HedonicEscalator
This is a delightful article about one of my favorite topics! I (somehow!) hadn’t heard of Provenge before this, and I enjoyed the dive into the previous failed attempts at cancer vaccines.
There is a secret third kind of target antigen besides TAAs and neoantigens, which I assume you left out due to space constraints: xenoantigens, or viral antigens from cancer-associated viruses. BioNTech currently has ongoing Phase 2⁄3 trials for BNT113, targeting HPV-associated cancers. Not sure how that’s going.
Three other trials were run using the same cancer vaccine design process. In early-2025, it failed in first-line metastatic melanoma. In mid-2025, it stalled in adjuvant muscle-invasive bladder cancer, after a “safety event [was] observed in the safety run-in population”. Finally, in November 2025, BioNTech disclosed in its third-quarter report that the trial in adjuvant colorectal cancer had crossed the boundary for futility at its first interim analysis, though this trial continues with the customary “the data are not yet mature enough to draw reliable conclusions about efficacy”.
Your links in this paragraph are broken, by the way.
Finally, for anyone interested, I wrote a post discussing Rosie’s story a few weeks back.
I was involved in the online art subculture before and during the early emergence of AI art. “Anti-AI” was a hyperstitious cascade enforced by the usual mechanisms of cancel culture; whenever an artist said they liked or used AI, they were harassed by online mobs until they backtracked. For practical reasons that I’ll get into, some degree of anti-AI sentiment was probably inevitable in the majority of online art communities. But I resent the narrative that a hardline anti-AI stance was ever anything close to universal. No, the artists that disagreed were simply bullied into silence, or No True Scotsman’ed as infiltrating techbros.
I actually find the anti-AI stance more sympathetic now, mostly because I matured enough to recognize what you pointed out, “they cared about the social role of the tech, and they didn’t care about the technical details.” Most people mad at AI art were upset because of practical complaints, like “AI art makes it harder to make money” or “AI art looks good, and that makes them feel bad.” Some of these complaints are selfish or embarrassing, but at least they’re understandable.
But people rarely ever cited these understandable complaints in their arguments against AI art. Instead they dressed up their grievances in incoherent ideological stances on, of all things, copyright.
I don’t doubt some people have deeply held principles about intellectual property. But before AI art, most online art communities hated copyright. Some of the nascent pro-IP crusaders made a living off selling fanart in defiance of copyright law. Then the moment something they don’t like rolled around, they started grasping at any flimsy excuse to declare it not only unethical, but also illegal, dearth of established case law be damned. I watched this all go down in real time, and as you can tell, I still have a chip on my shoulder over it. I was into AI art when it was still a dreamlike, distorted tech demo, and I was into AI art when temporarily embarrassed professionals on Twitter decided to launch a doomed crusade against it, and I was an artist before all of that. Copyright? Beauty wants to be free.
This is an interesting perspective! No, I wasn’t thinking about AI gender specifically while writing this article, although I’m working on an upcoming article about AI as the children of humanity that I expect you’ll enjoy.
The immediate context for why I wrote this article is that, per relevant xkcd, I saw a person claim that it was impossible for anyone to produce both sperm and eggs, thought “That doesn’t square with my understanding of reproductive biology,” and investigated to discover 1. I was right and 2. many people, including rationalist-adjacent people knowledgeable about biology, were surprised that I was right. I correctly predicted that others on LW would also appreciate learning about this.
The big-picture context for why I wrote this article is that I am a biology nerd, and a biotechnology nerd in particular. As my username suggests, I am interested in how we can use technology to create new kinds of beauty, pleasure, and purpose. Assisted reproductive technology (ART) is something that has mostly been a “normal medical treatment” for people with fertility issues, and for almost tautological reasons (most people aren’t sick), normal medical treatments are rarely socially transformative. (Exceptions involve treatments given to healthy people, such as vaccines and birth control, or particularly versatile inventions considered in aggregate, such as antibiotics). But augmentative ART is growing more popular: through egg freezing/IVF used to extend the fertility window of healthy women, through surrogacy, through the emerging practice of embryo selection, and more speculatively, through germline editing, IVG, artificial wombs, and lab-grown organs.
People talk about AI as a “profoundly abnormal technology,” and if the predictions of those people are correct, it certainly would be. ART will also be profoundly abnormal. Perhaps not as abnormal as artificial superintelligence, but abnormal enough to be an inflection point in history. Sufficiently accessible IVG and artificial wombs would render sexual reproduction obsolete. It would irrevocably alter fundamental aspects of parenthood, population dynamics, and gender that predate human evolution. This idea scares some people—it scares some people a lot—but as you might guess, while I am wary of potential negative effects, I am optimistic about the possibilities for new kinds of flourishing.
Studying gametogenesis in naturally occurring human hermaphrodites can tell us more about how to accomplish IVG, and how to artificially induce novel reproductive capabilities in human bodies. For those of us not working in reproductive biotech, it helps us develop better models of the plausible timelines of such emerging technology, which is useful to know about any potentially transformative technology. And it’s cool.
Did you read the article, or are you assuming its contents from the headline?
The article does not dispute that women are physically weaker on average, and is instead criticizing social norms discouraging women from exercise.
By ages 8 to 12, girls are already about 20% less active than boys, a difference that translates into lower cardiorespiratory fitness, weaker hand-eye coordination, and lower perceived competence in physical education. By adolescence, girls drop out of sports at six times the rate of boys, further widening these gaps. … One recent survey of over 400,000 adult Americans found that only 33% of women met the weekly recommendations for aerobic exercise, and 20% for strength training, compared to 43% and 28% of men, respectively.
I likely disagree with the author on the degree to which traditional gender norms are responsible for these differences, but it seems uncontroversial to say that gender norms play at least some role in discouraging women from exercise, and that exercising more would be good for women (as it would be for a somewhat smaller percentage of men).
It wouldn’t be unusable, vice grips would just become a default kitchen tool.
On her website she wrote:
Some forty years ago, when I was an undergraduate at Oxford, I had to have two teeth out. Because I seriously disliked the idea of having my consciousness invaded, even by the sensations of injection and anaesthetised extraction, I thought about the psychology of pain. As a result, I had the two teeth out with anaesthesia and without any unpleasantness to myself.
People who wish to relate this to something already know about and hence of no interest for further research often suggest that this was self-hypnosis. Actually it was not. In the case of the extractions it would be possible to maintain that I had been giving the matter so much thought that some sort of self-hypnosis had resulted, but several years later, with no preparation of any sort, the technique worked just as well when I had several holes drilled for fillings.
On this later occasion, the dentist (a new one) said afterwards, ‘Well, do you feel pain or don’t you?’ He said that he could usually tell when he was drilling on a nerve by the salivation, but he could not in my case. This illustrates the fact that this technique, or techniques that might be developed from it, eliminates some of the physiological side effects of pain and may eliminate the shock reaction associated with severe injury or major operations.
I don’t believe there has been any independent corroboration of this claim, so make of it what you will.
The AI angle, since many people are hyping the “AI for pancreatic cancer” line:
Yes, deep learning is used in the development of the mRNA vaccine. NetMHCpan, a small neural network, is used to help select immunogenic neoantigens (choose the most promising mutant proteins to target) for autogene cevumeran.
Autogene cevumeran paper https://www.nature.com/articles/s41591-024-03334-7
Most of the computational pipeline detailed in the paper consists of traditional tools, in line with what I’ve written about AI and personalized mRNA vaccines before. Deep learning is extremely useful for some things, but it should be understood as a specialized tool, not a silver bullet, for now.
AI & mRNA blog https://hedonicescalator.substack.com/p/did-paul-conyngham-really-use-ai
I’m not sure if deep learning was used in the development of daraxonrasib. A brief glance at the paper and previous work shows plenty of references to traditional computational tools, but nothing that stands out to me as modern DL.
Daraxonrasib paper https://pubs.acs.org/doi/full/10.1021/acs.jmedchem.4c02314
Previous work https://pmc.ncbi.nlm.nih.gov/articles/PMC10474815/
The company behind daraxonrasib, Revolution Medicines, is quite enthusiastic about ML. They recently made a deal with AI drug discovery platform Iambic Therapeutics. I don’t doubt Iambic’s tools will soon prove useful, but it’s fair to say “deep learning for drug discovery” is still in the early stages of development.
Iambic announcement http://iambic.ai/post/revolution-medicines-collaboration
Two exciting recent innovations on pancreatic cancer!
First, the personalized mRNA vaccine autogene cevumeran, developed by BioNTech and Genentech, just reported 6-year follow-up results from their Phase 1 clinical trial. 16 patients were treated, 8 were responders (showed signs of immune reaction to vaccine), 8 were non-responders.
7⁄8 responders (87.5%) survived 6 years after surgery, 2⁄8 nonresponders survived (25%).
AACR meeting notes https://www.aacr.org/blog/2026/04/20/live-updates-from-the-aacr-annual-meeting-2026-monday-april-20/
The most important result in this small trial is that vaccine response is strongly correlated with better outcomes. But for context, the trial was restricted to patients with operable pancreatic cancer. Patients diagnosed with stage 1 or 2 pancreatic cancer have a 5-year survival rate of 12%. Patients who get their pancreatic cancer surgically removed have a 5-year post-surgery survival rate of 20%. This makes the overall 6-year post-surgery survival rate of 56% among the 16 trial patients pretty impressive. Keep in mind that the trial patients may have been healthier than average for other reasons, and small n is small n, so we shouldn’t be too hasty until we see Phase 2 and 3 data.
Source on survival rates https://www.pancreaticcancer.org.uk/information/just-diagnosed-with-pancreatic-cancer/if-you-can-have-surgery-to-remove-the-cancer-early-pancreatic-cancer/prognosis-if-you-can-have-surgery/
Second, a small molecule drug I almost missed in the mRNA hype but arguably even cooler, the tri-complex ras inhibitor (!!!) daraxonrasib, developed by Revolution Medicines. Ras is a protein involved in many cancers, with PDAC (the most common pancreatic cancer) being especially dependent on ras, but it has historically been considered impossible to target due to its chemical properties. Daraxonrasib is, as far as I’m aware, the first drug to target generic forms of ras. It does so with an exotic “tri-complex” strategy involving gluing a different protein, cyclophilin A, to ras in order to disable it. Crazy stuff!
Phase 3 results found that daraxonrasib doubled survival time among patients with metastatic pancreatic cancer, 6.7 months to 13.2 months (p < 0.0001). Side effects are kind of nasty, but “well tolerated, with a manageable safety profile” by advanced cancer standards.
Revolution Medicines announcement https://ir.revmed.com/news-releases/news-release-details/daraxonrasib-demonstrates-unprecedented-overall-survival-benefit
Derek Lowe coverage https://www.science.org/content/blog-post/progress-against-pancreatic-cancer-part-one
Lecture by executive/scientist at Revolution Medicines https://www.youtube.com/watch?v=bU3IwuDJx24
Many of the incredible advancements we’ve made in oncology has been won slowly through another pathway we can target, another drug buying a few more months, another targeted technique to mitigate side effects. Progress is made one step at a time, and these are big steps. I’m excited for the future of biotech.
Nanopore basecalling is a great example of a technology that genuinely depends on deep learning! (It also fits my characterization of problems suitable for deep learning: predicting bases from electric signals is highly complex, and you have massive amounts of labeled training data).
When it comes down to it, “usually preferred” is hard to quantify. What I’m trying to get at is that when I read the methods section of (post-2020) papers in this field, I almost always find that they only mention the use of traditional algorithms, not deep learning ones. Even fewer discoveries are dependent on deep learning. Meanwhile, there are scores of papers floating around enthusiastically presenting DL algorithms which don’t ever see use in the lab. It’s easy to overestimate the role of “AI” by skimming through abstracts.
It looks like DeepConsensus is being used, but it’s not obvious to me how much better it is than traditional methods. Is it more of an incremental improvement useful in specific situations, or more of a paradigm shift no analyst can ignore? I’d love to hear someone working in a related field explain how often deep learning tools are used in practice.
Thanks for the response!
Can you explain to me what your counterfactual vision for “the world if non-sunburn exposure causes cancer” looks like? For example, do you expect that oncologists would track whether melanoma patients are in the subgroup of “white outdoor-working sunscreen-abstainers [without any sunburn history]”, and report those statistics? Please give some specific examples of what data you’d expect to see, and how we’d be able to gather and verify it.
Can you also explain to me what background you have on carcinogenesis? For example, are you familiar with initiators vs. promoters, the two-hit/multi-hit model, common oncogenes and tumor suppressor genes, etc.? It would be helpful to understand what your internal model of how cancer develops looks like currently before I write any more about the topic.
It’s not controversial. I don’t think anyone knowledgeable in oncology would dispute that sub-sunburn UV exposure significantly increases risk of skin cancer. My aim here was to explain why we can be confident of this belief in as straightforward of a way as I can.
I haven’t looked at your links in detail, but the first and second cover the same ground as the first and second points I made (observable DNA damage + cancer from indoor tanning). As for the third, I wanted to avoid resting any points solely on cohort studies because I didn’t want to get into too much of the controversy there. Though I ended up getting pulled into it on Reddit anyways.
Of course that cuts both ways, since if it were a pure win to be protected from the sun, we would all be dark skinned.
Not necessarily.
We know that humans need sunlight for optimal Vitamin D production, so this does happen to be true. But your logic here doesn’t hold up. Synthesizing melanin is metabolically costly. Cave creatures, which have no use for pigmentation but also incur no obvious direct disadvantages from it, often evolve albinism because of the benefits of repurposing melanin precursors for other needs.
Humans aren’t cavefish, but the principle applies. If you don’t need melanin, making it is a waste of energy, and benefits from sunlight aren’t necessary to explain the evolutionary pressure against pigmentation.
Cohort studies claiming benefits from UV exposure are not credible. It is impossible to separate UV exposure from positive traits correlated with UV exposure, such as exercising or having a social life, and no, you can’t just control for it.
It is true that your risk of dying from melanoma is low, and not worth freaking out too much about.
You should do a blinded test of your sun exposure with identical placebo sunscreen to see if the UV exposure itself is necessary for positive effects on your mood, for science.
IQ is normalized to mean 100, standard deviation 15 from a sample population of test takers, usually matched by age. The mean is set to 100, by design.
Technically, since widely used norming populations are usually drawn from developed countries and exclude people with severe disabilities, the “actual” mean IQ, if you were to test everyone in the world, is lower than 100.
“Miraculous oocyte epigenetic reprogramming” doesn’t have quite the same ring to it as “miraculous virgin birth.”