This is technically correct. However, this framing of the NordICC trial is kind of misleading. It is indeed true that it did not find a significant all-cause mortality reduction, but it was insufficiently powered to do so.
It did however reduce colorectal cancer relative incidence at 10 years by 30%, and CRC relative mortality at 10 years by 50%.
But because the corresponding absolute numbers are quite small, they don’t show up in the all-cause mortality given the sample. While there’s some morbidity associated with polypectomy, the numbers can’t fill the entire gap.
Sylvain Ribes
Karma: 0
Re: Multi-lingual capabilities
I have not reviewed the recent literature but last I checked (summer 2024), it seemed to me that the literature overwhelmingly drew the conclusion that English for the most part acted as a sort of mentalese, with translation neurons layered across the network, but mostly on the outermost layers on both sides (close to input and output)
It seems in this way that Haiku might be distinct from the heretofore studied models?
Do you think this is just a natural consequence of scale, with the understanding of a sort of a “universal grammar” naturally emerging, or is there anything to Claude’s training that’d make it qualitatively different from other models in this regard?
The procedural mortality from colonoscopy is well-characterized at roughly 0-1 deaths per 10,000 (NordICC saw zero in 12,574).
The per-protocol CRC mortality reduction was 15 [7-22] per 10,000 absolute. So even at the low end of the CI, known procedural harms eat at most ~15% of the benefit.
For the net to come out neutral or negative, you’d need an unidentified mechanism of death an order of magnitude larger than every characterized harm. Theoretically unfalsified, yes, but not actually how one should reason about established procedures.
Take for example appendectomy, there’s no powered ACM RCT of appendectomy vs no treatment either, and there never will be (although appendectomy vs antibiotics is very much on the table, other issue!).
The ‘cause-specific reduction doesn’t prove ACM benefit’ idea is usually fine imo, and applies in cases like PSA screening or mammography in 40-year-olds, where the cause-specific benefit is small, overdiagnosis is large, and downstream treatments carry meaningful mortality. Colonoscopy isn’t structurally that case: the cause-specific benefit is large in relative terms, polypectomy harms are well-mapped and small, and removing adenomas isn’t analogous to treating overdiagnosed prostate cancer with prostatectomy.
Pattern matching from one to the other doesn’t work. For similar reasons, the breast analogy doesn’t hold so well.