Original blog here.
Here’s a Lewis Black–style, black-humor rant-essay built on the “meta-overfitting” PDF you attached—with nods to the two rival LLMs, Notebook LM and ChatGPT-5.
Overfitting:
Or, How I Learned to Stop Worrying
and Love the Data Tsunami
You know what really cooks my neurons?
When people say “AI will save precision medicine!” as if all we had to do was dump a million-by-million-dimensional matrix into a GPU the size of Cleveland and—boom—out pops the best drug.
That’s not intelligence. That’s what you get if you cross a slot machine with a spreadsheet.
And yet, every week, some multimodal prophet waves a PowerPoint about how “early fusion of histology and transcriptomics” will banish overfitting. EARLY fusion. Not late. As if the problem all along was we were just showing the data to the model at the wrong time of day.
Two Gurus Walk Into a Bar…
In the blue corner, Notebook LM—the guidance counselor of AI.
It pats you on the shoulder and says, “Yes, dear, overfitting is a valid concern… but don’t fret! We’ve got multimodal data, calibration, foundation models, weather metaphors—everything but a therapy dog. We’re moving from ‘point’ predictions to ‘path-level’ journeys! It’s all going to be fine.”
In the red corner, ChatGPT-5—the grizzled statistician.
It leans over a sticky bar table and growls, “Kid, you’ve got ten million features and ten thousand patients. That’s not a dataset, that’s a mugging. And by the way, your perfect ‘survival’ label is leaking hospital ZIP codes. Fix your damn cohorts.”
Both claim to follow the gospel of Fukushima—Tempus COO, multimodal evangelist, patron saint of “lab-in-the-loop.”
But while Notebook LM sounds like it’s running a TED talk for venture capitalists, GPT-5 is the guy in the corner muttering, “Check the residuals before they check you.”
Dimensionality: The New Original Sin
You can almost hear the thunder in Lewis Black’s voice:
“We’re not just overfitting—we’re hyper-fitting! We’ve got ten-to-the-seven predictors for ten-to-the-four patients, and we’re shocked—shocked!—that the model memorizes who was treated at Sloan Kettering in 2016!”
Notebook LM waves MOBER at the problem—Look! We aligned cell lines, PDX, and clinical tumors!
GPT-5 says, “Sure… you aligned them. Right into the same ditch.”
Label Leakage: The Joke That Never Dies
Fukushima, to his credit, hints at it: sometimes your model isn’t predicting survival, it’s predicting which clinic still uses fax machines.
Notebook LM politely ignores this, preferring to talk about the elegance of probabilistic forecasting—GenCast for oncology!
GPT-5 says, “Great, now we have a perfectly calibrated probability… of over-treating the patients from the Midwest.”
Interpretability: The Warm Security Blanket
We get endless sermons about interpretability.
SURVPATH links EMT pathways to image tokens—fantastic.
But as Lewis Black might bark:
“You think a saliency map that lights up in fuchsia when the model cheats is going to save you? It’s like catching your dog chewing the sofa and saying, ‘Look! It’s interpretable!’”
Path-Level Forecasting: The New Flying Car
Everybody now wants to “model the journey.”
Notebook LM loves this—comparing it to traffic jams and weather systems.
GPT-5 says, “That’s just regularization with a pretty story. You still have to drive through the damn storm.”
Benchmarking: The Siren Song
Both admit that open-source benchmarks like Virchow and PRISM are a double-edged sword.
Lewis Black would call it a circular firing squad: everyone optimizes to the same leaderboard until the models can diagnose the dataset but not the patient.
Bottom Line (Delivered While Pointing Wildly and Yelling)
Look—overfitting isn’t a minor technical footnote; it’s the statistical equivalent of termites in the foundation.
- Notebook LM croons that our new multimodal mansions are termite-proof.
- GPT-5 warns the wood still rots if you don’t keep an eye on it.
The truth, as Fukushima himself implies, is that we need both:
-
The visionary idea that multimodal, interpretable, path-aware AI can actually advance precision medicine.
-
And the grumpy auditor who demands external validation, drift checks, cohort de-confounding, and a Plan B when your “foundation model” face-plants on Tuesday’s biopsy.
Or as Lewis Black might snarl, jabbing a finger at the screen:
“If your survival model is so clever it can predict the weather, the traffic, and the patient’s death—all at once—maybe it’s just memorizing the zip code.
And if that’s your definition of intelligence… God help the patients.”
Citation: Built on Bruce Quinn’s Meta-Overfitting blog comparison of Notebook LM vs ChatGPT-5, plus references to Fukushima’s essays on MOBER, SURVPATH, calibration, and multimodal translation.
