If I had to choose between Derek Lowe (author of the anti-amyloid-research article who is also highly experienced and skilled in pharma) and Scott Alexander/David Schneider-Joseph (psychiatrist and AI engineer, respectively), all my priors suggest Lowe gives better advice.
"I am David Schneider-Joseph, an engineer formerly with SpaceX and Google, now working in AI safety. Alzheimer’s isn’t my field, but I got very interested in it, spent six months studying the literature, and came away believing the amyloid hypothesis was basically completely solid. I thought I’d share that understanding with current skeptics."
6 months of reading literature when you don't know how to read biomedical literature isn't very confidence inducing. I know this site really likes it when smart outsiders come in and disrupt the status quo, but... probably not in this case.
This frequently comes up as a critique of my article, but I don't claim to be disrupting the field as a smart outsider. Rather, I looked at the field and concluded that the experts seem to know what they're doing. Derek Lowe is very much in the minority on this matter.
No, he's not (I work in pharma at a company that does basic and applied research on Alz). It's more correct to say there are several camps, but the camp promoting amyloid plaques as the causative/driver for Alz has struggled greatly to come up with evidence supporting its position.
Is your view that amyloid is actually a minority view among researchers? That seems completely wrong based on basically every conference proceeding I've viewed and the volume of papers and citations I've examined.
If your view is merely that there is a "camp" of experts that disagrees, then sure, but in that case, I do not think it is honest to frame this as a choice between believing in the authority of a single expert from that camp, vs. the (lack of) authority of me, a non-expert.
(I also think your read of the evidence is wrong, but I won't restate the arguments in my article.)
My opinion is that amyloid-as-cause moved from a majority to a minority view over the past few years, but it's not yet reflected in the literature (the entire amyloid establishment isn't going to give up its dominant position easily).
Also, I didn't say anything about the evidence (I don't have a "read" on the evidence, because I don't read Alz literature). My point is entirely that my priors indicate that Derek is a more reliable reader than you.
>My opinion is that amyloid-as-cause moved from a majority to a minority view over the past few years, but it's not yet reflected in the literature
>I don't have a "read" on the evidence, because I don't read Alz literature
these two sentences seem contradictory to me. i am not sure how you would keep up on the research (to know it's moved from majority-held to minority-held view), and know that the move is not reflected in the literature, without reading the literature.
Most scientists who are not experts in their field don't read the literature for a field directly. Instead, they synthesize their opinions about the field by consulting experts, and weighing various sorts of evidence. In my case, I work in an adjacent field and see presentations from scientists, have casual conversations with them, and read the news articles in major journals.
The raw literature for alzheimer's, as well as biomed in general, is not really easily interpretable. It's rife with errors, misleading statements, and intentional obfuscation.
You said the camp promoting the amyloid hypothesis has struggled greatly to come up with evidence to support its position. What did you mean by that if not a read of the quality of the evidence?
Why do you continue to frame this as a choice between a single cherry-picked expert's opinion, and my own non-expert opinion? Either fairly represent the spectrum of experts' views, or decide based on the actual evidence and arguments.
My estimate of the quality of the evidence is based on daily discussions with people who work in that field and reading summary articles in major journals. I typically don't read raw scientific articles directly- those are aimed at people in the field. Instead, my understanding comes from a synthesis of expert opinions weighted by my own priors (based on 30+ years in the field). Derek's opinion is now the prevailing one that I hear from a wide range of researchers.
I've seen this happen before, btw- overturning establishment paradigms, especially ones where the underlying etiology is complex- is extremely hard and often takes decades of experimental results.
What started as an argument to ignore arguments and evidence and instead rely on authority, seems now to have morphed into an argument that we should ignore the authority of the establishment, because of your own personal assessment of the evidence (which you have not yourself read) and your own personal synthesis of conversations you've had with researchers you've personally come into contact with (despite this being apparently unrepresentative of objective measures of typical researcher opinions).
Arguing from authority really only takes you so far when it ends up as an appeal to your personal experience. I'd rather you either address the arguments directly, or drop the dubious appeal to authority.
Your making an argumentum ad verecundiam which even if you are right means we have to discredit it.
It’s poor science to make an argument on authority, if you know the science then you should be quoting the published research and not relying on others so called expertise.
That's the dream (all science communication should be based in the raw scientific data as published in the literature) but it's not how things work now, nor is it even practical. Instead, we rely on experts (authorities) because our priors tell us that experts are usually the most able to interpret the complexity of literature.
One thing that has gone mostly unsaid in this thread is that scientists lie when they publish. Not every scientist, and not every publication, but significant fraction of papers contain true errors or omissions intentionally added by the paper authors. Learning how to read a paper and translate the bullshit takes some time, and usually requires a fairly deep understanding of the state of the art of the field.
I don't think you're obliged to discredit any argument made from authority (we're not making true logical arguments here, we're working in a real world space with ambiguouity).
Then the next question becomes 'which expert to trust'? which is a subjective judgement; personally, after polling many different experts, the "go look for other causes of alzheimer's" experts seemed to have the most compelling biological narrative.
Wow, it sure didn't take long to show a complete lack of familiarity in the field. It seems like that's going to be a real weakness with LLMs based on volumes of material that are later discovered to be semi-fraudulent and unmotivated by scientific principals.
I'm responding to a random comment: I was in molecular biology, but >20 years ago. Your article immediately presents as someone who's acquired reading expertise in a biological/medical subfield. Second, your initial survey presents as a "brittle x": AB is the proximal cause all by itself; but, it can also be the secondary cause from many vectors. Diseases like that are (essentially) impossible to explain to the public. Also, the biological principle function is a standard trope for "good for X in the short run, bad for the person in the long run".
I have direct experience of research into Alzheimer's disease, including specifically surrounding the amyloid cascade hypothesis.
The strength of the amyloid hypothesis is that it is currently the only way to unify early onset AD that is caused by mutations in APP and presenilin with the pathology of both early and late set AD.
The weakness is that experimental mice expressing mutated APP do not get neurodegeneration, despite showing amyloid accumulation and behavioral defects.
Mice expressing mutated presenilin in contrast do get both behavioral defects and neurodegeneration, despite showing no accumulation of amyloid.
"Perhaps mice are different" is the usual response/excuse.
This defense is considerably weaker now, given the very modest benefits of removing amyloid from the human brain, as shown in recent clinical trials.
So.... when considered rigorously, the amyloid hypothesis remains to be proven.
However, it will always have its supporters until there is an alternate explation for the convergence of mutations in APP and presenilin on precisely that region of APP that generates amyloid.
Those are amyloid-only mice. It's an amyloid+tau disease, with tau the proximate cause of neurodegeneration. Normal mice don't get tau pathology, whereas even healthy human beings do, however it stays localized until the presence of widespread amyloid pathology.
Causal intervention on amyloid+tau mice is consistent with causal mediation from longitudinal human neuroimaging data: the amyloid pathology greatly accelerates tau pathology, and then this causes neurodegeneration.
I appreciate your comment. However, a straight-forward test of the hypothesis that amyloid is the toxic (etiological) agent in AD fails when tested in mice.
One may then ask, what is being remedied in the many, many, studies that claim to successfully target amyloid toxicity in mice? And is this relevant to the processes that occur in AD?
Human pathology studies are limited in ability to determine causal agents because they are primarily observational, i.e. they find correlations, show that changes in certain other proteins or processes are associated, such as tau that you mention, inflammation, etc. Or as you mention, show that the pathological hallmarks of AD have a stereotypical order of appearance.
However, the only human studies that can demonstrate cause in AD are genetic studies.
One may then ask, what is being remedied in the many, many, studies that claim to successfully target amyloid toxicity in mice? And is this relevant to the processes that occur in AD?
I don't think studies rescuing cognitive deficits in amyloid-only mice are convincing evidence for the amyloid hypothesis, precisely because we know amyloid is not the proximate cause of neurodegeneration in actual Alzheimer's disease, and that proximate cause does not exist in those mice.
In other words, these mice are not faithful recapitulations of the full disease. They have their amyloid production turned up so far that their amyloid pathology seems to cause cognitive deficits, but that's not what's happening in humans. They are, at best, a good vehicle for testing specific narrow hypotheses about amyloid production and clearance. The field has largely moved on from amyloid-only mice as a direct predictor of clinical efficacy, and that was the right call.
Human pathology studies are limited in ability to determine causal agents because they are primarily observational, i.e. they find correlations, show that changes in certain other proteins or processes are associated, such as tau that you mention, inflammation, etc. Or as you mention, show that the pathological hallmarks of AD have a stereotypical order of appearance.
Here is some data in living humans, besides genetics, that has relevance to causation, in my opinion:
- The location and severity of amyloid pathology is a poor spatiotemporal match to the sites of neuronal volume loss, and to the severity and nature of clinical deficits. However, the location and severity of tau pathology is a very good match to both of these things. Of course, since these observations are correlational in nature, they don't absolutely prove a specific causal theory. But they do rule out, for example, the idea that amyloid is proximately connected (by which I mean nearby somewhere in the causal graph) to the process of neurodegeneration, whereas tau seems to be very proximately connected. From this observation alone tau could be downstream or sidestream rather than upstream, but it does then suggest that whatever causes tau pathology is itself upstream of neurodegeneration, since correlations always have a cause (the correct statement that "correlation ≠ causation" simply means "correlation between A and B does not imply that A causes B", but the explanation must be either A causes B, B causes A, or C causes both A and B).
- Anti-amyloid antibodies which remove plaque in humans cause downstream reductions in tau pathology in humans, and, separately, have clinical benefits in those humans.
- The spatiotemporal progression of amyloid and tau pathology is highly consistent with the hypothesis that amyloid pathology greatly worsens the tau pathology, but not vice versa. And there's not an alternative explanation I've come across for this fact than that amyloid pathology worsens tau pathology.
All of the above facts are generally true in combined amyloid+tau mouse models as well as in vitro human cell studies, which is some reason to believe these are closer to faithfully recapitulating the disease than the amyloid-only models. Once we believe that, we can then do more causal interventions on those models which we couldn't do in humans, and learn more about causality. For example, we know that intentionally worsening amyloid pathology in amyloid+tau mouse models also causes tau pathology and neurodegeneration to worsen in mouse models. And because these models look closer to the full disease than the amyloid-only models, this is at least relevant causal evidence, though we always have to be open to the possibility that the disease models are still missing some important elements.
I'm not aware of an alternative hypothesis to the (ATN) amyloid → tau → neurodegeneration model which synthesizes all of the above facts, along with the genetic evidence for amyloid's causal role which you referred to. By contrast, I'm not aware of any evidence inconsistent with the ATN model.
Yes, there is a clear sequence of how AD pathology develops, starting with amyloidopathy and progressing to tauopathy, but 1) there is as yet no established molecular connection between the two, and 2) one should not conflate pathology with disease mechanisms.
So, taking the amyloid hypothesis itself (putting presenilin aside for the time being).
We know that mutations in APP do cause AD. How? And if amyloid is not the "proximate" cause of AD, how do mutations in APP cause AD? Include in this Down syndrome, where >90% of cases develop early onset AD by age 50. They have an extra copy of APP that is not mutated.
Furthermore, people can accumulate large amounts of amyloid in the brain without having any notable dementia.
Adding tau to the equation does not help much in explaining how APP mutations cause AD. All people have tau. Furthermore, mutations in tau do not cause AD, they cause different neurodegenerative diseases (e.g. frontotemporal dementia).
Combining APP mutations with presenilin mutation and/or tau mutations in mice does lead to worse outcomes, but the same could be said for combining any other random set of neurodegeneration-associated gene mutations.
Yes, there is a clear sequence of how AD pathology develops, starting with amyloidopathy and progressing to tauopathy, but 1) there is as yet no established molecular connection between the two, and 2) one should not conflate pathology with disease mechanisms.
I agree that the specific molecular mechanism(s) is/are currently unknown. I've seen a number of proposals, but to my knowledge there isn't smoking-gun evidence for any one of them. But there can be causal evidence that A causes B (such as which I list) which exceeds a mere sequence of "A first, then B", and without knowing the specific mechanisms by which A causes B.
We know that mutations in APP do cause AD. How? And if amyloid is not the "proximate" cause of AD, how do mutations in APP cause AD? Include in this Down syndrome, where >90% of cases develop early onset AD by age 50. They have an extra copy of APP that is not mutated.
A bit confused by these questions, and I suspect the confusion may have to do with the term "proximate". By "amyloid is not the proximate cause of neurodegeneration", I simply mean it is upstream, mediated by another cause (namely tau). I think that clarification answers these questions.
Furthermore, people can accumulate large amounts of amyloid in the brain without having any notable dementia.
As predicted by the ATN model, at least for some time. But there is a threshold of amyloid pathology that does seem to guarantee progression to tau pathology and dementia.
Adding tau to the equation does not help much in explaining how APP mutations cause AD. All people have tau. Furthermore, mutations in tau do not cause AD, they cause different neurodegenerative diseases (e.g. frontotemporal dementia).
Sure, there are different tauopathies, each with a characteristic fold. All people have tau, but there's a specific AD tau fold emerging apparently from the locus coeruleus, then spreading to the hippocampus and entorhinal cortex, and it's this that seems heavily accelerated by the presence of amyloid pathology in humans. (By the way, a notable fact is that autosomal-dominant AD -- clearly caused by APP/PSEN1/PSEN2 mutations affecting amyloid production -- has the same tau fold as sporadic AD, even though the large majority of other tauopathies do not.)
Combining APP mutations with presenilin mutation and/or tau mutations in mice does lead to worse outcomes, but the same could be said for combining any other random set of neurodegeneration-associated gene mutations.
Note I didn't just say it "leads to worse outcomes". It's specifically that amyloid pathology worsens tau pathology, and then neurodegeneration occurs colocated with the tau pathology. This cannot be said for other random sets of mutations, in general.
(By the way, basically all of these points are discussed in the article I wrote which got linked above. You're under no obligation to read it but it might save us some time.)
These points still don't explain how mutations in APP cause AD.
Note that not all AD-causing mutations in APP also cause amyloid accumulation, for example APP-Osaka (loss of APP residue E693) results in familial AD without any accumulation of amyloid [0]. (One can ignore claims that this mutation increases Abeta oligomers, since the evidence is that Abeta oligomers are found at far too low concentrations in the human brain. They would have to be more toxic than ricin if they were etiological for AD). The oligomers seen on gels are an artefact, see the controversy surrounding Tessier-Lavigne).
As you state, and I agree, APP is upstream of tau in natural AD pathogenesis, but does not cause neurodegeneration in mice. So we still don't know from direct experimentation how APP leads to tauopathy and neuodegeneration. The evidence that this is through Abeta per se is tentative at best.
[0] A Second Pedigree with Amyloid-less Familial Alzheimer’s
Disease Harboring an Identical Mutation in the Amyloid
Precursor Protein Gene (E693delta) https://pubmed.ncbi.nlm.nih.gov/25743013/
Note that not all AD-causing mutations in APP also cause amyloid accumulation, for example APP-Osaka (loss of APP residue E693) results in familial AD without any accumulation of amyloid [0].
This is interestingly similar to the Arctic Mutation, and in the same codon no less: no plaque, but still autosomal-dominant AD due to an APP mutation. I had previously taken the Arctic Mutation to be evidence that it's not plaque per se, put more likely protofibrils (which are components of plaques in normal AD, and still present under the Arctic Mutation) or precursor aggregates which are pathogenic. The fact that the Osaka Mutation blocks protofibril formation underlines the uncertainty, that you and I agree exists, on the detailed molecular mechanisms. I would be inclined to point then to oligomers, but you say the oligomers are found at far too low concentrations to be relevant — what's your source for this?
As you state, and I agree, APP is upstream of tau in natural AD pathogenesis, but does not cause neurodegeneration in mice. So we still don't know from direct experimentation how APP leads to tauopathy and neuodegeneration. The evidence that this is through Abeta per se is tentative at best.
Not only APP, but also PS1+PS2 mutations of course, can cause ADAD, and the relevant mutations all seem to cause more Abeta42 production. In the sporadic case, production usually seems unchanged, but clearance is usually impaired (especially with ApoE4). What they all seem to have in common is amyloid production or clearance. I'm curious if you know of another pathway they have in common besides this. Otherwise it's hard to see what the alternative hypothesis is, which could explain the etiology of seemingly highly-similar disease trajectories (ADAD + sporadic AD).
I’ll add as an addendum: APP mutations do cause neurodegeneration in mice, if those mice are combined amyloid+tau models. This seems most faithful to the human disease.
As you have demonstrated an interest in this topic, but are not an active researcher, I suggest that you become familiar with Alzforum [www.alzforum.org]. It provides reputable summaries and comments from leading researchers on topical issues and papers in Alzheimer's and related neurodegenerative diseases.
I actually read your article some time ago, having found it via SSC. One comment I would make is that the pathology and genetics of AD implicate APP without much doubt. In the field, there isn't much argument against it, even by those who work on tau. Divergences in opinion more often focus on which is the better target for drug therapy.
For example, Derek Lowe's articles on AD do not argue against an APP-tau-neurodegeneration pathway (I am a regular reader of his blog), but argue that targeting amyloid is unlikely to work as a drug therapy. This argument is sometimes extended by Lowe and others to questioning whether Aβ per se is the agent that triggers the downstream effects of either mutating APP or otherwise disrupting its biology. It is possible that Aβ is a by-product of the "true" APP-triggered pathway, see the arguments in favor of C-terminal fragments I linked to earlier.
A fundamental problem the field faces is that it is difficult to experimentally induce actual AD in animals, and so it is difficult to scientifically test hypotheses as to what might cause it. Compare perhaps to something like diabetes. Exactly why mice do not develop realistic AD is unknown.
In addition, there have been literally hundreds of publications (minimum 300) claiming to reverse "Alzheimer's disease" in mice, and the only ones that have lead to any kind of meager success are those using the antibodies. The arguments made in the 2010 review below by Zahs and Ashe [0] remain mostly true sixteen years later.
The recent clinical trials, e.g. Lecanemab, are of interest not only because they might provide at least some benefit, but also because they might represent the first actual scientific experiments showing that removing Aβ can ameliorate the disease, even if only slightly because of the late stage at which they have so far been tested.
Currently ongoing trials using anti-Aβ antibodies and targeting earlier phases of the disease in those with inherited mutations are of keen interest. If these demonstrate clear prevention of disease progression, then the Aβ hypothesis can be considered tested and proven. If not, then the Aβ hypothesis should be abandoned. This would however still not disprove an APP-focused hypothesis.
No APP-alone mouse gets neurodegeneration resembling that seen in the human AD brain, i.e. with so-called "neuritic" plaques, tauopathy, spongiopathy and widespread neuronal death.
This is why researchers now most often use the 5XFAD mouse, which has APP with three mutations, and presenilin with two mutations (hence 5 FAD mutations) [0]. Note however that mutated presenilin alone is enough to cause neurodegeneration in mice, such mice however do not accumulate amyloid, which is why mutated APP in added to make the pathology more "realistic".
As to Aβ42, there are mutations in APP which cause familial AD, but produce exclusively Aβ40, e.g. APP A673V [1]. Note also that most studies report alterations in the ratio of Aβ42 to Aβ40, precisely because effects on the levels of Aβ production are inconsistent across APP mutations.
Nevertheless..... and despite my obvious skepticism towards the amyloid toxicity hypothesis, the mutations in APP that cause AD all cluster in or near the region of the protein that is Aβ. There must be a reason for that. It is also in contrast to presenilin, where the mutations are distributed throughout the molecule, indication a loss of presenilin function causes AD.
One alternate explanation to Aβ or oligomer toxicity is proposed toxicity of the immediate precursor to Aβ, i.e. the APP C-terminal fragment (CTF), see for example the recent paper below and references therein [2].
At this point you should just go to grad school in a molecular neurology program.
You clearly have the passion for this sort of research, and it would probably be useful to immerse yourself in the process to get more experience and judgement.
Grad programs in bio usually have journal clubs; bring one of your papers, and see what people think of it.
When the first drugs targeting HIV arrived the results were undeniable. Yes the drugs sucked for various reasons and yes HIV would evolve resistance. But the data demonstrated a very clear link that these drugs suppressed HIV and suppressing HIV made people live longer. Or consider mRNA and COVID, a great success story where the technology was put to good use and the results are obvious.
On the flip side we have certain cancers like certain breast cancers, melanoma, etc that never had a "wow" moment where some miracle turned them from highly fatal into treatable but we have seen decade after decade treatments improve and survival rates march ever upward such that what were once almost guaranteed death sentences are now often very treatable.
These are two disease treatment models worth keeping in mind. Sometimes major leaps are made. Sometimes progress is slow.
Now if we consider amyloid beta therapies: we have treatments that target amyloid beta with varying degrees of success but at least some show definite reductions in amyloid beta plaques. To the best of my knowledge that has not shown to improve outcomes in Alzheimer's patients to any meaningful degree.
That concerns me and I think justifies some skepticism of the amyloid hypothesis. The data is messy but if amyloid beta were a symptom not a cause that could certainly fit the results we are seeing. That doesn't mean the amyloid beta hypothesis is wrong but I think skepticism of the "state of the art" in the field is warranted given the pathetically ineffective progress made to date.
Now if we consider amyloid beta therapies: we have treatments that target amyloid beta with varying degrees of success but at least some show definite reductions in amyloid beta plaques. To the best of my knowledge that has not shown to improve outcomes in Alzheimer's patients to any meaningful degree.
This is false. They slow down disease progression by about 30%, as measured by cognitive outcomes. This is discussed in the article.
I work in drug discovery. For the past twenty years or so, my personal analogy for this hypothesis has been a fantasy story around the days after the bombing of Dresden, when a new civilization suddenly visits Dresden and has no priors about what may have happened there. The aliens see bricks all over the place and assume that the bricks were the cause of the catastrophe. They take great efforts to pick up the bricks and save a couple of lives from the people who were covered in the ruble. The aliens build better systems to pick up bricks in the future and get ready to act next time. When a nearby city gets bombed, they quickly visit and help recover bricks saving a couple more lives. A different civilization could have instead focused on reducing the bombs or detecting and defending against the attacking airplanes.
Our immune systems are complicated, much more so than airplanes and bombs. The amyloid deposits are very likely part of an immune response, and although in principle immune responses going wild are horrible and can be fixed, it is very important to work on identifying and addressing the causal factors of this disease. There have been more therapies tested on the amyloid hypothesis that mere statistical fluctuations could explain away. I don't always agree with Derek, but I'm with him on this one. New ideas are urgently needed here, or this horrible disease will be an increasingly common end state for our aging populations.
The hypothesis that amyloid is simply a downstream effect, not a cause, is of course worth considering, and where my mind was at when I first approached the literature skeptically. The widespread presence of amyloid-beta plaques in Alzheimer's disease has been known since 1906, but the field didn't adopt the amyloid hypothesis for decades precisely because of this possibility. But then in the early 1990's, strong genetic causal evidence emerged, which is why the amyloid hypothesis emerged at that time. Other important causal evidence has emerged since, especially that tau pathology (the proximate cause of neurodegeneration) is causally downstream of amyloid pathology, which we know from many lines of evidence now. (See the article for a lot more detail on all this, if you are curious.)
As for the possibility that the successes of amyloid therapies might be explicable by chance, this is highly implausible. Only three (aducanumab, lecanemab, and donanemab) of a dozen or more amyloid therapies successfully cleared plaque, and it is precisely those three that achieved a slowdown of cognitive decline in phase 3 trials (with aducanumab succeeding in only one of its two, but with the others succeeding in their only phase 3), several of which with p-values below 0.001. This is not p-hacking or reporting bias.
Regarding lack of p-hacking. The placebo arm of blinded trials breaks when your brain can detect a medication. The effect is tiny in these studies; approval was rushed to give hope to patients. The drug was discontinued later.
You're thinking of aducanumab. Lecanemab and donanemab have been in widespread use for several years now, and open-label extensions vs. external controls showing increasing benefits over longer treatment durations.
Cease, yes, if the cause is removed early enough. But if you intervene too late (once symptoms are detectable), then the downstream tau pathology, which is what directly kills neurons, likely spreads on its own via a prion-like mechanism.
So far, no clinical trial has completed prior to clinical onset for an antibody which actually removes plaque. This is probably the main reason only 30% slowdown has been achieved so far. The donanemab prevention trial is due to complete next year. That will be an important one to watch.
Yes, and the fact that something like oral hygiene can influence AD would support your thesis. Often an infection in the gums/teeth can result in boils or worse in distant parts of the body. Add a dysfunctional blood brain barrier and you are screwed.
Amyloid plaques could be part of a disease cascade. Not the root cause, in other words, but an additional downstream problem than probably still needs to be addressed.
> This is false. They slow down disease progression by about 30%, as measured by cognitive outcomes. This is discussed in the article.
Perhaps I am just not well-informed but 30% slowdown in progression translates to sufferers have some mild improvement in cognitive tests and live a few months longer.
Maybe it is simply too early to tell but I would naively expect something much more significant. Perhaps this is the sort of thing that requires much earlier treatment to demonstrate better results.
I'm not saying amyloid beta research should be terminated. Merely that everyone in the field should be willing to entertain other ideas.
Perhaps I am just not well-informed but 30% slowdown in progression translates to sufferers have some mild improvement in cognitive tests and live a few months longer.
A few years longer. It's obviously frustratingly far from where we'd like to be. It was only a direct response to the claim that Alzheimer treatment isn't even in the "sometimes progress is slow" category, but rather in the "no meaningful benefit at all" category.
I'm not saying amyloid beta research should be terminated. Merely that everyone in the field should be willing to entertain other ideas.
We agree that people should be willing to entertain other ideas! I don't think anyone is saying otherwise.
But without any training in experimental methodologies, molecular biology, protein mechanics, pharmacology, or any of the other specialized fields that make up the world of Alzheimer's research, how do you view yourself as qualified to make that conclusion? What body of knowledge are you drawing on to conclude that the experts' reasoning is sound, they are properly controlling their experiments, they are drawing the correct conclusions based on the underlying mechanics? AFAIK even people who do meta-analyses are qualified in the field they are doing the analysis for.
This is a reasonable challenge. I won't and shouldn't be able to convince you that my process was satisfactory, but I'll describe it anyways.
I agree that I don't have the qualifications to check whether, for example, a particular cryo-EM study was conducted properly. But I can check whether those who do have such qualifications disagree on the methodology or findings of that particular study. There's a lively debate within the Alzheimer's research community; it's not hard to find dissenting opinions on just about anything, and I actively seek them out, and when such disagreement exists, I avoid weighting any evidence too heavily, unless the disagreement is about broader matters of synthesis or specific statistical or methodological questions in which my non-biological scientific background permits me to reach my own conclusions.
I am also careful not to heavily weight a single assumption-laden preclinical study conducted by a single lab, for example, but instead to look for "smoking gun"-style evidence, in those few cases that it exists, or to look at the bulk of evidence across many studies from many labs, where the specific conclusions do not seem to be seriously in doubt by experts. In general, I've been skeptical, considering alternative explanations wherever it seemed crucial to the bigger picture, and avoiding trusting anything that seemed like it involved knowledge which was heavily in the weeds on stuff that I couldn't understand. I had a personal motivation to understand the genuine truth here, and enough scientific background that I usually know when I'm out of my depth on a specific matter.
I think it's reasonable of you to say: that all sounds well and good, but I just don't know your process well enough to trust it, and you don't have formal qualifications on the matter, so I'll ignore what you say. I certainly wouldn't expect you to take anything on my authority. I see my article essentially as an act of science journalism, and scientific journalists often lack formal training in the field they report on. You can read it and see if the reasoning makes sense and the evidence is convincing, or you can reasonably ignore it and fall back on expert opinions.
I did the investigation precisely because the majority expert viewpoint was being called into question by a lot of non-experts, and I had a personal motivation to find out, genuinely, whether this critique was warranted. If you don't have that motivation, then it's probably not worth your time to do the same. I did, and I came away satisfied.
Is putting your thumb on the scale against Lowe. When a few replies down from here some commenters have provided an article demonstrating the exact fraudulent science in favor of what Lowe is saying.[0] It seems you may very well be disrupting it because he has a minority opinion. So you’ve possibly spent 6 months understanding an incorrect and fraud supported thesis. That seems like an outsider trying to disrupt it by using their “Google/SpaceX” creds to claim authority on the work of insiders.
1. I don't say Derek Lowe is wrong because he's in the minority. Minorities are sometimes right. But since the parent comment was arguing on authority and my lack thereof, I point out only that one shouldn't cherry-pick one's choice of authorities. Either accept the majority opinion of the experts, or come to your own opinion based on the quality of the arguments and evidence.
2. I would never want anyone to believe what I say because of "Google/SpaceX creds" (I didn't even write that line, Scott added it, and only to provide a brief biography and acknowledge that I do not work in the field, not to lend an air of authority to my words).
3. There's no need to cite the fraud to me, since I already discuss it in my article. You are welcome to read that article and form your own opinion about the arguments therein.
This really has become the new physics now, right where they think they can invade any given field in six months because that’s how long it’s supposed to take physicists to learn AI
In some ways physics is different from biology and medicine, I do think outsiders to physics can pick up and contribute a bit more easily (although it depends on field). Biomed just has an absolute insane amount of ambiguous knowledge that mostly gets picked up through diffusion across decades of learning. And many of the results in the literature are just wrong (one of the reasons I stopped being a researcher was seeing just how bad the publication record in biology is).
BTW, many physics people pick up the mechanical bits of machine learning/AI very quickly since they have all the foundational mathematics. The harder parts are understanding all the methods/tricks/complexity that got us to the state of the art- similar to biomed, you just sort of have to immerse yourself amongst knowledgeable people and let their knowledge diffuse in.
I don't know anything about alzheimers but I'd hope someone that does would engage with David SJ's points instead of dismissing him on lack of authority alone.
Oh FFS what does any knowledge matter any more if we can now say that someone who has no medical training has the ability to turn the entire field of Alzheimers on its head?
Makes me wonder if this is the typical correlation != causation argument where amyloids are produced and are simply a marker/symptom rather than the cause.
Most of my career success has been based on my tendency to be relentlessly proactive and it does not surprise me in the least that frontier models would start to pick up on these strategies (I'm pretty sure each of the individual things you list above are available in the codeoverflow parts of the training corpus, and combining them to achieve a goal seems ... like a fairly obvious result of the type of training these models go through.
About a year ago I remarked to people that despite all my attempts to make data more programmatically accessible, the most effective way for AI to interact with a modern computer is to use the built-in accessibility interfaces driving actual desktops with full applications. IE, the best API for an AI is the UI (mainly because that's what most humans use).
Typically what happens is that the new treatments with bad side effects are given to the sickest patients (who have exhausted all other mechanisms), rather than to the bravest souls with less dire current circumstances.
This makes some sense in terms of compassion and matching new experimental techniques with patients with no hope, but it skews the results highly negative because the patients are already very close to death's door. It does not provide an accurate signal for what the results would be if we gave them to less sick people.
I don't think any of this can be changed without large-scale social acceptance of greater risk in clinical trials and significant support from the government.
> It does not provide an accurate signal for what the results would be if we gave them to less sick people.
It provides an excellent signal because we want to prove that these drugs are doing something that the standard of care is unable to.
There's this sense that medicine is easy and some evil cabal are limiting health to their cronies. Most medications never get to trial for their intended indications, and most fail trials. There's no reason to believe oncology medications are somehow uniquely unlikely to go through this well-described process of failure.
I think you misunderstood my point. Results in extremely ill patients don't correlate well with results in mildly ill patients (in cancer), so the signal you get from extremely ill patients is not predictive- it may be a drug fails in extremely ill patients, but would have worked great on the mildly ill patients- basically a false negative.
> I don't think any of this can be changed without large-scale social acceptance of greater risk in clinical trials and significant support from the government.
I agree with you about significant financial support from the government if that support is financial and given to smaller groups.
I disagree about societal acceptance though. I feel like your point of view may be missing clinical trials and treatments in the USA. The laws we have are written in blood, and the laws more sophisticated than I think you appreciate.
There are various options for fast-tracking drug development if the conditions are serious enough or the drugs promising enough. I suggest actually reading about how these processes work, and the history of what happens when we don't have these protections, before deciding that approvals are overly cautious.
You are explaining clinical trials to somebody who works in pharma (speccifically a company with many cancer drugs that already got approved). I'm pretty familiar with the area already- for example, I know the drug thalidomide, which was part of the origin story of the FDA, is still prescribed widely as an anti-cancer drug.
Seeing and hearing Karateka for the first time on an Apple II+ was a life-changing experience for me. Along with Ultima III, it made me want to be a game developer (I was in 4th grade, so around 12, at the time). Everything about the game is just so smooth and well-done- it has a plot, a progression, good animation and realistic sounds. I was pretty unhappy for years around the fact that I didn't understand the technology (machine language instead of BASIC, Apple's very funny graphics implementation, doing sound and animation simultaneously) to make games like that.
If anybody wants to see how that entire creative process went, there's a "game" called "The Making of Karateka" on Steam that is a nice interactive experience telling the story of Jordan Mechner's start in the games industry and how Karateka came to be.
It's a fun media experience with a lot of playable prototypes.
I think if Damore had limited his criticism to the corporate DEI (which I found to be a bit heavy-handed and prone to critical theory groupthink) in a short doc, he probably would have been a bit more successful (in helping Google rein back its DEI training).
One rarely stated thing I learned over time working there is that managers read eng-misc and will prevent you from transferring to their team if they didn't like what you said, or how you said it, or who you said it to.
I once interviewed for a company and the interviewer was telling me how he (a vc) funded a project to generate large streams of random numbers; you would select an index at random, share that private key with somebody, and then the subsequent text could be used as a one-time-pad. NSA would be forced to buffer/save the entire stream, which could be generated at GB/sec, if they wanted to decrypt.
I wonder if we could mess with NSA-style surveillance by having a good chunk of the population streaming lots of random data over the internet. Essentially, Alice piping her /dev/random to Bob's /dev/null over netcat or something. Make a slick looking app that does it 24/7 in the background using excess bandwidth and tell people it sticks it to the NSA.
Spy agencies would not only have to store it all in case it was something valuable, but at some point they may try to crack it because it's indistinguishable from encrypted data and waste resources on it. If enough people did it, total web surveillance could become impractical.
That's known as cover traffic and is a tactic employed by many of the anonymity oriented overlay networks.
I'll note that any observer already has this problem to the extent that video streams are also encrypted. However most observers presumably recognize the endpoints as well as being able to classify the traffic by means of statistical analysis.
What might be useful would be a tool to generate arbitrary user data of various forms, including HMTL, video, audio, and various message formats. Then it could assemble a convincing traffic stream full of gibberish to exchange with peers at random. You wouldn't even necessarily need all that much of it to overwhelm any would be observers when considered relative to the volume of streaming service traffic that already exists.
> generate large streams of random numbers; you would select an index at random, share that private key with somebody, and then the subsequent text could be used as a one-time-pad.
Yep- this is what I do. I use a high quality VLM to generate labelled boxes (in my case, around tardigrades in a microscope image), do some light editing to fix the small number of errors, and then train YOLO26 with it. Works great, saved me tens of hours of labelling. It's a bit scary that there is a VLM that works as well as my fine-tuned model (although much slower).
thats a fantastic strategy thank you, and thanks to all the other helpful posters as well here. do you have any tips for how to choose the base yolo model? or just any generic one will do?
"I am David Schneider-Joseph, an engineer formerly with SpaceX and Google, now working in AI safety. Alzheimer’s isn’t my field, but I got very interested in it, spent six months studying the literature, and came away believing the amyloid hypothesis was basically completely solid. I thought I’d share that understanding with current skeptics."
6 months of reading literature when you don't know how to read biomedical literature isn't very confidence inducing. I know this site really likes it when smart outsiders come in and disrupt the status quo, but... probably not in this case.
reply