He also asked the bedside thread to generate a dense handoff document “for a new attending physician” capable of supporting cutting-edge care. Compressing the accumulated history into a few thousand structured words made it reusable in fresh model sessions without discarding the medical through-line.
Data acquisition remained the bottleneck. Important results arrived inconsistently through the portal, paper printouts or documents apparently degraded by faxing; phone photos of irregular tables demanded manual checking because a subtle transcription error could change which markers appeared on the cancer cells.
For crisp website screenshots, direct multimodal input generally worked. For difficult photographed documents, Labenz found Gemini 2.5 Pro better than GPT-5 Pro at transcription, though still not reliable enough to skip line-by-line verification in a high-stakes case.

13. AI connected oncology to a household mold problem no specialist owned

The second parental lever was preventing infection while chemotherapy suppresses Ernie’s immune system. Viral illness concerned the team, but bacterial and fungal infections could be more dangerous, including organisms already present in the mouth, skin or gut.
The family could adopt a temporary COVID-style exposure protocol because both parents work at home and only one other child attends school. Nathan suspended AI-event travel and on-site speaking, at least for several months of the six-month treatment roadmap.
A basement flood added a cross-domain problem: about two inches of water arrived just as the family returned from New Orleans, raising mold concerns for an immunocompromised child. GPT-5 Pro helped identify local remediation companies, interpret comparative indoor and outdoor air samples, and connect those findings to Ernie’s specific oncology risk.
The resulting plan was to isolate Ernie from active basement work, perhaps stay with family for several days, and run roughly eight to 10 new HEPA filters throughout the house. Slight basement elevation, clean results elsewhere and filtration effectiveness gave Labenz confidence that permanent relocation was unnecessary.

14. Relapse planning surfaced ctDNA before the doctors recommended it

The third lever is preparation for the smaller but devastating outcome in which the cancer returns. Relapse usually happens within six months and almost always within two years; if it occurs, Labenz says bluntly that most affected children die, though not all.
He is gathering the still-pending genetic subtype, remote second opinions from perhaps two leading centers and information about their trials. Formal reviews costing under $1,000 could both validate current care and establish relationships before an emergency requires CAR-T cells, bispecific immunotherapies or another experimental line.
The proposed nonstandard step is circulating tumor DNA. A laboratory would fingerprint stored tumor tissue, then test blood for that exact signature—at sensitivity Labenz believes may be around five cells per million—during treatment and surveillance.
Clinicians acknowledged ctDNA but called it unproven or questioned what they would do with a positive result before symptoms or imaging. Labenz’s answer is preparedness: ordinary monthly blood work and quarterly scans often detect relapse only after sickness or substantial growth, which is especially inadequate for a daily-doubling cancer.

15. Early detection links relapse biology, model competition and medical reform

The mechanistic case for ctDNA is that next-generation treatments can trigger cytokine storm, whose severity appears to rise with disease burden. Fewer cancer cells should also mean fewer “at bats” for evolution to discover another drug-escape mechanism; Labenz preserves the uncertainty but finds the causal argument compelling.
A model he believed to be a Gemini 3 checkpoint sharpened that plan with responses he found five to 10 times faster, shorter and more “pitch perfect” than GPT-5 Pro. He still used GPT-5 Pro as a checker and had to request citations explicitly, but within hours Gemini had become the output he most wanted to read.
The economics behind that preference are extreme: $200 a month for GPT-5 Pro beside an estimated $500,000–$1.5 million treatment bill. Labenz compares the subscription cost with somewhere between under a tenth of a percent and 2% of the total medical cost, calls the surplus “absolutely insane,” says he would pay $10,000 monthly in this situation, and argues that avoided complications, infections or unnecessary relocation can even make AI “GDP destroying.”
His policy conclusion cuts both ways. Delaying an AI oncology researcher costs lives, yet he still supports banning uncontrolled superintelligence; meanwhile, near-term medicine needs better data access, liability reform and a stronger right to try because AI-supported patients may increasingly bring well-reasoned ideas that could beat static standard-of-care rules.
Labenz’s preferred geopolitical race is therefore a “cancer cure Olympics” between the United States and China, extending eventually to healthy lifespans of 150 years. His closing claim is not that every AI proposal is right, but that “standard of care is not the end of history”—and he argues that the medical system’s current structures are becoming handcuffs. The experience also left him feeling there are “no decels in the pediatric oncology unit”: the costs of delaying useful AI progress are unusually vivid when a child’s life is at stake.

Nathan Labenz

Hello and welcome back to The Cognitive Revolution. Today’s episode is going to be a little bit different—by far the most personal episode that I’ve ever done. It’s definitely not an episode I ever expected or wanted to be doing, but here we are. It’ll be a little raw. I’ve taken a bunch of notes and prepared for this, but I really haven’t scripted it out, so it’s going to be a lot of recollection mixed in with practical anecdotes about AI, and we’ll see where it all goes.

The title says it all from a family standpoint: my son has cancer. Six-year-old Ernie is an otherwise totally healthy kid. Everything was very normal. We’ve gone back and looked at the history in quite a lot of detail at this point, and as of October 11, he had a huge weekend jumping up and down on trampolines and in bounce houses, just having a great time. He seemed totally normal, and we never could have expected at that time how quickly things would have gotten really serious—and, in all honesty, at one point, extremely scary—in just the 2½ to 3 weeks following that time.

The good news is—I won’t keep you in suspense—that he has been diagnosed with Burkitt leukemia. He was initially diagnosed with Burkitt lymphoma, but they did find some of it in the bone marrow, and that makes it Burkitt leukemia, as we now understand it. This is one of the fastest-growing cancers of all cancers. It has a doubling time as fast as 24 hours, which is really crazy.

The good news, though, on the flip side of that, is that it is also one of the most responsive cancers to treatment of any known cancer. Because the situation was getting pretty scary and his condition did seem to be worsening by the day, they actually started chemotherapy at Children’s Hospital of Michigan, in our hometown of Detroit, a day before getting the final pathology report back and having what they considered to be a definitive diagnosis. They felt that time—and we, as parents, were very much pushing for this at the time as well—was really of the essence, and it was important to get him into treatment, even if we didn’t have complete certainty about what was going on at that point. It was pretty clear that it was cancer.

The good news is that it’s now been about 10 days, and he has responded extremely well—fortunately, basically a best-case scenario in terms of his response. He had a big mass in his belly, lesions all over his abdomen, some even above the diaphragm, and cancer in the bone marrow, possibly in the cerebrospinal fluid, although that’s still a little ambiguous. But after the first wave of chemotherapy, there was something like an 80% reduction in the central mass of the cancer, and lots of the lesions and colonies that it had established around the body were totally gone. Not all of them are gone; there are still a few left.

The best update that we got in the last 48 hours before recording this is that, as of the last test, his bone marrow and spinal fluid were clear. That’s basically a best-case-scenario response to treatment, and it means that, at this point, the overwhelming odds are that he will be cured—which means the cancer will go away, it’ll never come back, and he’ll go on to live a hopefully long, totally normal, healthy life.

If that weren’t the prognosis, I don’t know that I would be sitting here recording this now. I really wanted to get to this milestone before sharing anything. Very fortunately for us and for him, he does seem to be responding about as well as one could hope.

So that’s the situation. Now I’m going to try to tell you what I’ve learned about using AI in a medical crisis situation. I’ve said many times that if it were really serious, I would use both human and AI doctors aggressively. That is what I did, and I think we definitely got a ton of value from it. I also learned a few things about how we could have used it better, and certainly some ways in which AI, like human doctors, is not perfect. Nevertheless, the bottom line is just an unbelievable amount of value.

I hope this will be interesting to the AI-obsessed regular listener, but I also hope it will be valuable to other people who maybe don’t have as much AI experience and find themselves confronted with a medical crisis for themselves or for a family member, as we just have, and are thinking, “How can I make the best use of AI?”

The first observation I want to make is just that exponentials are crazy. People say this a lot in AI, right? Where are we on the exponential? It looks flat behind you and vertical in front of you, regardless of where you are.

Cancer is also an exponential. There are obviously a ton of different kinds of cancers, a ton of different mechanisms, and a ton of different nuances to all the many different kinds of cancer. That’s really an important caveat that’ll come up again and again throughout this narrative. But at heart, cancer is the out-of-control growth of cells. That’s how I explain it to my son.

We had talked about cells before this, so he knows he has little microscopic cells in his body, and he knows that there are different kinds. I’ve explained that there are skin cells, bone cells, cells that make up your eyeballs, and cells that make up all the different parts of you—all those many different kinds of cells. One cell, which in his case was a B cell, part of the immune system, started growing out of control, and that is what started to cause him all of his problems.

This is an exponential process, right? When one of these cells becomes cancerous and enters into this out-of-control growth phase, it just divides and divides. They say that the doubling time of this Burkitt leukemia is as little as 24 hours, which shows that, in an exponential like that, things move very quickly. What seemed to be fine, or starting to get mildly concerning, can very quickly turn the corner into a real crisis, and that was a very surreal experience to live through.

It is most extreme in the kind of cancer that he has. The doubling time is certainly much lower in other kinds of cancer. Again, the flip side of that is that, because most of the chemotherapy drugs available today target the growth mechanisms of the cell and specifically kill cells that are in the process of growing and dividing, this type of cancer is very responsive to that treatment.

Other kinds of cancer that grow more slowly are less responsive. You may have more time to figure things out, but you also have a harder treatment challenge, it seems, than the one my son was facing. Exponentials are really a strange thing to grapple with.

We honestly don’t know when it started. When we look back at the beginning of the school year—let’s say early September, just 2 months ago—there was 1 week or so when he was complaining about his tummy hurting. It didn’t seem to be bothering him too much, but he did complain about it a number of times.

One day, I got a call from school, and they said it was bothering him. We’d heard that enough times, so I said, “Okay, I’ll pick him up and take him to the doctor.” They said at the time that it was probably constipation. Kids going back to school this time of year often don’t want to use the bathroom, or whatever. They told us to give him some prune juice and said he should be fine.

We did that, and it seemed like he was fine. It seemed like the complaint of the tummy ache went away. Now we wonder: Was that the very beginning of this exponential? Was it just a little murmur, the sort of equivalent of seeing some of these emergent AI capabilities pop up, but they’re not as big of a deal as maybe that 1 symptom suggested? Or maybe the whole thing was actually a much bigger deal than that little preview suggested.

We still don’t know if that was actually related to this, and I don’t think we ever will. The next symptom he had was knee trouble. He was complaining of pain in his knee, and because he started complaining about that after this big weekend with tons of bounce-house and trampoline activity, we basically again felt like we had a really good explanation for it. He had just done hundreds of jumps on trampolines and bounce houses over the weekend. He probably just hurt his knee a little bit.

He was complaining about it for a little longer than it seemed like we would expect. He hadn’t really complained about a knee like that before, and he’d certainly run, jumped, and done all kinds of things. So it seemed like it was lingering.

I would say my wife was definitely more concerned sooner than I was. I’m generally the kind of person who says, “Yeah, he’ll be fine.” And he’s always been fine. She worries a lot more and is much more inclined to take him to the doctor, and she did take him to the doctor with this complaint about his hurt knee.

Even at that time, she had it in her mind that maybe this was leukemia. I think, to some degree, that notion came from AI. She’s definitely a regular AI user, as I am, and that was the first occurrence of the thought that maybe something was really wrong. That was about 4 weeks ago from when I’m talking to you.

Nevertheless, the doctor said, “No, I think it’s fine. The most common cause of this sort of thing is indeed jumping on the trampoline. Come back in a week if it’s not better.”

We happened to have had a family vacation scheduled, and a couple of days after that, we went on vacation. The last episode that I recorded for the podcast before recording this one was on October 21. I had this vacation planned and happened to record a bunch of episodes, so I had them ready to go.

In the intervening weeks, we’ve just been publishing things that I recorded on October 21 or before. That turned out to be fortunate—not that I would have prioritized the podcast over my son’s health in any event, but we happened to have enough inventory that we were able to keep putting episodes out even while I was dealing with all of this.

We left for vacation on October 22nd, and at that time, I would say we did not have any reason to be really concerned. My wife had this idea that maybe it could be something like this. She, in all honesty, is a worrier. Obviously, many parents are about their kids, and that is appropriate. I think definitely one of the big themes of the experience is that you do need to advocate for the patient, whether it's yourself or your kid, in the process.

If you really feel in your gut that something is wrong, that is a powerful signal that doctors don't always listen to, but they should. And yet, I would say, at least for me, I think she had a little bit of a different view by that time. But at least for me, when we left for vacation on October 22nd, I didn't have any sense that there was anything really wrong. The tummy thing had passed. The leg was still bothering him a bit, but I thought maybe he had tweaked it a little bit more than we thought. I was like, "If we come back and it's still bothering him, we'll have to maybe go get a knee MRI or something."

He was also tired the week before we went on that trip, but we'd had a big week. My brother was in town. We stayed up late a couple nights. There were a few different things that sort of explained why he would be tired. So, in my judgment, at the time that we left for vacation, there was really nothing to worry about.

But I'm pretty confident that at the time we did leave for vacation, in fact, the cancer was underway. And with the exponential growth of the cancer in his body, we started to hit threshold points where it started to become undeniable that, okay, something is going on here. It came in waves. It wasn't all the time, but at times he was just really in pain—not sleeping well, moaning in the night, restless, not able to sleep, clearly uncomfortable at night, and better at times during the day. Motrin treated it well, so it wasn't super severe pain, but there were some times when he was just screaming in pain, in agony, and it was like, "Okay, we've never seen a behavior like this before."

It was October 24th, a Friday, and we were driving home from a big day's worth of activities that he had participated in mostly pretty well. He had a decent time and was even playing on the playground and running around a bit. We went to a food festival in Larose, Louisiana, and it was a fun time; he enjoyed it. But on the way home—oh my God—I had just never heard any kid scream in pain like he was crying out. He was just like, "My tooth, my head." It seemed like the pain came from a bunch of different places.

That is, as we've learned, a classic sign of a potential cancer. The tooth—we were like, "Maybe he has a cavity." I don't think that was it either. Somehow, the body is registering pain in all these different places. But he was clearly so bad in that moment, crying out, screaming, so upset, and even asking to be taken to the hospital.

We were like, "Should we take you to the hospital now?" My wife was like, "Let's take him to the hospital now." We asked him, "Do you want to be taken to the hospital now?" He said, "No, take me in the morning." We trusted that, for lack of a better way to decide, and in the morning took him to an urgent care. At the urgent care, they did an X-ray of the knee, so that looked fine. They kind of checked him over, and by that time he was much calmer and seemed much more normal.

I'd say one of the tips to parents in general navigating a mysterious medical situation would be to take video of their worst moments and show that video to the doctor, because I think there was a disconnect between what we had seen and the way he was presenting in the urgent care clinic that day. He just seemed pretty normal, pretty fine. But if the provider had seen that video, they might have thought about it quite differently. The other thing to do is definitely log the symptoms and be pretty thorough about it. We hadn't done this yet at that time, but we started to do this later and then present a medical professional with a detailed history of what had happened.

I don't think you need to do that for every sniffle, obviously. But if you're starting to suspect that something might really be wrong because of a confusing constellation of symptoms, then writing them all out and having the provider read through that in black and white, I think we found to be pretty effective in terms of just getting the case out there, getting them to take the survey-level view before we get into questions or before we start going down any particular rabbit holes on this symptom or that symptom.

But we hadn't done that yet. So, as of the urgent care visit on Saturday, October 25th, it was, "Yeah, the knee looks good. Don't really see anything wrong." They tried to get a blood draw there, but it's not easy to draw blood from a 6-year-old sometimes, and my son certainly did not make it easy on them. Actually, a real proud-dad growth moment as he's gone through this is that he has become much better at taking a blood draw. Even in this extremely difficult time for him, while he's been feeling bad and having all these things happen to him, he has been able to show growth and learning, and now he can stoically sit there, take a poke, and get the blood draw.

But at that time, he had not developed that skill yet. So the folks at the urgent care were like, "Yeah, we can't really take his blood, but we don't really think it's too important. If you want to do that, you probably have to go to Children's Hospital. They're specialists in taking blood." So we left there and went on with our day. He was definitely not feeling good—low energy, not in terrible pain, but there were ups and downs. He was clearly not himself. It was becoming clear that something was going wrong.

But they sent us away without anything too clear, so we carried on for another couple days. Then, on Tuesday, which was our last full day there before flying home again, we were just out trying to enjoy the afternoon, and he was screaming in pain, demanding ice. He wanted to ice his face, and he's never done anything like that before, with all these different things coming up. He was also starting to run a low-grade fever.

We had also been texting with our pediatrician back home to get their take on it. They are really good about engaging with us via text message at any random time. Finally, they said, "I think you should take him to the ER. There's enough stuff going on here at this point. Better safe than sorry."

So we audibled our plans for Tuesday, took him to the ER at Children's Hospital of New Orleans, and they basically came to the same conclusion that the urgent care did. Again, by the time we got there, he had settled down and calmed down and had the tablet in the waiting room. They weren't seeing the screaming-in-pain, "What is wrong with my baby?" kind of behavior that we had seen. They did take blood, and what they found was that basically everything was normal. All the cell counts, all the chemistry, all this kind of stuff pretty much came back normal.

There was just 1 red flag: LDH, lactate dehydrogenase. His was elevated. The reference range, in the units that they use, is a couple hundred to a few hundred. His was up around 1,500, so it's like 3 times the reference range. This is a marker of cell turnover. When cells die, lactate dehydrogenase ends up in the blood. When they measure the level of this LDH in the blood, they use that to infer how much cell turnover is going on in the body.

Nathan Labenz

And some of it is certainly normal; it's never expected to be zero. Our cells are always turning over. But if you have a cancer developing, then that can really elevate it. So that should have been a red flag.

However, it came back from the lab at the hospital with a note that said there was some hemolysis of the sample. What that means is that the lab technician noticed that the cells in the sample were already starting to break down when they did the measurement. When that happens, they seem inclined to throw out that test, because when cells break down, they give this stuff off. When blood gets drawn, after a while the cells start to break down, so the next thing to do is just repeat it.

Well, they did repeat it. They got the same number again, but it came back with the same note again: hemolysis of the sample. Every other test looked good, with this one outlier. Both times they ran it at the children's hospital, it had this note that there was hemolysis of the sample.

You probably don't think about it too much. The doctors there—and it was just my wife, Amy, and Ernie at the hospital at that time—said, “We've ruled out anything life-threatening.” I remember that language very distinctly. “This test, we think, is an artifact of the hemolysis. You should follow up with your pediatrician when you get home, but we don't think there's anything serious going on. We've ruled out anything life-threatening, so you can go home.”

For the moment, we took that and had the last couple of hours in New Orleans that evening. The next day, we flew home. When we got home, this was the first time we had real data. Everything else was: a couple of weeks ago, his tummy was hurting for a week; it seemed like that passed; then the knee; then he's been complaining about the tooth; then he's got this kind of restlessness and apparent discomfort at night; and then he's got this low-grade fever—but it wasn't even a fever. It was 99-point-something degrees, not technically a fever.

And then there were these moments where he seemed to be so unwell, just crying out in pain. But there was no data. What does that story add up to? The doctors weren't making too much of it. My wife, in her gut, I think, had this sense that something was wrong with our baby, which I think is hard for doctors to know what to do with at times.

It means something—mother's intuition, certainly not without false positives, but with these blood tests, now we had some data. When I got home that night, after the kids got to bed, I stayed up late and compiled all the information that we had. We got the health portal from the New Orleans hospital and logged into that. This stuff is not great. The data infrastructure, as anyone who has spent time in the medical system knows, is not great, and data portability is not great.

The first thing I had to do was click into a bunch of different test results and copy the full view, because I didn't know at that time if I should trust the screenshots alone. I could have just taken screenshots, but I didn't know if that would lead to worse performance from the models. This was really important. I was also thinking, “Maybe I can have an agent go through and click on all these things and grab the more detailed view for each test.”

But my thought was, “No, this is N of 1.” If I were doing this all the time, maybe. But I'm not. I just want to compile this one time and get some good AI eyes on this case. So I did it myself, purely manually—copying and pasting and visiting every one. I trust myself more than I trust an agent on a one-off basis to do a good job of that.

I pulled all that together and wrote up the medical history. This was the first time I wrote up the medical history. Amy and I went back and forth on that a little bit. I wrote everything I could remember, she added a few things, and then we put that together with the test results.

I opened a tab for every model. I did GPT-5 Pro, Deep Research, Claude 4.1 Opus, and, of course, Gemini 2.5 and Grok 4—all across the tabs. I put the same big prompt into each one: “Here are the test results. Here's the history from the parents. Analyze the case.”

This was a major pivot moment for me. My wife would say, “You should have been listening to me more,” and maybe I should have been. I was getting concerned for sure, but I wasn't jumping to conclusions, especially because at the hospital they had said, “We've ruled out anything life-threatening. It doesn't seem like a serious issue. Follow up with your pediatrician.”

I was in that mindset, but still concerned enough that it was the first thing I did when we got back: compile all this information and put it into the AIs. This resulted in a real freak-out moment for me, to be honest, because the AIs all came back with different tones, which was interesting, but they pretty much all came back saying, “You cannot ignore that LDH. That is a red flag.”

“Hemolysis of the sample would raise the level, but it probably wouldn't raise it that much. This is a real concern. You need to seek additional medical attention in the next 24 to 48 hours and make sure that you get a proper workup done on this kid.”

They were starting to use leukemia as one possible cause of all this, while still saying it probably wasn't the most likely thing. It was at the top of the list, not necessarily in terms of likelihood. I wasn't asking them to give me percentage guesses at that point anyway, but it was at the top of the list in terms of: obviously, if it is this, it's really serious, and your whole life is about to change if it is this.

Ruling out leukemia was the first thing they framed it as, except for Claude, which was basically like, “This seems like leukemia.” Claude was quite alarmist in its response, quite emotional, and I've gone back and forth honestly on whether that was appropriate or not.

I think about how Amanda Askell talks about Claude being always a stranger in whatever context it's operating in. What would a virtuous person do? A virtuous world traveler who's a visitor to all these places and just trying to be as helpful as possible? I don't know. I think Claude might have had it right.

It was jarring. The Claude response was by far the most alarming. It caused me to freak out. I never have anything like this, honestly, but I did have a mild panic attack or anxiety attack or something. I'm not sure that's really the right term for it, but I was freaking out. I'm sitting here at my desk reading all these things, thinking, “Oh my God, is this really happening? Could this really be happening?”

Claude doesn't know me. It doesn't know how I'm going to think about this. I found that I preferred the GPT-5 Pro response. Gemini 2.5 was always good, but I wouldn't say it really added too much. I wasn't super thrilled with Grok 4. I didn't try Grok 4 Heavy, to be fair; that might have been better.

I found that I preferred the GPT-5 Pro response: a little more long-winded, more detailed, more clinical, and less emotional. It was still basically conveying the same message—that this is a red flag and you need to pay more attention to it—but not in a highly alarming way like Claude did. That kind of tone was just too much for me, and I honestly moved away from Claude and favored GPT-5 Pro during this time.

But I also think Claude was right. It was really urgent. If it was talking to somebody like me, who isn't by nature particularly concerned and is thinking, “Oh, it'll probably be fine. It'll pass,” then maybe that kind of alarming language really is what was necessary to get me over whatever hump or denial I might have been engaged in and actually doing everything I could.

As it turned out, this is in fact the fastest-growing cancer known to man, and the doubling time is 24 hours. It really was important that I get into a different gear. So I was put off by that from Claude. I found that I didn't want that in future interactions, and I ended up moving away from the model for most of the rest of this episode.

Yet maybe it was what I needed in that one moment, just to make sure: “Okay, this dude needs to take this seriously. How am I going to make sure, as Claude, that I get him into the right headspace so there are no more delays and he does everything he needs to do?”

It did work. It's interesting to think about what would have happened if I didn't have Claude and only had the other ones. I think I still would have been in that headspace. Some of the episodes that we did see with him just being so obviously, deeply unwell were haunting. There was something there; we weren't going to just ignore that.

So I don't think I had to have that from Claude, but I could see slightly different situations where that kind of real alarm-bell ringing from Claude might be the right thing to do for the patient, even if it's not the kind of tone that I wanted to hear and it ended up scaring me away from the model a bit for most of the rest of this episode.

Overall, at this point in the story, we still didn't have much information. I would say the AIs were more accurate than every doctor we had seen up until that point. The New Orleans doctors were wrong to say that they had ruled out anything life-threatening.

The AIs did not make that mistake in my prompting, across the board. They all said, “You need to go on and try to rule that out.” They did not consider the results that we had as having ruled that out. Quite on the contrary, they considered it to be a live possibility.

Interestingly, my wife Amy was using GPT-5. She had not yet upgraded to Pro at that time. I guess it's on me, as the local AI influencer, for failing to convey to her, “Use Pro for anything that really matters to you.”

She was using GPT-5 and did have the paid version, just not the $200-a-month version. One thing that she tried was a very similar thing.

In fact, she extended my initial prompt and response, and she pasted in screenshots from the portal showing all of the results. Whereas I had copied them as text, she tried the screenshot approach. Interestingly, that did work in the sense that I think the model had no trouble interpreting the content of the screenshot.

But one thing that was in the screenshot version that didn't get copied into my notes was this hemolysis thing. I had made a note of it in my prompt, in the medical history that I wrote. I had said that we took him to the ER, that the doctors did all these tests and deemed them normal, explaining the LDH being high with this hemolysis explanation. But that had come in my voice, and the results as they were presented in black and white just showed the actual levels.

When she put in the screenshots, the UI in the portal itself had, in addition to the results, this text saying, “Hemolysis of the sample. Don’t take this super seriously,” or whatever exactly the language was. In response to that, it actually did make the same mistake that the human doctors had made. It basically said, “We throw that out. We look at everything else. It’s normal”—basically the same reasoning path that the human doctors went down.

Going back to one of my earlier comments, if it’s something serious, absolutely use humans and AIs. You’re going to need them both, and they both have a lot of value to add. You can do and ask much more of one with AI, whereas with the human doctors, you’re obviously very limited in the number that you can speak to, for how long, how many different ways you can ask the question, and how many different ways you can present information.

I would say use Pro is one lesson. We did get better results from GPT-5 Pro versus just regular GPT-5. Also, do multiple takes with multiple different variations on the way that the information is presented, and then look at the totality of what you’re seeing.

You can get a lot of second opinions. You can get n opinions from AIs. While some models, in some cases, with a certain presentation of information, did make the same mistake that the human doctors in New Orleans did, other models with slightly different presentations definitely did not make that mistake. At least one, Claude 4.1 Opus, really tried to leave nothing to chance when it came to making an impression on me—the user, the father of the patient—that you need to act with urgency because there is at least a decent chance that this is really serious.

So that was a pivot moment for me. It was like, “Okay, even though sometimes it may be saying everything seems fine, as the human doctors had, other times it’s saying you need to act with urgency. It’s time to act with urgency.”

The next day, we went to the pediatrician and saw him. He again said, “I still think the most likely scenario is that this passes. You never really have an explanation.” And that’s pediatrics. The kids have these complaints, and they come and go. As long as they pass and the kids seem normal, we don’t worry about them.

But I do think you should go to Children’s Hospital and see an oncologist, just because I would hate to think that I’m missing something. Better safe than sorry. Any kid who’s been complaining of pain for this long deserves a thorough look.

He did a nice job of advocating for us, and he actually made the phone calls that needed to be made. He got us an appointment first thing the next morning at Children’s Hospital. This pattern almost played out again. We saw the doctor, and she reviewed things. She basically said, “I don’t think this sounds like cancer. My guess is it’s more of an autoimmune-type of thing, but I’ll do a physical exam.”

Actually, this is something that I would also advise parents especially, but even for yourself. One thing that I had not had the intuition to do at any point in this process was to ask AI to interview me for more information, or to ask AI to guide me through a physical exam of my kid. I think that might have actually been really helpful, because another hinge moment in the case—which I think could very plausibly have been missed by all the humans in the room—was when the pediatrician touched a spot on his lower-left abdomen during the physical exam.

He really recoiled in pain from that. The rest of his body had been touched and palpated in all the different ways, and he was fine. He doesn’t like that stuff, but he was tolerating it. Then, on this one part in his lower-left abdomen, all of a sudden it was like, “Ow, that hurt.”

We didn’t think too much of it at the time of the pediatrician appointment, but when we saw that happen again with the oncologist, Amy and I both said, “The exact same thing happened yesterday.” If I could coach myself earlier, I would say, use AI to talk through a physical exam. Do it on yourself or do it on a kid, and see what you can turn up.

All the talk, all the record of the symptoms—which we did bring in at that point—and even the blood tests, which had the red flag, never quite added up to “We need to get really serious about it,” because he was presenting so relatively calmly and normally in those moments in the office. That was until that moment and that touch. We said, “We saw that exact same thing yesterday, Doctor.” Then she said, “Okay, we need to get an ultrasound.”

At this point, again, you’re like, “Let’s get the ultrasound. How soon can we do that?” This was Friday, naturally. It was Halloween, October 31, Friday. “We’ll see when we can get that scheduled.”

This is a moment where Amy gets credit for doing some forceful advocating. She was basically like, “I’m really worried. Can we do this right now?” And they were able to do it. I give them credit as well for listening to her and for going ahead and doing that.

They sent us down. The technician didn’t tell us anything, so we went back up to the doctor. She saw us again. We were there for 5 hours, I think, between seeing her and going to other parts of the hospital and what have you. Her response was, “Okay, now I am concerned. They do see something.”

My guess is that at this point they were kind of sugarcoating it for us. It was just the preliminary read, which had been happening a lot in our experience. Some procedure, some diagnostic, whatever, some scan happens, and there’s a preliminary read. The doctor managing your case talks to the radiologist or whoever is doing whatever procedure to get their preliminary impression, but you don’t really get real data at that time. You don’t get a report. You don’t get anything that you can really take home.

We weren’t recording all the appointments. We probably should have been. That would be another tip: record all these appointments, have AI transcribe them later, and glean what information you can with the help of the AI transcription.

But we just had this little conversation. She said, “They do see something. We can’t diagnose on this basis, but I am concerned, and so I want to get you in for an MRI as soon as possible. We’ll see you back here in the clinic next Thursday.”

It still didn’t seem like it was a total emergency situation at that time. Again, they weren’t even saying it was necessarily cancer. It could be a benign cyst or something like that, but they were definitely going to take a harder look at it. An MRI was the next step.

When we did get the full ultrasound report, which popped up in the patient portal over the weekend, running that through the AI made it pretty clear that, okay, we had a serious problem here. At that point, it was starting to really look like our kid might have cancer, which is a hard thing to get your head around, especially because he seemed totally fine 2 weeks prior.

Nathan Labenz

But you read things like, “There’s a 5 cm x 6 cm x 4 cm mass in his belly, and there’s also something on the liver, and there’s also some other stuff in other places.” There was probable involvement of the spleen and probable involvement of the kidneys. You’re like, “Okay, this seems like it’s maybe really happening.”

So, they scheduled the MRI for Monday. That weekend, he wasn’t suffering too badly, but the debate at home was like, “Should we take him into the emergency room even sooner and try to accelerate this even more, or do we just wait for the MRI on Monday?” With kids, they do the MRI under full anesthesia because they just can’t sit there long enough. It’s a full hour that they have to sit in the tube, and I don’t think our kid would have been willing to do that, and most kids aren’t. So they do it under general anesthesia.

Amy was like, “I think we should take him in sooner.” I said, “They said Monday.” At this point, we were asking, “How urgent is this? How short is this timeline?” because it had all seemed like it was happening pretty fast. But again, it was about to start happening a lot faster. We didn’t know what was in store.

They were saying these cancers grow pretty slowly. At that time, their best guess—and AI also concurred with this—was neuroblastoma, and they were kind of like, “They grow relatively slowly. They change on the time scale of weeks or months. Certainly, we definitely wouldn’t wait a month, but a couple of days probably isn’t going to make a big difference. It’ll be fine.”

Since he was going to be put under general anesthesia and he’d never had that before, and I’d never even had general anesthesia prior to these last couple of weeks, never even seen a person go under general anesthesia, I was like, “Let’s do it Monday under controlled conditions.” There’s a plan for that. Everybody’s got their plan. There’s a schedule. It seems like it’ll be safer to do it on the schedule and according to plan rather than go into the emergency room and try to push for them to do it there. Hopefully they will, but it might be more chaotic or whatever.

Ultimately, the attending oncologist, whom we called over the weekend, had a judgment that was more consistent with mine, which basically amounted to, “Take some Motrin. If he’s feeling okay and he doesn’t have a big fever, come in on Monday. It’ll be fine. These things don’t move that fast.”

The weekend wasn’t too bad. He took some Motrin, and it was controlled. His pain was controlled. He was walking. He was still able to walk up the stairs. Certainly not his best, certainly low energy, certainly something was not right with my baby—we knew that—but it still seemed very plausible that it really wasn’t that urgent. These things don’t move that fast.

Had it been neuroblastoma, it would have been both slower-growing but also, as alluded to earlier, a lot harder to treat. Actually, it would have had a lot worse prognosis. We didn’t know all those details at the time. But if I had to critique the doctors there, I would say there was high uncertainty, and we did start to ask AIs for percentage chances at this point in time. We were getting neuroblastoma at maybe a 30% chance, lymphoma at maybe a 25% chance, and a benign mass at this percent chance, and whatever.

But were we effectively acting according to the worst-case possibility? I would say the human doctors probably anchored a little bit too much on the most likely scenario and didn’t probably give enough weight to other still-not-much-less-likely scenarios, still very plausible scenarios that, in fact, would demand more urgent attention. And sure enough, that is where we ended up.

We went in Monday for the MRI, and they said, “Okay, with that, we’ll now do a biopsy.” They were able to schedule the biopsy for Tuesday. So we went in Tuesday, and he had a whole bunch of procedures. The biopsy itself was a little surgery: 3 holes in the belly to get at the main mass. They also took a part off the liver. They did a bone marrow aspiration, they call it, to get some bone marrow to see what’s going on in there. I think he had a CT scan that day. The whole thing involved a lot of procedures.

After the biopsy, we were directly admitted to the hospital. At that point, we were very glad to be directly admitted to the hospital because he was just clearly getting worse. It was one of these things where, again, time has this weird feel: looking back, it’s flat, and in front of you, it’s vertical. But if you zoomed out and said, “Where were we 2 weeks ago? Where are we now? Where were we 1 week ago? Where are we now?” then it was like, “Oh my God, this is actually moving pretty fast.”

At some point on Tuesday, the day of the biopsy, and Wednesday, he was really looking bad. We were just like, “This is going downhill fast.” At that point, I also started to get into it. My wife is a natural advocate for her child at all times. I’m a little more chill. But at that moment, I was like, “Okay, what we are seeing here is now getting really scary.”

I would say those were the scariest days of my life. By far, the biggest cries I’ve had in the last 20 years were during that period. The biggest one was the first time they put him under anesthesia. I really wasn’t ready for that. It was so sudden, and it just made the whole thing feel so real. It was like, “Oh my God, I’m handing my baby over to these people to literally cut into his body and do all these things. Oh my God, this is going to be the beginning of such a long road, and we still don’t even know what’s going on or what the prognosis is.”

So, I had a good cry about it at that time, but I was also really scared enough to move into a pretty strong advocate mode. This is when I was like, “You guys have done well so far,” was my message to the team at the children’s hospital. I think overall they have done really well. I’ve talked to an oncologist friend at a bigger center, and she said that at bigger centers, sometimes it can even move more slowly.

This is a pretty big center, but they did seem to move pretty agilely and coordinate schedules, and it’s a huge operation, obviously. So, to get the kid in from a Friday visit to a Monday MRI to a Tuesday biopsy was pretty good, and yet for a minute there, it seemed like it still might not be fast enough.

On Wednesday—we were counting the days, and this was now November 5th, so this was 10 days ago—we were just filling time. They didn’t have anything scheduled for Wednesday. There was going to be a draw of cerebrospinal fluid the next day, but there was nothing really on Wednesday. We were just sitting there in the hospital, and he was getting bloated. They were giving him fluids, but he was all puffed up and swollen, could barely get out of bed, was really unsteady on his feet, and was as weak as could be.

We were just like—you are looking, I didn’t say this to him, of course, but in our minds, we were just like, “You are looking terrible, and we are really scared.” So, we began to push the doctors: “Look, it now seems clear that every day is of the essence. From yesterday to today, what we’re seeing is so scary. We need to absolutely be prioritizing this as fast as we possibly can.”

Their message was very consistently, “Okay, don’t worry too much yet. We need to get to the diagnosis. We need to get to the diagnosis. Before we get to the diagnosis, we don’t know what the treatment plan is going to be. With the diagnosis, then we will refer to what the standard of care is, and then we can execute on that. We understand that you’re in agony here waiting, but we can’t treat appropriately without the diagnosis.”

So, the point of advocacy became, “We need to get this diagnosis.” But I was like, “The weekends—obviously, the cancer doesn’t care about the weekend, but the medical system functions quite differently on the weekends.” That was one of the reasons I had been hesitant, and the oncologist on call at the time had been hesitant, to rush into the emergency room and try to get a weekend MRI.

But now I was thinking, “Oh my God, it’s Wednesday today. If this slips at all, we’re not going to have a diagnosis till Monday.” I told the oncologist there at his bedside, “5 days from now seems like a really long time. I think it’s imperative that we get this started by Friday.”

Right then, my son said something like he couldn’t move his legs. He needed to pee or whatever. I was like, “Okay, get up.” He was like, “I can’t.” The oncologist was startled by that. He took a sharp breath or said, “Oh my.” Clearly, that was a pivotal moment for him where his concern jumped up.

I think, in fact, my kid could use his legs at that point. I think he was just really tired and not feeling good at all, obviously, and didn’t want to, and said, “I can’t.” But him saying that was enough, I think, to take the oncologist from, “Standard protocol: we’re going to get the diagnosis, then we’re going to treat. Yes, we’re trying to accelerate everything we can, but we can’t get ahead of the diagnosis,” to, after hearing that and going and conferring with the team or whatever and thinking it over himself, the next time he came back, he said, “I think we should go ahead and start the chemotherapy even before we get the official diagnosis. I agree with you. Time is of the essence.”

They had narrowed it down enough at that point that he was pretty confident it was going to be some kind of lymphoma. From the biopsy, the surgeon—they don’t diagnose by eyeballing it in surgery—but he came out and said, “It’s definitely not neuroblastoma.” He basically felt, “I know what that looks like, and this did not look like that.”

So, that put them much more in the lymphoma category. With that, there are different subtypes of lymphomas, and you might treat them somewhat differently depending on where it’s all found in the body. All that was still unclear. But at that point, the doctor switched over into, “Okay, we’re going to break the rules a little bit here. The treatment for any of these different types is similar enough, and we’re pretty confident at this point.”

Nathan Labenz

It is one of those things. Let's go ahead and get the treatment started, and we'll refine our decision as the final reports come back. But it is better to start now with somewhat imperfect information, given the trajectory that he seems to be on, versus waiting and potentially having it slip past the weekend.

They started the first drug of the chemotherapy that Wednesday, which is just a steroid that is very commonly used, but it is part of the chemotherapy regimen. Then the next day, on Thursday, after they did a few more procedures, that's when they started with the more serious, classically chemotherapeutic drugs. It was good they did, because he really was on an exponential trajectory, with a doubling time of 24 hours. He was getting weak, looking swollen, and could barely move.

As we've learned more, obviously, this is all terrible luck on a macro level. But in some ways, I think he was kind of lucky in that the cancer had grown a lot by this point and it was all over: a big mass in the belly, part of the liver, all over the kidneys, lots of different places, in the bone marrow, and possibly in the spinal fluid. Again, that still remains a little bit ambiguous. Basically, they found some in there at one point, but it was a very low amount, and they were worried maybe it was contamination. They weren't quite sure what to make of it.

Nothing in his brain, thankfully, or spine, thankfully. But regardless, okay, start the treatment. He's clearly on a very bad trajectory. Start the treatment and catch it over the next several 24-hour doubling times. I honestly don't know how long he would have lived. It's clear that without treatment, he wouldn't have lived very long, but exactly how long, I don't know, and I haven't asked that question. It certainly seemed plausible that if things had slipped that long, from that Wednesday to the following Monday, he maybe barely could have made it to that Monday.

I think one of the threshold effects we were approaching at that time, I believe, was organ failure of various kinds, given how much the cancer had grown. Again, it's like good luck, bad luck. How do you think about it? The good news was that all of his organs were functioning well when we went into the hospital. The kidneys, the liver, all the enzymes, the markers they look at—they were all good.

That was part of why, in New Orleans, they said, “We've ruled out anything life-threatening. As the cancer grows, it starts to choke off the organs. The flow of fluids in and out of organs is critical. If fluid can't get in or it can't get out, you start to have a real problem.”

He never really had that problem. I think he maybe started to flirt with it that Tuesday and Wednesday before the chemotherapy started, but he never really had it. He never really had anything approaching serious organ failure. But I do think in the next few days, we probably would have started to see one or multiple organs start to fail, just as they were getting choked off by the growth of this cancer throughout his body.

So, we started the chemo. The next day, we did get the information back. This would now be Friday, November 7, and at that point they said, “Yep, it does look to be a B-cell lymphoma. We believe it is Burkitt lymphoma. There's still some more fine-grained testing. We'll look at all the different cellular markers and all that kind of stuff, and there's a genetic component to it as well. But everything we see so far makes us think that this is Burkitt lymphoma.”

That means it is super-fast-growing, super-aggressive, but also super-treatable and super-responsive to treatment. The most likely scenario by far is that he is cured. It should work. The cancer should never come back, and he should live a normal, long, healthy life.

However, if it does come back, it's really bad, so that's also important to understand. It's definitely very important to follow the chemo protocol because we want to get every last cell of this cancer. If we miss any, the longer we let it go on in there, the more the cancer itself can mutate and evolve and figure out how to evade these drugs. If it does come back, the prognosis is not good.

In some ways, it was a major relief: “Hey, good. The overwhelming chance is that we're going to have a positive outcome here.” But it was still obviously very scary to contemplate that long-term future.

That diagnosis moment, though, is really the critical before-and-after. Before this, you're in total panic, total fear, especially if it's happening so fast. You're like, “Could another 24 hours take my kid?” After you have at least a sense of what's going on and you have a plan, you start to move into more of a managed situation.

Again, there are so many different kinds of cancers, and they're all unique, so I don't want to overgeneralize from our experience with this particular cancer to anything else. But because they do have a very effective treatment protocol for this cancer, at least for us, we moved at that point into more of a managed situation: okay, here's what we're going to be doing, and here's the schedule.

It is a lot on the body. The first phase of chemotherapy that they give is less intense than the next 2 rounds. The reason it is less intense is because when the tumor breaks down so fast in response to the drug, that process can itself overwhelm other systems of the body. They call that tumor lysis syndrome, and it has to be carefully managed or it itself can prove deadly. But they do have experience doing that, and they seemed pretty confident about what they were doing.

As we got into this phase, it became clear that there were probably 3 big things we could do as parents, and where AI could help with all 3 of them, to improve the quality of care that he would get and improve the overall odds of a good outcome.

The first one is just double-checking the doctors' work nonstop. They were taking blood every 6 hours, so we were getting lab results 4 times a day. Not every test was done every 6 hours, but basic ones were, and some were done just daily. So, a ton of information was flowing in. He was getting all these different drugs, and his condition was changing in all sorts of different ways.

He got bloated when he started getting all the fluids. They gave him something to get fluid off the body. He peed an unbelievable amount. It was close to a gallon, which is crazy. He weighs 50 pounds normally, so you're talking about taking more than 10% of his weight off just in urine in just a few hours when he got that drug.

It was a pretty dynamic situation, with a lot going on, a lot of moving pieces, drugs he's never had before, obviously, and this constant flow of information. We also didn't really know at that time: how good is this team? We want, as any parent would, the best possible treatment for our kid. Are we getting it? How would we know?

The role of GPT-5 Pro in particular has been just outstanding. It's really been an amazing AI doctor, essentially, that has ridden shotgun with us every step of the way. I have a long-running thread where I'm just focused on the bedside: what's going on now, what are the latest labs, what do we need to do, and is there anything they might be missing or anything that they're doing that might be wrong?

Access to information is tough. We don't have the same view that the team itself has, which is annoying, to say the least. But we do have the patient portal. We can get the labs out of there, and we can write notes about the updates, what he's feeling, what's going on, and what medicines they're giving him.

It would definitely be better if we had all that documentation readily exportable. Nevertheless, with what we do have, I have a long-running thread where, at every lab draw or any moment when there's something going on that's of concern, or they're planning to give a new drug that wasn't part of the plan or whatever, I put it in there.

That happened once with the bloating. It happened another time when his blood pressure went up and it was just over the threshold where they wanted to treat it. I was like, “Okay, should they or shouldn't they? It seems like he's already got a lot of drugs.”

In all these different moments, the patient portal has a print function that spits out a PDF. So, I just grabbed that and put it in. I've probably done that close to 20 times now over the last 10 days since the treatment started. We got on more of a regular cadence, and it has really allowed me to educate myself tremendously.

I've learned an absolute ton from this process, understood what the doctors are doing and why, and seen if what they're doing really makes sense. I came into this with pretty high confidence in the latest models to do a pretty good job. I would, again, only use the best.

GPT-5 Pro from OpenAI doesn't really have a bedside manner. It does have nurse-practitioner-level capabilities, and the GPT-5 series was at least competitive with, if not beating, Claude a little bit, even though Claude did score higher on the overall GPQA specifically. Sam Altman did note when they launched GPT-5 that they've worked really hard on the medical use case specifically.

It did seem to be, across the medical tasks, the highest-performing model. As I said earlier, I preferred its long-winded explanations and more neutral clinical tone, and it's been really good. It has been absolutely on the level of the human doctors in terms of their analysis, and it has also given me a ton more confidence that they are in fact doing a good job when it comes to the main overall treatment protocol that they have.

The AIs have confirmed, yes, this is absolutely state-of-the-art, best care you can find anywhere. They've explained to us that this is highly protocolized: anywhere you go, if they're doing the right thing, they will be doing this.

One of the questions we had was whether we should move to a different center, or whether we should try to move, or whether there was better treatment available somewhere else. The AIs have been pretty consistent in saying, “Not for this case.”

It could be very different if you needed surgery to remove a tumor and you were looking for the best surgeon to do this sensitive surgery, or if you just had a really rare condition. This Burkitt lymphoma has a little bit over 1,000 cases in the United States per year, so it is fairly rare, but there is a protocol for it.

In other cases, where you're just an N of 1, a total outlier, and nobody has seen something quite like this before, then again, I think you might want to seek out, if you have the means to do it, the single clinician in the world who could give you the best guidance on this particular case.

But in our case, because this was a known cancer with an established protocol that is known to work well, the AI was very reassuring that we were getting the best care. Then, on a day-to-day basis, with all the things that come up—the blood pressure is up, the heart rate is down, why is this going on, do we need to treat it or not—it was all broadly very much in line with the doctors and, by and large, really supported their analysis.

So that gave us a couple of things that we advocated for. ChatGPT or GPT-5 Pro was more likely to suggest supplementing or replenishing key minerals. He's been low on calcium, and he's been low on magnesium. Probably that's just because he's getting so much fluid that it literally dilutes the blood to some degree, but you don't want that stuff to get too low.

The doctors, for the most part, were like, "It's not that low. Oh, it's a little low. We expect it to be low. No need to do too much with it." GPT-5 Pro was a little more like, "I'd like to see that higher." So we pushed for that a little bit, and I think on the margin there was probably one time—or maybe they started doing it a little earlier than they otherwise would have.

It also didn't quite agree with the doctors on how quickly to treat the blood pressure. When the blood pressure went above the threshold that they had set for treatment, it was just slightly over, and GPT-5 Pro was like, "Most clinicians, I think, would take a wait-and-see approach here. Measure it again, measure it more often, then treat—especially if it goes up more—but probably not treat yet."

They wanted to treat, and at a minimum, GPT-5 Pro gave me enough information and understanding to have a conversation about what was going on and how this works. The AI explained that to me, and with that information I was able to have a conversation with the doctor, which I did find compelling because I had enough information to go into that conversation and feel like I could parse what he was saying back to me. I was able to be confident that, okay, this seems reasonable, and I said, "Okay, I trust your judgment. Let's go for it."

Even though GPT-5 Pro would have waited, GPT-5 Pro is also a little bit more aggressive about wanting more information. There were a few different tests that it wanted to run, basically to rule things out or to confirm its understanding, and the doctors didn't always want to do that. They were kind of like, "Yeah, that enzyme, that liver enzyme, it's up, but we expect it to be up. He had a biopsy that touched the liver. It's always up in these cases. We don't need to do more testing just for that."

I think overall these were very marginal differences in how to proceed. Because they were so small, and because the logic that was being expressed to me by ChatGPT was so consistent with what the doctors were saying, even if on these very fine-margin points they might have done something slightly different, overall it gave me a ton of confidence and a ton of peace of mind that the team here is doing a good job.

Not just that they're using the right protocol for this particular disease, which is well established and everybody would use if they knew what they were doing, but also that, at the bedside and on a real-time basis, they were generally on top of it and had good judgment that we could trust. I gained a lot of confidence in that from AI, so I would absolutely recommend doing that.

If you are in the hospital yourself and you have the ability to do it, or if you're helping anybody else—kid or other adult—definitely be on top of that patient portal as results come in and as new medications are being introduced. Take the time to do it. The first time is hard because you're figuring out the portal and where all this information is, and you've got to assemble this context.

I saw a tweet while I was at the hospital from Amanda Askell saying that she has multiple 100-page prompts that she uses regularly, and I was like, "What is that?" I guess, first of all, that does reflect something that I have experienced in this process, which is that the models are really good at long context. Forget about what you might have learned at some point in the past about overwhelming the context. That problem, I think, is pretty well solved at this point.

If Amanda Askell is using 100-page prompts on the regular, you can be pretty confident that it's not wasted information, and it's not overwhelming or confusing the model. My experience has been consistent with that. I hadn't really done anything like that myself, but with all this medical stuff, now I really am starting to get there.

My new standard prompt is: "Here's the prehistory from the parents of the initial symptoms that brought the kid in the first place. Here's all these different reports from the MRI, from the CT, from the PET scan, from the spinal fluid analysis, from the bone marrow analysis—deep stuff with particular cellular markers: how many, what percentage of cells expressed all these different markers—the pathology report itself as they worked up the tissue, and other stuff I'm forgetting."

They've got about 10 PDFs that can go into a single prompt that really present an overall, pretty holistic look at the case. This is getting into the 100-page range, and I am using it pretty regularly now. So that's a general AI tip, but definitely with medical stuff, you want to be making sure that the whole picture, as much as you can represent it, is represented in the prompt. You don't just want to show up with, "Oh, here's what's happening right now," and haphazardly type it in.

Another interesting trick, especially as we've started to move into some other ways that we can help the kid—and I'll unpack that in a second—is using this really long-running bedside analysis thread. One of the useful things to do is say, "Okay, now can you convert all of this history into a handoff document for a new attending physician so that they can provide the most cutting-edge care possible?"

What we got out of that was just a couple thousand words, but it was extremely information-dense and super useful to then put into another prompt. So that's the first bucket: get ongoing second opinions at the bedside. This gives you information to understand what's going on, and it gives you the ability to advocate when you need to. If the team is doing a good job, the AI is probably going to say that, and it's going to give you a lot of peace of mind.

I imagine myself without AI, and what I imagine is constant Googling, constant anxiety, never feeling like I have enough information, just being totally overwhelmed and totally stressed out, and never quite sure if they're doing the right thing or not. In contrast, with the AI to help me, I felt quite empowered. Obviously, it's still a stressful, scary situation, but I had much more confidence in my ability to advocate effectively and more confidence in my belief that the team is, in fact, doing a good job for him, which was obviously great to be able to believe with confidence.

The other two big things that we can do for this particular situation are minimize infection risk. This chemo that he's going to get is going to be pretty intense. The immune system is going to be really wiped out at times, and infection is one of the big things that can get you.

You don't have full ability to avoid that because some of the infections people develop just come from bacteria that are already in their mouth or on their skin or in their gut. So it's not like you can drive that risk to zero, but you can do some things.

Because Amy and I both work from home, and because we only have 1 other kid in school, we're in a position to do a pretty good COVID-style social-distancing, quarantine kind of protocol for our family for a while. One thing we're going to do is just try to minimize as much as we possibly can the risk of exposure. That means I won't be doing any traveling to AI events for a while or any on-site speaking engagements for at least a few months.

The treatment roadmap is 6 months. We also happened to have a flood in the basement the same day we came back from New Orleans concerned about his health. I walked down into my basement, and there were 2 inches of water in parts of the basement.

Unfortunately, there wasn't a lot more because the water was coming in fast. Fortunately, our drain was flowing, so it was also flowing out pretty fast, and it didn't really turn into a bad situation. Nevertheless, you've got to deal with that, right?

One of the big concerns after basement flooding is the rise of mold. It's also a big concern for kids when it comes to infection. The doctors actually aren't that worried about viral infection. They are quite worried about bacterial infection. The viral infection is bad, but they don't think it's going to be life-threatening. Bacterial infection can be, and fungus can be as well.

So now you're in a really tricky spot if you're me. This is a great example of where AI can shine so much. What do we do about this mold? How do we deal with this? How would we know if we've dealt with it? Is this really a risk, and how can we deal with the risk?

I've run so many deep-research reports with GPT-5 Pro over the course of these last couple of weeks. It's a lot. One of the ones that I ran was, "What companies in my area can help me with mold prevention or remediation in the wake of a flood?" I noted that I do have a reason that I need it to be extra well done, and it helped with that.

GPT-5 Pro was also really good at taking the results of a mold test that a mold-remediation company came out and did, where they literally just take air samples—one outside in the backyard, some down in the basement where the flooding was, and others in different parts of the house. It was really useful.

Try to imagine getting your oncologist and your mold guy to talk to each other. This is not going to be an easy meeting to arrange.

They usually operate in a very different cultural context. How are they going to have a real meeting of the minds and get you the productive synthesis that you need to know? Can I bring my kid home to my house? Do I have to move out of my house for the time being? Is there something I can do to make this work because of this finding of a little bit of elevated mold in the basement?

Spoiler alert: we have a little bit of elevated mold in the basement. The rest of the house seems fine. Because of that, would we treat prophylactically for mold differently than we would otherwise? In general, they don't treat kids with antifungal drugs because they don't think it's worth it, and there can be side effects. But if you're going to be living in a mold environment, would they treat prophylactically?

Bringing knowledge of basement remediation and mold prevention and oncology, with all the specifics of my kid's case, together is just something that you basically would not be able to do otherwise. But I was able to do it and get to the point where I'm confident that I know what the plan is going to be. The plan is basically going to be to buy a bunch of HEPA filters and run them all over the house. When they're doing work in the basement, don't let him go down there. Certainly, maybe we'll go to my parents' house for a couple of days while the work is actively being done. GPT-5 Pro did recommend that.

But it seems quite clear, given the fact that the mold level in the basement was only slightly elevated relative to outside anyway, and the overwhelming effectiveness of HEPA filters at filtering mold out of the air, that we're going to be fine. I'm probably going to buy 8 to 10 new HEPA filters and have them all over the house, running constantly, which would certainly be overkill if I didn't have my son's health to worry about. But if we do that, the analysis, I think, is very solid, and this will be fine. We're actually going to have a lot fewer mold spores in the air in the house than one would be exposed to just by going outside here in the fall, a season when all the leaves are coming down.

That was huge. What a nightmare, right? I got this flood at the same time as my kid has cancer, and then they interact there. This is not just 2 problems I have to deal with. It would be fine to just put the basement on the back burner, except for the fact that it creates a specific danger to him right now.

I don't think I would have been able to get very confident about the situation. I might ultimately have chosen to move out of the house out of general fear, but with the quality of analysis that I'm able to get from GPT-5 Pro, I know what to do. I know that these HEPA filters are going to keep him safe, and we might still take a couple of days out of the house, but otherwise it's a fine situation. So, that is huge.

So, I have a second thing: reducing infection. The third thing goes back to this kind of weird bimodal outcome where, unfortunately, one mode is much bigger than the other mode. But with this Burkitt leukemia, most of the time it responds to treatment really well. Fortunately, that's been the case for my son. There was an 80% tumor reduction or so, with lots of lesions just plain gone after the first week.

The bone marrow was clear. Amazing stuff. Basically a best-case scenario. So glad that's the update. But for some kids, it goes away and then it comes back. And if it does come back, it comes back with a vengeance. The prognosis is not good.

To put it bluntly, most of those poor kids die. Not all of them, but most do. So that's the third lever: what can we do to be as prepared as possible for that particular eventuality? Here, maybe, is AI's biggest contribution, which happens to coincide with a new model popping up in Google AI Studio. This has been posted about all over the internet and is thought to be a Gemini 3 checkpoint.

It's one of these situations where the model is under a code name and you're told, “Check it out. Don't publish your results.” Although I am seeing people do that, I'm not going to publish results out of respect for that request. But there are lots of purported Gemini 3 results out there that look pretty impressive. So I started using GPT-5 Pro and what I believe to be a Gemini 3 checkpoint in parallel to address the question: What if it does come back? What can we do now to be ahead of that?

What can we do to catch it as early as possible? What can we do to give him the best prognosis at that time? Where would we go? Because that would be the time to go to a leading medical center and try to get the absolute cutting-edge, state-of-the-art treatment, clinical trials, whatever it may be. And again, I do have months, hopefully, especially now that he's responded well.

It seems very likely that what's going to happen is we're going to continue with this chemotherapy regimen. The cancer's not gone yet. It is expected to be pretty much gone after the first 2 rounds. He did the prephase, which they also call the debulking phase. That's easing into chemotherapy, and that alone took the cancer volume down 80%.

The next phase he's starting now is the first of 2 pretty hard-hitting phases. You might see me bald in a future episode, by the way, because my son is going to lose his hair, and I'm not sure how he's going to react to that. I might shave my head in solidarity with him. We'll see. I'll probably actually let him decide.

So, if you do see me with a shaved head in future episodes, that will probably mean that my son opted me into going through that part of the experience with him. Anyway, after those 2 intensive rounds of chemotherapy, you're basically expected to look clear. What exactly “clear” means, we'll unpack that a little bit more later. Then there are 2 more rounds of moderate intensity that they call the mop-up phase, for anything we missed.

There are 2 final rounds that are relatively mild. Those are the consolidation phase, which is basically, again, “We really think there's nothing here at this point.” But because we know that if it does come back, it comes back hard, and because we really do have to get every last cell—because if not, you get back into the exponential growth, and it starts again—we've got to do all these things, right? We've really got to make sure it's gone.

But the AIs have been incredibly helpful for this task of orienting to, when it comes back, what does that mean? One thing that they've highlighted has not come up in conversation with the doctors at all at the hospital. In fact, doctors generally don't like to get ahead of themselves, but I want to get ahead of this because I want to be ready. I basically want to have an emergency plan where I know exactly what we're going to do.

If and when we ever get the terrible news that the cancer is back, I want to be able to act on that plan immediately and have all of the work done. I think this is something that parents, or anyone, can do for themselves now in a way that would have taken just an unbelievable amount of work before. Now you can actually do it easily.

What are the different classes of treatments that are out there for this recurring or relapsing cancer? There are different kinds. First of all, I should also say: what kind of cancer is it? There actually is still some ambiguity around exactly what kind of cancer it is. We know that it's this Burkitt leukemia, but there is another type of cancer that has very similar properties and has only recently been distinguished from the classic Burkitt case, but has different genetic rearrangements that are causing the cancer. That actually does lead to a different prognosis.

So, we still haven't gotten that data back. They send this out. It goes to labs in faraway places, and they've got a whole production line. It takes a while to get this stuff back, but we're still waiting for that. The AIs have identified the different classes of treatments. First of all, I think Gemini 3, honestly, is excellent. It really impresses me with how sharp it is.

It seems to be every bit as smart as GPT-5 Pro, but also much faster. Its responses are much more focused, but they still feel like they're getting the core information to you and not missing anything. Whereas GPT-5 Pro feels like talking to somebody who's a world-class expert in the subject because they're so thorough and have so much information, it's giving you this massive information readout.

I found that to be true definitely with Deep Research, which sometimes I want—the comprehensive view—but even with normal GPT-5 Pro responses, they're often repetitive, with a lot to deal with. It feels like it's giving you this big information dump, and it's on you to work through it and make sense of it. I wanted that because I did believe it was the smartest model that I had access to.

Again, I had a little bit of reluctance to use Claude for this because of what I found to be an overly emotional experience the first time. But Gemini 3 is much faster, probably 5 times faster, maybe even more than that—maybe 5 to 10 times faster. It gives shorter responses that are more focused and just feels like you're talking to the most brilliant person you've ever met, who really gets your question and frames the answer to your question for you in a way that is really sophisticated.

It has the right information, but it's also much easier to digest than GPT-5 Pro has been. So, I think Gemini 3 is going to be—assuming this is a Gemini 3 checkpoint, which I think is very likely to be the case—a huge deal. It might be, interestingly, the first time Gemini has had a persona advantage relative to other models.

I've loved Gemini 2.5 Pro for many things. By the way, the transcription of this information is so bad. These tests show up in the portal when they come back from the lab. Sometimes other things show up in the portal; other times they don't.

We're still waiting for certain key test results to show up in the actual portal. They're not good at moving this information to patients, in large part because I don't think patients historically have had much to do with it. Most people probably aren't even asking for it, but I have been asking for it.

A lot of times, they've brought me paper printouts, and I'm like, “Oh God, can you give me a PDF of this or send me the original?” And it's like, “Nope.” So there I am at the hospital with paper printouts half the time, some of them containing the most important information. What do I do? I don't have a scanner. I'm literally just taking a snapshot of the document with my phone, and that's obviously not the best representation of it.

Some of the documents have weird artifacts on them where they might literally have been faxed from the lab that did the analysis. Clearly, there's been document lossiness by the time it's handed to me, so it's not super easy to read at times, even with the naked eye. Then you've got the photo, and then you're trying to get AI to do that.

Your mileage may vary, but if you have a screenshot of a website and the text there is super crisp, as rendered on a screen, you're probably fine from an accuracy standpoint to put that in directly. However, if you're taking a picture of a text document, GPT-5 Pro is actually still not that great at transcribing something that challenging. Gemini 2.5 Pro is better. I would say it is clearly the best at transcribing documents like that, but it can still make errors on subtle things when you've got tables and an ad hoc, weird layout. Some of these documents are gnarly.

For those sorts of things, I would use Gemini 2.5 Pro to do the transcription. And then you still have to systematically and very carefully check it. When you're talking about something like what kinds of cell types are active in your son's cancer, you want to make sure that information is 100% correct.

When you're given this sort of seemingly previously faxed and then printed document that's not easy to read, I don't think you can really trust any of the transcription at this point quite well enough to nail that. Gemini 3.5 Pro would be my go-to, but you're still going to have to review.

Anyway, that's a bit of a digression because I was on Gemini. I think Gemini 3 is going to be a huge deal. I think it's going to be a persona and style advantage, and maybe also a raw intelligence advantage. It has seemed very smart, but I would say GPT-5 Pro has generally had very similar content. I wouldn't say it's missing things.

When I asked for lists—how would you prioritize? How would you taxonomize?—generally speaking, the information that I'm getting back can be reconciled to be the same information. But the presentation and the analysis are just so pitch-perfect coming from this model that I think is Gemini 3 that I think it's going to make a big splash. I think people might soon find themselves going to Gemini by default a lot more than they currently do.

As much as I have loved Gemini 2.5 Pro, I trusted GPT-5 Pro more. But Gemini 3, over the course of a few hours—literally last night, as I was really getting into this and asking, “What should my plan be in case I ever get this terrible news, and I want to be able to act immediately on getting the news?”—the model I believe to be a Gemini 3 checkpoint is just giving me exactly what I want.

It's giving me really detailed, really sophisticated analysis, packaged up in a way that is really easy to read and pretty opinionated. In part because I do have GPT-5 Pro there to check it against, it really is earning my trust. Just in my first session, it has taken over as the model I'm most excited to read the outputs from. GPT-5 Pro is more of a checker: Do these things line up? Whereas before, GPT-5 Pro was the number-one source, I think the model I believe to be Gemini 3 is taking the top spot.

That is pretty notable from just a pure AI standpoint. So what are we going to do? First of all, we've got to get this additional genetic information to figure out which subtype of Burkitt lymphoma my son has at the genetic level, because the prognosis actually does differ based on that.

Then it recommends—and I think this is really smart, and an oncologist friend had suggested something similar—getting a second opinion. What does a second opinion mean? I think for us, we have pretty high confidence that the treatment that's happening right now is the best available treatment. But getting on file and establishing a relationship with, I think, 2 leading oncology centers that are doing clinical trials with the different kinds of drugs used to treat this cancer when it comes back—that's a step we can take now.

Some of these centers do have formal remote second-opinion services where, for not a huge amount of money, like under $1,000, you can submit all your documents. They'll have one of their oncology team members work through your case and say whether they think you're getting the right treatment. Then you're also on file with them, and you can establish a relationship so that, if and when this happens, we want to come and be part of this or that clinical trial. They can already start to steer you in the right direction as part of that second-opinion process.

But Gemini 3—the model I believe to be Gemini 3—is also doing a phenomenal job of that, and really fast. It's using search. I have found that I need to tell it to cite sources if I actually want it to cite sources; otherwise, it blows past that. But when told to cite sources, it does a good job of it.

It did a really great job of giving me the taxonomy of different drug types. There are these engineered cells called CAR T cells, which are basically part of the immune system. They take these cells out of the patient, engineer them to attack the cancer cell based on markers that the cancer cell has, and then put them back in the body. Then they go attack the cancer cell. This is cutting-edge stuff. It is logistically challenging.

So I asked a ton of questions: Can I do that now? Could we even go to that treatment now? This chemo is going to be rough, right? Could we do anything that's sort of an immunotherapy? It seems like maybe it would be easier on the patient.

But I think Gemini 3 made many good points around, “Look, this protocol works. Follow the protocol. There's almost nothing in all of oncology that has a better outcome than this. So do the thing. Don't distract.”

Again, really good analysis and a mechanistic understanding of how this drug could interact in different ways. One of the problems with the immune therapies is that the chemo knocks your immune system down so badly that the immune therapies can't work until the immune system has had a chance to come back.

There are a lot of nuances to it, and I think this model has done a phenomenal job of explaining to me why all the different ideas I've had for different ways to customize my son's treatment regimen are not, in fact, good ideas at this point. In many cases, it's, “Save that for later. Save that for later. That could be a good idea later, but don't do it now.”

Very compelling. I've come away quite confident in its analysis, time after time, as I've tried to probe: Is there anything different that we could or should be doing?

Another class of things is drugs that attach to a marker on the cancer cell and recruit other immune cells within the body to come and attack them. None of these things are super effective, but these are the trials that are pushing the frontier for what can be effective in the case of recurrence.

The thing that the AI surfaced for me that I had not heard of anywhere else, that the doctors never talked to us about, that I think we absolutely do want to do, is a thing called ctDNA. This is circulating tumor DNA.

Basically, what happens is, because we have a tissue sample, we can request that the tissue sample be sent to the lab, and they do a genetic fingerprint on it. Then later, you can do blood tests, and they're very sensitive. I believe the number is 5 cells in 1 million that it should be able to detect.

So if this cancer is coming back in a meaningful way at all, it starts to show up on the ctDNA before it would show up with other diagnostics. Then you can get a jump on doing some treatment.

Our oncologist was aware of that but said, “Yeah, it's not really standard, whatever.” A friend of mine who's also an oncologist said, “Yeah, I don't really recommend that for my patients. It's not really proven. And if it does come back positive, then what are you going to do? Would you treat based on that if you can't even see it?”

But she happens to work in a more solid-tumor realm, whereas, because these lymphomas grow so fast, the model I believe to be Gemini 3 and I are both on the same page that, for this particular situation, the literature shows that what they recommend is basically inadequate.

What they recommend is monthly appointments with blood draws. But again, the blood draw is just going to show the LDH, which, at that point, means you're probably starting to get sick. And then quarterly scans, which you don't want to have too much radiation from.

By the time they can see it on a scan, it's starting to be kind of large. Again, I think that varies a little bit depending on whether it's a solid-tumor cancer versus a lymphoma. I think they can spot small things there. But for the lymphoma, again, it's growing so fast.

Anyway, the bottom line is that the monitoring regimen they recommend doesn't seem to work that well. Most people who do have a relapse come back because they're sick. They don't start treatment because a diagnostic test identified the issue before they were sick.

Nathan Labenz

They come back because they're sick, which makes sense because, again, if it has a 24-hour doubling time, you're going to be sick pretty quick. So one thing we are probably going to do that we wouldn't have done otherwise, and that I didn't even know about otherwise, is this ctDNA testing. You send in the tissue, they do the fingerprint, and later on you do the blood test. You can do them periodically, however often you want to.

If you get a positive, especially after a bunch of negatives, we should also be able to use this throughout the treatment process to confirm with even more confidence than the usual scans that the cancer is gone. Maybe we could even—I haven't thought about this until right now—save our kids some radiation exposure from the scans if the ctDNA tests are coming back negative, because that should be the most sensitive test there is for whether this cancer is present in the body or not. If it's not at that level, you're probably not going to see anything on the scan either.

In fact, they say that sometimes, again, I'm just learning all of this from AI; I'm not Googling at all, with a mass like my son has in his abdomen, you can be cancer-free, but there's some scar tissue that's still left there. It will show up on the scan as though there might still be something there. It's kind of hard to say—we don't know. A lot of times, I guess, they have to do an extra biopsy to get that tissue, and then they come back and say, “Oh, no, it was just scar tissue. You're okay.” Or maybe it was cancer. But the point is, they don't know.

The ctDNA should be trustworthy enough to distinguish this from scar tissue on the scan. If the ctDNA shows no cancer in the blood, then you should be good to go. You should be confident enough that you don't need another biopsy, and maybe you could even avoid the scan. Regardless of that, I think this is going to be the main thing that I will do differently from what the doctors would probably advise.

Assuming—and it seems very likely, knock on wood, at this point—that we get through this treatment and he gets to the other side and he's well, then we're just, “Okay, God help us—don't come back, don't come back.” Another notable thing about this Burkitt lymphoma is that if it does come back, it typically comes back pretty fast, usually within the first 6 months, almost always within the first 2 years. Beyond that, you're pretty much thought to be in the clear. So we wouldn't even have to do this for that long of a time.

I think this will become standard. My guess is that this will become standard, but as of now, it is not standard. I think we will do something that the doctors have not advised and have advised against, and that is just get these tests so we know at the earliest opportunity—ideally before he becomes sick—if the cancer is, in fact, coming back.

Then, because we'll have these relationships already established with these major centers and we'll know what kind of clinical trials they have going on and, based on those early conversations, what they recommend, as Gemini 3—or the model I believe to be Gemini 3—put it: if ctDNA comes back positive on a Tuesday, you want to be infusing one of these next-generation drugs on a Thursday. You have to have that early monitoring to do that; otherwise, you're going to be really sick.

Another thing that it helped me understand is that all these drugs, including these next-generation CAR T cells and these bispecific immunotherapies, have side effects, and the side effects seem to scale with the amount of disease that there is. The reason they give this gentle chemo first is to debulk the tumor so you don't have too fast of a tumor breakdown and overwhelm your body systems.

These next-generation drugs cause something called cytokine storm syndrome. That's bad—really bad. But it happens because of the interaction between the drugs and the cancer cells. It's something that is very tough to manage clinically, but the literature Gemini 3 helped me understand is very clear that the greater the burden of disease, or the more cancer there is in your body, the more this problem can get out of control. The process of being treated itself causes a lot of the medical harm.

So catching it really early with the latest diagnostic tools and treating it as fast as possible should mean the best chance of successful treatment. Often, the doctor will say there's no data that proves that, but Gemini 3 is very confident, and it makes a ton of sense to me that you would want to be treating as early as you can detect any meaningful amount of cancer coming back. That's going to give you the best chance in multiple ways.

First, the side effects: this cytokine release syndrome issue is not going to be nearly as severe. Also, these cells evolve, right? Everything evolves. Cancer cells evolve to evade drugs. That's what makes the bounce-back cancers so bad, right? Basically, some cell in there—you might have gotten all the cells, but one of those cells evolved a mechanism to evade your treatment, and now it's coming back.

They have these additional lines that can potentially get it, but it might evolve a way to avoid that treatment as well. The fewer cells there are, the fewer at-bats that class of cells has to evolve that next-generation escape. So it seems like early detection is good. But in something with a 24-hour doubling period, which has this tendency, in these rare, extremely bad cases, to evolve additional ways to evade the latest and greatest treatments, you want to do that with the smallest detectable amount of cancer that you possibly can.

So that's basically our plan. We haven't done the second opinions yet, and we haven't chosen which class of drug we think is more or less good. Part of that is going to depend on the genetic information that we still need to get back. But I'm confident already, and we're only 8 days from when he got a pretty confident final diagnosis, which is notably 9 days from when he started chemo because, again, he was moving so fast. They started the chemo even before the full diagnosis.

Already, while my wife has been living at the hospital 24/7, huge thanks to my parents, as well as my sister and her husband and my wife's sister and her husband, for coming over and staying with our other 2 kids at various points, I've been at the hospital about 18 hours a day. There's been a ton going on there. With all the fluids, he's peeing every 45 minutes, and just helping him get up and use the potty every 45 minutes to an hour has been a lot unto itself. That's basically a round-the-clock job.

To say that I haven't had a lot of time to do other things is an understatement. And yet, already at this point, I have been able to make this much sense of what's going on. I wouldn't say I'm done; I'll certainly be doing more. But to be at day 8 from diagnosis, having started with a knowledge base of zero, and to understand the dynamics, the different kinds of this cancer that there can be, what those mean, what the different types of drugs are that are used to treat it, how they work, why they can't necessarily be used right now, which specific centers have trials going on, and which of those places have these second-opinion remote offerings that allow you to get into their system ahead of time—to have all of that mapped out at this point is just incredible.

To say that AIs are starting to accelerate science, I think there's a lot of evidence of that. But certainly, they are accelerating medicine. Just think about any new area of science you want to get up to speed in. I knew nothing 2 weeks ago about this, and now I'm conversant with true experts. Even at the bedside, with basically the AI doing it for me, I can hold my own in conversation and ask intelligent questions that the doctors don't feel insulted by. I can really make sure we're having a good meeting of the minds for my son's care. That has been invaluable—really invaluable.

A couple of final reflections. One thing I would say that this makes very clear to me is that the idea that AI is a bubble, that the value isn't there, that it'll never amount to anything—whatever different ideas people have—obviously, some VC investments are going to go to zero. Who knows? Maybe even OpenAI will default on some of its data center build-out payments at some point. Obviously, they're redlining that as much as they can. So there are a lot of different meanings of “bubble.”

But one thing I will say for sure is that AI has been indispensable in this process in all the ways that I've outlined, and the consumer surplus is absolutely insane. I'm paying $200 a month for GPT-5 Pro. I don't know what Gemini 3 is going to cost to access, but it won't be more than that.

Another deep-research report that I had ChatGPT do estimates that the insurance company is going to pay out $500,000 to $1.5 million for all this care that my son is getting over the course of his 6 months of treatment. Fortunately, I have health insurance. The cost of the AI, even at $200 a month relative to that, is somewhere between under a tenth of a percent and 2%. That's paying $200 a month, which is the most expensive AI subscription that there is, at least mass-marketed today, and it's under a tenth of a percent up to maybe 2% of the total medical cost.

It is GDP-destroying, I think, in various ways. The higher costs come with more complications. By using the AI all the time, I am helping to minimize mistakes and complications. By getting on top of the mold problem, I'm minimizing the risk of infection, which is another complication that could drive costs up.

We might have moved, if we weren't able to gain confidence in our current care—which, fortunately, I have now—to another city and gone to an even more expensive center, rented a house, and taken on whatever other expenses we would have had associated with that.

Nathan Labenz

And that very well might have been worth it if we didn't have the confidence to know that what we are getting is basically as good here as it would be at any of those places. At least for now, while we're in this kind of main protocol.

It's paradoxical. The economic situation here is weird. Consumer surplus is insane. It's GDP-destroying.

What would I have paid if they were price discriminating? That's how you'll know if they ever really get evil. That's when you'll see it, when they really were like, “Nathan, we know your son has cancer. What are you really willing to pay for ChatGPT Pro during this time?” I genuinely think I'd pay $10,000 a month and feel pretty good about it. And even then, it would only be 5% to 10% of the overall cost of his treatment.

The AI doctors are basically here. They can't do everything, obviously. They can't do procedures or physical exams, and they obviously just can't order and prescribe stuff. But in terms of analysis, quality of reasoning, consistency with what we've seen, and explanations, you really do have AI doctors now, and the consumer surplus on them is absolutely insane.

I think all anyone needs to know about AI is: if your kid has cancer, do you use the AI or not? The answer is, you use it a lot more, and you get a lot more value out of it in the most stressful, scary, overwhelming context that you've ever been in in your life. You turn to AI a lot more than you do otherwise, and I think that just makes clear that the value here is truly undeniable.

There's no revealed preference stronger than what you do when you're trying to take care of your kid. I can tell you with 100% confidence and sincerity that the AI came through huge for me, for us, and for him in this process.

Another kind of higher-level meta-reflection there is that there are no decels in the pediatric oncology unit. I've always been of mixed mind on AI, and I still am. I've always said, “I want my AI doctor. Nobody better get in the way of my AI doctor.” I feel that as strongly as I ever have.

What I think I feel even more strongly now is the idea that delays in AI progress come at a real cost to real people. One of the more painful outcomes of this would be if the worst were to happen and we were to lose our son, which, again, I think is very unlikely at this point, but is still scary and definitely still in the back of our minds. For my wife, it's in the front of her mind, I think, all the time. I'm able to keep it maybe a little bit more in the back of my mind.

But if something like that were to happen, how much would it suck to be like, “This all happened 1, 2, or 3 years before an AI might have been able to figure out how to cure this disease”? That's going to be even more the case with real N-of-1 conditions.

We have the AI doctor. We don't yet have the full-on AI oncology researcher who I could say, “Here's everything that there is to know about my kid. Go find a cure for a case that is sort of case-report-level rare.” We don't have that yet, but we're probably not going to be waiting too much longer for it. I always recognized this on an intellectual level, but I have definitely come to feel in a much deeper way that delaying that means people will die who wouldn't otherwise have to die. That's something that we should take very seriously, and we should not delay unless we really have to.

At the same time, of course, I do stand by my decision to sign the recent call for a ban on superintelligence. I think, at least for my son, given his favorable prognosis, that an uncontrolled recursive self-improvement loop—an intelligence explosion—is probably more likely to kill him, and a lot more people besides, than his cancer is at this point.

I'm not saying we should let the companies run wild, or have no governance, or enter into recursive self-improvement loops without a much better understanding of what's going on internally in the systems than we have today. But I do think it's important for me to be intellectually honest about the fact that delays in the timeline to the AI oncologist who can take your N-of-1 case and have a real chance of going out and figuring out what to do about it for you—the cost of delaying that has really been brought home to me and made so real. I don't want to see that delayed any more than we absolutely have to for existential security reasons.

I've got to have 1 comment on China. Can we race China to cure cancer? Seriously, I've said this before, but this is a race worthy of 2 great civilizations, and everybody benefits. The whole rest of the world would benefit.

I seriously want to see a gold-medal tracker for cancer cures. We know that the Chinese government loves to compete for gold medals. The U.S. obviously doesn't want to be anywhere but number 1 on the leaderboard. I want the cancer cure Olympics, and this is the race that we should be having.

Obviously, it goes beyond cancer as well, but we should have that race. There was already a war on cancer, and it didn't necessarily amount to as much as we might have hoped. But this time, I think if we do declare a war on cancer, we might win that war. Then there's more beyond that that we obviously will want to do.

I think let's not demonize Xi for speaking about extending healthy lifespan. Obviously, there's a lot to not like about Xi. There's even more not to like about Putin, and there's even more still not to like about Kim Jong-un. It's certainly counterintuitive to see those 3 in conversation with each other and feel like there are good ideas coming out of it.

But when they were understood to be talking about how, with all this technology, AI, and biotech, we could imagine extending the human healthy lifespan to 150 years, that's not crazy. They are not crazy for talking about that. They should be talking about that. I want the president of the United States to be talking about that.

Why are we so behind in terms of our vision, our expectations, and our standards for progress that we have Xi, Putin, and Kim Jong-un leading the way? It's an absolute failure of imagination and a failure of leadership that we don't have our own version of that.

Not only is that, in my view, a shameful omission from American and broadly Western political leadership unto itself, but we then risk, because these guys are obviously so bad in so many other ways, that an idea like extending healthy human lifespan gets understood as a bad thing, like only the kind of thing that desperados or authoritarians would care about. That could not be further from the truth.

We should look back on the progress that we've made in extending healthy human lifespan and view that as an absolute, unalloyed good. These are the fruits of all of the prior generations of humans working together to get us to this point. It's only been the last 1 to 2 generations that my son's condition could be cured, so that's an incredible thing.

But the idea that we're potentially going to get negative about that idea—the very idea of extending healthy human lifespan—because we're only hearing about it from some of the world's worst people is dangerous. The right thing to do there is not to throw out the idea that we want better, that we want better for ourselves, for our kids, and for our parents.

The right thing is to start articulating our own positive vision and actually start trying to deliver on it with urgency, so that when they talk about it, they sound like also-rans, as opposed to being ahead of us and our sort of mimetic immune system feeling the need to downgrade and dismiss the idea because of its source. That has also come home to me in a really powerful way.

I guess my final thought is that I think the medical system has served my family very well here. Again, just the fact that there is a protocolized treatment that works so well, the fact that they were able to expedite so much of the scheduling to get enough clarity, and the fact that they were willing to flex the rules and do the chemo—to get it started before the final diagnosis—have all been important.

A couple of mistakes were made along the way, or you could at least have imagined catching it earlier if somebody had been a little bit more astute. But overall, I do think we've been served very well.

At the same time, I do feel like the structures on which the medical establishment runs will soon be handcuffs. There is just so much information out there. The ability for AI to power through that information with one particular patient in mind absolutely is already more than what human doctors can do for any single patient.

I think this means that standardized medicine, while it has been good—because how many people really could deviate from the best-proven protocol that was double-blind and all that good stuff, and actually get more expected value from it? Very few people can do that historically. I flatter myself; maybe I could have, just with Google and 6 months' worth of work.

But now we're entering an era where anybody with Gemini 3 access can do that and can really start to come up with their own plans that are going to be well-grounded, conscious of side effects, and conscious of drug interactions. Because the patient cares so much more than anyone else, obviously, and because the AIs have so much more processing power than you can otherwise bring to bear on anything like this, you're going to see patients showing up with ideas that very plausibly do beat the standard of care.

I have not been able to find anything like that in my case, to be clear. That's in direct relationship to the fact that the protocol is so well-established and the prognosis is so good. If the prognosis were not nearly so good, I think even now I might be able to use Gemini 3 to come up with a better plan than the standard of care.

You're just going to have this happening over the next year or 2, but honestly, just 1 year.

You're going to have this happening by the millions across the country, and the medical system is not prepared for it. I remember the classic line, “Rage, rage against the dying of the light.” I would say, “Rage, rage against the lack of a right to try.” This has always been true, but it's becoming absolutely critical now that we allow people to take more ownership of their own medical care and allow them to drive some experimentation.

Yes, some people will use that badly, but if they just use the best AIs that are available at retail today to come up with a plan, often enough they're going to have a good idea. And I think often enough the doctors will come to recognize that, too. At this point in time, I wouldn't say you want to be talking too much to your doctor about the fact that ChatGPT told you this, because they've got a lot of baggage. But a year or two from now, they're going to understand, because patients are going to make them understand that the AIs do have the ability to reason about this stuff at a high level and to add value to many standard-of-care protocols.

I guess, actually, I do have one, right? The ctDNA one is one, because that's not standard and I think it's an obviously good idea. But even more—and not just diagnostic, not just monitoring, but active treatment plans right now—that's going to be brought home to the doctors.

But then the rules are going to really stand in the way: the rules about how that's only available in this clinical trial, that's not approved, and all these liability concerns around, “If I, as a doctor, do the standard of care, I'm not at risk of being sued because I did the right thing according to the profession. That's the standard of care. That's what I can rest easy at night, and that's what I've been legally advised protects me.”

But that doesn't mean it's in the best interest of the patient, and I think that's going to become very clear. So I think we're going to need reform at the level of simply giving people more legal access to try things.

Those rules had a point. I'm not denying that. But with the AI ability to advise everyone, the quality of decisions that patients are going to be able to make for themselves is going to be dramatically improved. We're going to need reform on liability. We're going to need some sort of different standard, where just because you did something that isn't the standard of care doesn't mean you were negligent.

In fact, you probably, in many cases, were doing better than the standard of care if you took proper steps and had AI review the case. If you approached it the right way, standard of care is not the end of history.

And then data—we're also going to need reform of data. We opted our son into the research program where they're going to track him and whatever, but the battle to get data just from the hospital staff to us in real time has been a battle. I have no idea exactly what they're going to be sharing or in what format, but I can pretty much guarantee that it is dramatically suboptimal.

Especially if you imagine a sort of more right-to-try environment, where people are doing a lot more different things, that will create much more diversity of data that's out there. Making that stuff available for AIs to reason over gives us the opportunity to create better outcomes in the short term—a richer data environment that will inform even more and better analysis, which will inform even more and better outcomes.

The doctors aren't quite ready for it, but I think they will be convinced by their patients. Over the course of this last week, I think I have won the confidence of our oncologists. Again, I wouldn't be able to do this by myself, but because of the value that I'm able to get from AI, they now trust me.

I think the questions I'm asking are not just, “I Googled this term and I think this is the thing.” I am actually doing my homework, and I do have real expertise—not in my head, but at least supporting my efforts. I think the doctors will get there. The rules, though, on these various dimensions are going to have to catch up.

So that's it for now. Really, really crazy last 3 weeks. Scary and emotional at times, for sure. I'm really grateful for all of humanity's collective effort to get to the point where there is an effective treatment.

Can you imagine the pain of having this happen and there not being an effective treatment, knowing that just 2, 3, or 5 years from now, an AI might be able to sort it all out just for your individual case? We absolutely should not delay that.

Obviously, existential risks still matter tremendously, but the AI doctors are not just here—the AI medical analyst that can really deliver strong, personalized medicine is coming online, too. And it's an amazing thing.

Thankfully, the prognosis is really good. It's going to be a rough road. Chemo is not easy. That's another thing I think could get better for sure once we really understand this better.

I'm pretty sure that a less aggressive treatment would work for a lot of people. But for now, we don't quite have that information, and I feel like it is probably worth doing the full harsh treatment because that is the thing that's proven to work best.

Fortunately, for now, the prognosis is good. So wish us luck. I'm sure we'll have an update before too long.