One quick piece of semantic and linguistic housekeeping for the commenters…
Heritable != Molecular / Genetic Mechanism
There is a conflation of these terms in popular discourse that does a disservice to the field of statistical genetics, imo. There are mechanisms of inheritance that operate various length / time scales other than that of biological macromolecules. For example, if you tell me what language your parents natively speak I can tell you your primary language with >90% accuracy.
So before we start getting 3 replies deep into any thead, please remember that retrospective observational data measured with unqualified instruments is notoriously confounded and that we can barely infer causal structure in controlled functional genomics experiments (much less a GWAS of phewas). So let’s all please keep an open mind and not be so certain about our beliefs.
This comment reads as if it were dropped into a generic "genetics of lifespan" thread,. The Dynomight article is already making a much more sophisticated version of some of these same points. The article's central argument is precisely that heritability is a contingent observational statistic, not a Platonic form. This particular article isn't conflating heritability with genetic mechanism at all. It's interrogating a simulation model and its assumptions. The warning about "unqualified instruments" and "retrospective observational data" feels off as this paper isn't a straightforward observational study. it's a parametric simulation fitted to twin registry data.
This comment might be very useful in a Reddit thread full of people saying "50% of lifespan is in your DNA," but it's a bit off-target as a response to this particular article.
Agreed, the Dynomight article is on-the-mark. I work in this field and was really puzzled when I read this paper. Yes, it is obvious that excluding extrinsic causes of death will increase heritability estimates. But is death from influenza genetic or extrinsic?
The typo on the first page of the Science article is on the authors, not the editors.
The accurate version of the result would be something like: “if you model lifespan as aging + i.i.d. noise and dial the noise to zero, heritability of the aging component is ~40-50% in our model.” Which is barely a finding, since by construction reducing i.i.d. noise has to increase heritability of whatever non-noise remains.
This would require an accurate definition of ageing. What is ageing? How is it related to life span? Because in theory, there can be definitions of ageing that are not tied to life span. For instance, do bacteria age? Does this affect life span? What is the life span of a bacterium anyway? Does hydra age? (For those who don't know much about biology: literally everything ages, if you define ageing as functional decline over time. Even viruses would age, if you define it as functional infections possible plotted over time. Does DNA and RNA age? The definitions become blurry; almost no molecule is immune to changes and modifications, so just about anything would age. So it really depends on the definition, and we need to read the definition before we can accept assumptions based on it. Thus: what is ageing and how does it relate to lifespan, as definition?)
> For example, if you tell me what language your parents natively speak I can tell you your primary language with >90% accuracy.
According to the link above, the heritability of the primary language is zero, whereas the heritability of what language(s) a person speaks in general (whether primary or secondary) is not necessarily zero and varies by language.
I believe that your example of "what language your parents natively speak" is incorrect.
Some ways of measuring heritability would have trouble detecting this as environmental, but that is considered a deficiency in those measures, not part of the definition of heritability. Any serious study into heritability of language would quickly find it is largely due to the common environment.
We have many bits and pieces of causal structure for some human traits courtesy of GWAS and PheWAS but you are right that lifespan genetics of humans is seriously compromised by rapid changes in life styles and environments.
Heritability has a very specific meaning in quantitative genetics [1], which in many ways is not what your intuition would suggest [2]. It is this usage that the article talks about that.
That said, there are plenty of critiques of this definition of heritability, and not just because it is different from what a layperson would expect it to mean.
For example, the way it is used also usually has a big problem in that the standard formula assumes that Cov(G, E) = 0 (or at least is negligible), whereas in practice that is not actually true [3, 4].
This definition of heritability is also mathematically flawed in that it assumes (without evidence) that P = G + E, or at least can be reasonably approximated this way. Given that human development is the result of a feedback loop involving genetic and environmental factors, one would expect a model closer to something like a Markov chain. Proposed justifications of a simple additive model as an approximation (e.g. via the central limit theorem for highly polygenic traits) have to my knowledge never been tested.
More recent genome-wide association studies [5] have actually shown a considerable gap between heritability estimates from genotype data and heritability estimates from twin studies, known as the "missing heritability problem".
OP has another post on the definition of heritability, which I really liked: https://dynomight.net/heritable/ . I'm a layman, though, so since you seem knowledgeable, I would love to hear your thoughts on that article!
For instance, OP's definition H = Var[G] / Var[P] seems to bypass the issues you mentioned:
> For example, the way it is used also usually has a big problem in that the standard formula assumes that Cov(G, E) = 0 (or at least is negligible), whereas in practice that is not actually true [3, 4].
> This definition of heritability is also mathematically flawed in that it assumes (without evidence) that P = G + E, or at least can be reasonably approximated this way.
> For instance, OP's definition H = Var[G] / Var[P] seems to bypass the issues you mentioned:
No, this is exactly the definition I am talking about. The problem is that while theoretically you could work with Var(G)/Var(P) even if Cov(G, E) ≠ 0, studies are not designed to capture that.
In fact, the standard ACE model [1] used in twin studies explicitly assumes among other things that there is no gene-environment correlation. This means that it gets silently added to one or more of the ACE components; not because of any ill intentions, but simply because if you included covariance, the resulting system of equations would be underdetermined and could not be solved [2].
But to make matters worse, gene-environment correlation/interaction itself is disproportionately absorbed by the A and C components rather than E. All this can lead to inflated heritability estimates.
And to clarify, I am not making any pronunciations about how much relevance or magnitude that effect has; for all I know, this could in the end be a minor effect. My point here is that there is a lot of mathematical handwaving going on with very limited testability of the modeling.
[2] If you want to be precise, you need to actually distinguish between gene-environment correlation and interaction and use P = G + E + (G x E), but that makes the system even more underdetermined, because now we have both Cov(G, E) and Var(G x E) to worry about.
> Heritability has a very specific meaning in quantitative genetics [1]
Literally the first paragraph of that page is
> Heritability is a statistic used in the fields of breeding and genetics that estimates the degree of variation in a phenotypic trait in a population that is due to genetic variation between individuals in that population. The concept of heritability can be expressed in the form of the following question: "What is the proportion of the variation in a given trait within a population that is not explained by the environment or random chance?"
That matches what I assumed it meant, and it seems like OP and the post are arguing that that is some kind of surprising interpretation.
> OK, but check this out: Say I redefine “hair color” to mean “hair color except ignoring epigenetic and embryonic stuff and pretending that no one ever goes gray or dyes their hair et cetera”. Now, hair color is 100% heritable. Amazing, right?
Uhm, no. That is exactly what I (and I think most people) would expect the answer to be.
> That matches what I assumed it meant, and it seems like OP and the post are arguing that that is some kind of surprising interpretation.
The unintuitive part is that in quantitative genetics, heritability is defined in terms of variance in traits at the population level, not as the passing of traits from parents to offspring (that would be heredity [1]). Of course, I may have misinterpreted what you said in your OP when you cited the wiktionary definition of "[g]enetically transmissible from parent to offspring", and if so, I apologize, but at the time it seemed to me that you were talking about heredity.
> Uhm, no. That is exactly what I (and I think most people) would expect the answer to be.
What the article is talking about is that if you fix Var(E) = 0, then Var(P) = Var(G) in the standard heritability model, i.e. all phenotypic variance is explained entirely by genotypic variance (because in that model, Var(P) = Var(G) + Var(E)).
Fun fact (even if only tangentially unrelated): In Western countries, wearing glasses is a highly heritable trait, because wearing glasses is a strong proxy variable for refractive error [2], such as nearsightedness, which is highly heritable. It is often brought up as another example of how the quantitative genetics definition does not match conventional use of the word.
The heritability statistic that occurs in the literature is the ratio of genetic variance to phenotypic variance.
Two corrollaries:
* When discussing heritability results from the literature, we are discussing that statistic, not your intuitive understanding of what the word should mean.
* In the scientific literature, your conception of heritability doesn't operate. In the scientific sense, the number of hands you have has low heritability, despite being genetically determined.
I think you're going to find "let's check Wiktionary" is not the decisive move in these kinds of discussions that it is elsewhere.
Another great example of the unintuitiveness of heritability is the fact that earrings are highly heritable. Earrings are highly correlated to a specific genetics (being female), so they're very "heritable", even though that correlation is an arbitrary cultural fashion.
See my sibling comment. This is misleading for the same reason, but in this case the cause of misleading is narrowing the timespan under consideration to approximately now.
> In the scientific sense, the number of hands you have has low heritability, despite being genetically determined.
This is only a surprise because unlike layman the author of this joke insists on considering heritability among humans specifically. While "heritability among humans" sounds like a reasonable comment to a layman, the author of this joke is misleading the layman, because layman (before being mislead) correctly thinks of "heritability" as "heritability among all living things with genes".
There is a genetic component to alcohol use disorder, for example. But if one is in an environment where there is no access to alcohol whatsoever, then that person, despite their genes, will not develop an alcohol use disorder. The disorder can still be passed from parent to child, but it's more complicated than just genes.
I heared the same distinction as OP, but it is the other way around, it's the degree to what a trait is inherited from you parents which cannot be explained by the enviroment or Random Chance.
What you're expecting heritability to mean is essentially "are genes responsible for expressing this trait", which is very different from "can I get this trait from my parents?" which does not impose any particular method for passing on the trait.
If the study doesn't use sequenced genes of parents and children as input into the model, it can't make the distinction between genetic or non genetic influence by parents.
That is exactly wrong. The measure of heritability used in the scientific literature is very much tied to genetics, just not in a very direct way. That is, heritability is a measure of how much of the variance in a trait is explained by genetics vs environment. In this sense, wealth will have a relatively low heritability, because it is weekly tied to genetics, even though it is very much a trait most people inherit from their parents. Skin color will have a high heritability, because the variance in skin color is almost entirely explained by genetics.
The unintuitive part is that traits with almost no genetic variance at all, such as the number of arms, have very low heritability - since, in a population study, almost the entire variance in the number of arms will be explained by environmental factors (very very few families have 1 or 3 arms as a recurring trait - and there are way more people who lose their arms during life).
"Welcome to science hell, professor. This is IshKebab, he once saw something on the internet about your field of expertise and is going to spend eternity lecturing you on it."[1]
I don’t know about this. After some time sitting with it, I think that mid level and senior ICs - especially those slow to adapt - are going to be at risk of getting replaced by entry level “AI native” kids. Net on net it probably washes out to “normal” patterns of turnover and hiring once things settle.
Think “Smithers, we need to hire some of these kids who know computers!” Only fast forward about 30 years and str.replace(“computers”,”agents”).
That would only be true is AI usage experience was equivalent to domain experience, especially since the former keeps getting easier. If anything, companies might want to hold onto their seniors and midlevels, because they collectively decimated the process of creating new ones by refusing to hire and train younger workers. If later down the line they have a need for someone young and AI-experienced, they could just reach out into the endless job market and scoop up as much as they like.
In some ways domain experience can be a hindrance, with ingrained pathways and practices shaped by constraints that no longer apply. My personal opinion is that you probably want a mix of domain experts who are enthusiastic about AI and some kids who are free of preexisting dogma, and are willing challenge assumptions and try out things that the old heads might chafe at.
An example from software engineering is that all production code should undergo meticulous human review. Saying “no” to this sounds crazy to an experienced SWE, but might not actually be that crazy.
I think the constraints will remain in some fields, especially where there is a high price to pay for mistakes and consequently additional regulation. You can't vibe review code that will run on medical equipment, aircraft systems or industrial machinery. It doesn't matter how few people work in these fields, the fact that they shut off the tap to making new domain experts, while everyone and their grandma is learning to use AI will mean that the experts will eventually be at a shortage after retirements, while the enthusiastic AI users will be very abundant and underpaid.
The comment above is on to something. I find CarPlay to much more valuable and much more of a lock in to the iPhone than Siri. I do not think I could ever go back to using the infotainment systems that ship with cars. So makes sense why they might prioritize over Siri. And in the context of CarPlay, the simplicity of Siri is nice. I really only need it to execute a few simple commands like looking up directions, making calls, reading / sending texts, playing a podcast, etc.
I don’t quite follow the reasoning behind this being “unbiased”. Studies must be designed to answer specific research question. What is the question being asked by the study in question. If the question being asked is whether AB the right target (in humans), then it would need to consider other forms of human evidence, such as genetic evidence, alongside alternative explanations of failure — bad molecule, wrong isoform, bad cohort selection, etc.
The question asked in the Cochrane review linked to above is whether there is evidence that current treatments targeting amyloid provide relief. Their conclusion is that the effect is too small to be considered beneficial.
Especially when considering the undoubted side-effects, which include edema and micro-bleeding in the brain.
It doesn't ask whether amyloid is causative of the disease.
I wonder. Because if that was the question, why would they include anti-amyloid treatments that were never approved and are not part of any current treatment regimen and dilute the real effect of the approved antibodies? (Not that the approved antibodies are a silver bullet, but clearly they have some positive effect.)
The Cochrane review did not ask the question of whether Lecanemab (for example) worked, it asked the more general question of whether targeting amyloid worked. If targeting amyloid, generally, worked you might expect all approaches to targeting amyloid to show some efficacy. So they included all those for which clinical trials had been conducted. Obviously some treatments didn't pass the threshold required and so were not subsequently approved. If you only included successful trials you would bias the outcome.
It's possible, of course, that some methods of targeting amyloid might work where others failed. But even those that claim success (Donanemab and Lecanemab) have a very modest benefit.
Yes, the new anti-amyloids are not a revolutionary therapy. But they do have a small but definite effect on disease-progression. (something which in a highly noisy indication like Alzheimer's may actually mean a lot more to the patients than it seems by looking at the data. E.g. donepezil and memantine have very modest effects too, but patients and caregivers are still surprisingly positive about them).
And since these were developed using different methods and epitopes, it does not make sense to pool them together when doing a comparison. It's obvious that if you combine modestly beneficial compounds and compounds with zero beneficial effect then the mean effect will be worse than the modestly beneficial compounds.
And when you do so just because they were aimed at the same molecular target, not even considering if anyone claims that the old, unapproved antibody-therapies, with no positive studies to support them, have any effect, what you are doing is somewhere between extreme ignorance and deception.
All the trials included in the Cochrane review demonstrated removal of Aβ amyloid from the brain. Side effects were relatively modest, except for edema and micro-bleeds. So they are all legitimate tests of the hypothesis that removing Aβ amyloid would be beneficial.
No you would not expect all efforts to target AB to work. You can have the right target and the wrong molecule. Look at GLP1s as an example. Multiple pharma companies tried GLP1 receptor agonists for more than a decade and it was not until recently that Novo and Lilly got them to work.
But has only weak/moderate effect on weight loss compared to semaglutide, tirzepatide and retatrutide.
Alzheimer is a lot more difficult to measure than weight and amyloid-beta in the brain obviously significantly harder to target than the GLP-1 receptor in the body.
Also, a review pooling together all GLP-1's (including those that failed in development) and concluding that they as a class have only moderate or weak effect on weight loss, would obviously be badly misleading.
If you grouped all GLP-1 agonists that had entered clinical trials, they would be shown as successful [0]. No GLP-1 receptor agonist has failed to reduce appetite in clinical trials, but some were better than others.
The point about the anti-amyloid trials is that they all succeeded in removing amyloid. But they did not improve cognition, and only some resulted in a slightly slower rate of decline in cognition (people still declined).
>Forty-seven RCTs were included, with a combined cohort of 23,244 patients. GLP-1 RAs demonstrated a mean weight reduction of -4.57 kg (95% CI -5.35 to -3.78), mean BMI reduction of -2.07 kg/m2 (95% CI -2.53 to -1.62), and mean waist circumference reduction of -4.55 cm (95% CI -5.72 to -3.38) compared with placebo.
Your meta review has like 80%-90% of the studies on semaglutide and liraglutide. When finding random effect per drug, they even say this:
> Hence, the drugs eligible for meta-regression were semaglutide (subcutaneous injection, once weekly) and liraglutide (subcutaneous injection, once daily), which both showed a dose-dependent treatment effect in terms of weight and BMI.
And the class-effect is much more modest than semaglutide by itself would have shown, exactly as I said. What is the point?
Both meta-analyses above cover all eligible clinical trials for the respective strategies, targeting the GLP-1 receptor, or targeting amyloid.
For the GLP-1 strategy, Table 1 in the above paper shows that all drugs were clearly effective. Some were better than others.
For the amyloid strategy, only some of the drugs were effective, and that efficacy was very weak, even when all the drugs removed amyloid. This brings into question the hypothesis that removing amyloid would alleviate AD, i.e. the slender benefits of Lecanemab might not be caused by amyloid removal. The same cannot be said for the GLP-1 receptor agonists.
Nice work! Here is an article you may find helpful if you have not already come across it.[0]. You may also want to consider benchmarking against some non ML methods.[1]
This is going to catch some heat, but what if the most important professional “developer skill” to learn or improve is how to effectively use coding agents?
I saw something similar in ML when neural nets came around. The whole “stack moar layerz” thing is a meme, but it was a real sentiment about newer entrants into the field not learning anything about ML theory or best practices. As it turns out, neural nets “won” and using them effectively required development and acquisition of some new domain knowledge and best practices. And the kids are ok. The people who scoffed at neural nets and never got up to speed not so much.
Edit: as an aside, I have learned plenty from reviewing coding agent generated implementations of various algorithms or methods.
> what if the most important professional “developer skill” to learn or improve is how to effectively use coding agents?
Well, it's not. There's a small moat around that right now because the UX is still being ironed out, but in a short while able to use coding agents will be the new able to use Excel.
What will remain are the things that already differentiate a good developer from a bad one:
- Able to review the output of coding agents
- Able to guide the architecture of an application
> in a short while able to use coding agents will be the new able to use Excel.
Yeah, but there’s “able to use Excel”, and then there’s “able to use Excel.”
There is a vast skill gap between those with basic Excel, those who are proficient, and those who have mastered it.
As in intermittent user of Excel I fall somewhere in the middle, although I’m probably a master of knowing how to find out how to do what I need with Excel.
The same will be true for agentic development (which is more than just coding).
And the last two are much more important.
Don't forget that most decision makers and people with capital are normies, they don't live in a tech bubble.
If we know the outcome of that code, such as whether it caused bugs or data corruption or a crappy UX or tech debt -- which is potentially available in subsequent PR commit messages -- it's still valuable training data.
Probably even more valuable than code that just worked, because evidently we have enough of that and AI code still has issues.
I see this line of thought put out there many times, and I've been thinking: why do people do anything at all? What's the point? If no one at all is even reviewing the output of coding agents, genuinely, what are we doing as a society?
I fail to see how we transition society into a positive future without supplying means of verifying systemic integrity. There is a reason that Upton Sinclair became famous: wayward incentives behind closed doors generally cause subpar standards, which cause subpar results. If the FDA didn't exist, or they didn't "review the output", society would be materially worse off. If the whole pitch for AI ends with "and no one will even need to check anything" I find that highly convenient for the AI industry.
You could e.g. write specs and only review high level types plus have deterministic validation that no type escapes/"unsafe" hatches were used, or instruct another agent to create adversarial blackbox attempts to break functionality of the primary artifact (which is really just to say "perform QA").
As a simple use-case, I've found LLMs to be much better than me at macro programming, and I don't really need to care about what it does because ultimately the constraint is just that it bends the syntax I have into the syntax I want, and things compile. The details are basically irrelevant.
Code quality will impact the effectiveness of ai. Less code to read and change in subsequent changes is still useful. There was a while where I became more of a paper architect and stopped coding for a while and I realized I wasn't able to do sufficient code reviews anymore because I lacked context. I went back into the code at some point and realized the mess my team was making and spent a long while cleaning it up. This improved the productivity of everyone involved. I expect AI to fall into a similar predicament. Without first hand knowledge of the implementation details we won't know about the problems we need to tell the AI to address. There are also many systems which are constrained in terms of memory and compute and more code likely puts you up against those limits.
I don't disagree that code quality is currently more important than it's ever been (to get the most out of the tools). I expect that quality will increase though as people refine either training or instructions. I was able to get much better (well factored, aligned to business logic) output that I'm generally happy-ish with a couple months ago with some coding guidelines I wrote. It's possible that newer models don't even need that, but they work well enough with it that I haven't touched those instructions since.
I mean, sure, for programming macros. Or programming quick scripts, or type-safe or memory-safe programs. Or web frontends, or a11y, or whatever tasks for which people are using AI.
But if you peel back that layer to the point where you are no longer discussing the code, and just saying "code X that does Y"... how big is X going to get without verifying it? This is a basic, fundamental question that gets deflected by evaluating each case where AI is useful.
When you stop being specific about what the AI is doing, and switch to the general tense, there is a massive and obvious gap that nobody is adequately addressing. I don't think anyone would say that details are irrelevant in the case of life-threatening scenarios, and yet no one is acknowledging where the logical end to this line of thinking goes.
I mean, the promise of perfect AI and perfect robotics is that humans would no longer have to do anything. They could live a life of leisure. Unfortunately, we're going to get these perfect AI and perfect robotics before we transition socially into a post-scarcity, post-ownership society. So what will happen is that ownership of the AI and robots will be consolidated into the hands of the few, the vast rest of us will have nothing economically relevant to do, and we'll probably just subsist or die.
We're already seeing this today. Every year, thousands of people are becoming essentially irrelevant to the economy. They don't own much, they don't invest much, they don't spend much money, they don't make much money, and they are invisible to economics.
> They don't own much, they don't invest much, they don't spend much money, they don't make much money, and they are invisible to economics.
Indeed. Sometimes I think the so-called “lower classes” end up functioning more like crops to be farmed by the rich. Think, dollar stores that sell tiny packages of things at worse unit cost, checking account fees, rent-a-center, 15% interest auto loans and store credit cards with 30% interest…
I've definitely felt this kind of way in the past. But these days I'm not so sure.
Setting aside the AI point about it, the idea of people becoming essentially irrelevant to the economy is an indictment on society. But I'd argue that the indictment really is towards what constitutes measurement in the economy. Not an indictment on society itself, or technology.
Sure, someone may not spend much money or produce much money, but if they produce scientific research or cultural work that is intangibly valuable it is still valuable regardless of whether economists can point to a metric or not. Same goes for the infinite amounts of contributions to our world from nature: what is the economic value of a garden snake or a beetle? A meaningless question when the economy can only see things in dollars.
They will still be turning out the same problematic code in a few years that they do now, because they aren’t intelligent and won’t be intelligent unless there is a fundamental paradigm shift in how an LLM works.
I use LLMs with best practices to program professionally in an enterprise every day, and even Opus 4.6 still consistently makes some of the dumbest architectural decisions, even with full context, complete access to the codebase and me asking very specific questions that should point it in the right direction.
I keep hearing “they aren’t intelligent” and spit out “crap code”. That’s not been my experience. LLMs prevented and also caught intricate concurrency issues that would have taken me a long time.
I just went “hmmm, nice” and went on. The problem there is that I didn’t get that sense of accomplishment I crave and I really didn’t learn anything. Those are “me” problems but I think programmers are collectively grappling with this.
They are not intelligent. Full stop. Very sophisticated next word prediction is not intelligence. LLMs don’t comprehend or understand things. They don’t think, feel or comprehend things. That’s just not how they work.
That said, very sophisticated next word predictors can and sometimes do write good code. It’s amazing some of the things they get right and then can turn around and make the weirdest dumbest mistakes.
It’s a tool. Sometimes it’s the right tool, sometimes it’s not.
None of those things will be necessary if progress continues as it has. The AI will do all of that. In fact it will generate software that uses already proven architectures (instead of inventing new ones for every project as human developers like to do). The testing has already been done: they work. There are no vulnerabilites. They are able to communicate with stakeholders (management) using their native language, not technobabble that human developers like to use, so they understand the business needs natively.
If this is the case then none of us will have jobs; we will be completely useless.
I think, most likely, you'll still need developers in the mix to make sure the development is going right. You can't just have only business people, because they have no way to gauge if the AI is making the right decisions in regards to technical requirements. So even if the AI DOES get as good as you're saying, they wouldn't know that without developers.
For some definition of work, yes, not every definition. Their product is not without flaw, leaving room at for improvement, and room for improvement by more than only other AI.
> There are no vulnerabilities
That's just not true. There's loads of vulnerabilities, just as there's plenty of vulnerabilities in human written code. Try it, point an AI looking for vulns at the output of an AI that's been through the highest intensity and scrutiny workflow, even code that has already been AI reviewed for vulnerabilities.
> This is going to catch some heat, but what if the most important professional “developer skill” to learn or improve is how to effectively use coding agents?
If it does go as far that way as many seem to expect (or, indeed, want), then most people will be able to do it, there will be a dearth of jobs and many people wanting them so it'll be a race to the bottom for all but the lucky few: development will become a minimum wage job or so close to that it'll make no odds. If I'm earning minimum wage it isn't going to be sat on my own doing someone else's prompting, I'll find a job that involves not sitting along in front of a screen and reclaim programming for hobby time (or just stop doing it at all, I have other hobbies to divide my time between). I dislike (effectively) being a remote worker already, but put up with it for the salary, if the salary goes because “AI” turns it into a race-to-the-bottom job then I'm off.
Conversely: if that doesn't happen then I can continue to do what I want, which is program and not instruct someone else (be it a person I manage or an artificial construct) to program. I'm happy to accept the aid of tools for automation and such, I've written a few of my own, but there is a line past which my interest will just vanish.
What the people excited about the race to the bottom scenario don’t seem to understand is that it doesn’t mean low skill people will suddenly be more employable, it means fewer high skill people will be employable.
No one will be eager to employ “ai-natives” who don’t understand what the llm is pumping out, they’ll just keep the seasoned engineers who can manage and tame the output properly. Similarly, no one is going to hire a bunch of prompt engineers to replace their accountants, they’ll hire fewer seasoned accountants who can confidently review llm output.
And those that do have not yet understood what will happen when those seasoned workers retire, and there are no juniors or mid that can grow because they have been replaced by AI
> What the people excited about the race to the bottom scenario
I'm not excited about it. I just see it as a logical consequence if what people are predicting comes to pass, and I've thought about how I will deal with that.
The endgame in programming is reducing complexity before the codebase becomes impossible to reason about. This is not a solved problem, and most codebases the LLMs were trained on are either just before that phase transition or well past it.
Complexity is not just a matter of reducing the complexity of the code, it's also a matter of reducing the complexity of the problem. A programmer can do the former alone with the code, but the latter can only be done during a frank discussion with stakeholders.
A vibe coder using an LLM to generate complexity will not be able to tell which complexity to get rid of, and we don't have enough training data of well-curated complexity for LLMs to figure it out yet.
No kidding. So far the complexity introduced by LLM-generated code in my current codebase has taken far more time to deal with than the hand-written code.
Overall, we are trying to "silo" LLM-generated code into its own services with a well-defined interface so that the code can just be thrown away and regenerated (or rewritten by hand) because maintaining it is so difficult.
Yeah, same. I like the silo idea, I'll have to explore that.
I'm relieved to hear this because the LLM hype in this thread is seriously disorienting. Deeply convinced that coding "by hand" is just as defensible in the LLM age as handwriting was in the TTY age. My dopamine system is quite unconvinced though, killing me.
I have a silo’d service that handles file uploads of PDFs, images and so on. It was largely vibe coded.
It sits on an isolated tier and isn’t allowed to persist state or have permanent storage. We wanted to reduce the impact of a security flaw in this code.
We’ve ended up doing similar things for search and for an orchestration tool used for testing. The key thing is it’s non critical so we can live without it.
Yes, a retreading of the accidental vs. implicit complexity discussion is in order here. I asked an AI agent to implement function calls in a programming language the other day. It decided the best way to do this was to spin up a new interpreter for every function call and evaluate the function within that context. This actually worked but it was very very very slow.
The only way I was able to direct the AI to a better design was by saying the words I know in my head that describe better designs. Anyone without that knowledge wouldn't be able to tell the heavy interpreter architecture wasn't good, because it was fast enough for simple test cases which all passed.
And you can say "just prompt better" but we're very quickly coming to a place where people won't even have the words to say without AI first telling them what they are. At that point it might as well just say "The design is fine don't worry about it" and how would the user know any better.
I also remember a similar wave around 10-15 years ago regarding ML tooling and libraries becoming more accessible, more open source releases etc. People whose value add was knowing MATLAB toolboxes and keeping their code private got very afraid when Python numpy, scikit learn and Theano etc came to the forefront. And people started releasing the code with research papers on github. Anyone could just get that working code and start tweaking the equations put different tools and techniques together even if you didn't work in one of those few companies or didn't do an internship at a lab who were in the know.
Or other people who just kept their research dataset private and milked it for years training incrementally better ML models on the same data. Then similar datasets appeared openly and they threw a hissy fit.
Usually there are a million little tricks and oral culture around how to use various datasets, configurations, hyperparameters etc and papers often only gave the high level ideas and math away. But when the code started to become open it freaked out many who felt they won't be able to keep up and just wanted to keep on until retirement by simply guarding their knowledge and skill from getting too known. Many of them were convinced it's going to go away. "Python is just a silly, free language. Serious engineers use Matlab, after all, that's a serious paid product. All the kiddies stacking layers in Theano will just go away, it's just a fad and we will all go back to SVM which has real math backing it up from VC theory." (The Vapnik-Chervonenkis kind, not the venture capital kind.)
I don't want to be too dismissive though. People build up an identity, like the blacksmith of the village back in the day, and just want to keep doing it and build a life on a skill they learn in their youth and then just do it 9 to 5 and focus on family etc. I get it. But wishing it won't make it so.
Talented, skilled people with good intuition and judgements will be needed for a long time but that will still require adapting to changing tools and workflows. But the bulk of the workforce is not that.
This is so true... I am having issues with the change right now.. being older and trying to incorporate agentic workflow into MY workflow is difficult as I have trust issues with the new codebase.. I do have good people skills with my clients, but my secret sauce was my coding skilz.. and I built my identity around that..
The cure for me has been to write an agent myself from first principles.
Tailored to my workflow, style, goals, projects and as close as possible to what I think is how an agent should work. I’m deliberately only using an existing agent as a rubber duck.
Using a coding agent seems quite low skill to me. It’s hard to see it becoming a differentiator. Just look at the number of people who couldn’t code before and are suddenly churning out work to confirm that.
I think your argument is predicated on LLM coding tools providing significant benefit when used effectively. Personally I still think the answer is "not really" if you're doing any kind of interesting work that's not mostly boilerplate code writing all day.
Define interesting. In my experience most business logic is not innovative or difficult, but there are ways to do it well or ways to do it terribly. At the senior levels I feel 90% of the job is deciding the shape of what to build and what NOT to build. I find AI very useful in exploring and trying more things but it doesn’t really change the judgment part of the job.
How much of software programmer work is interesting? A fraction of a percent? I'd argue most of us including most startups work on things that help make businesses money and that's pretty "boring" work.
It absolutely is, but the fundamental misunderstanding around this seems to be that "effectively using coding agents" is a superset of the 2023-era general understanding of "Senior Software Engineer".
At least when you're talking about shipping software customers pay for, or debugging it, etc. Research, narrow specializations, etc may be a different category and some will indeed be obsoleted.
I don’t think it could be the most important skill to have. The most common, and the most standardized one for sure, but if coding agents are doing fundamental R&D or running ops then nobody needs skills anyway.
> As it turns out, neural nets “won”
> The people who scoffed at neural nets and never got up to speed not so much.
I get the feeling you don’t know what you’re talking about. LLMs are impressive but what have they “won” exactly? They require millions of dollars of infrastructure to run coming around a decade after their debut, and we’re really having trouble using them for anything all that serious. Now I’m sure in a few decades’ time this comment will read like a silly cynic but I bet that will only be after those old school machine learning losers come back around and start making improvements again.
Neural nets are used in way more applications than just LLMs. They did win. They won decisively in industry, for all kinds of tasks. Equating the use of one with the other is a pretty strong signal of:
> you don’t know what you’re talking about
Consider: Why did Google have a bazillion TPUs, anyway?
Not sure why this would catch heat rationally speaking. It is quite clear in a professional setting effective use of coding agents is the most important skill to develop as an individual developer.
It’s also the most important capability engineering orgs can be working on developing right now.
I'd offer an edit that the most important skill may be knowing when the agent is wrong.
There's so much hand wringing about people not understanding how LLMs work and not nearly enough hand wringing about people not understanding how computer systems work.
I'd say viewing it as most important is pretty unprofessional. But isn't it the point of this extreme AI push? To replace professional skills with dummy parrots.
> This is going to catch some heat, but what if the most important professional “developer skill” to learn or improve is how to effectively use coding agents?
Doing so will effectively force a (potentially unwanted) career change for many people and will lead to the end of software engineering (and software as a career), assuming AI continues to improve.
"Effectively" using agents means that you're writing specs and reading code (in batches through change diffs) instead of writing code directly. This requires the ability to write well (or well enough to get what you want from the agent) and clearly communicate intent (in your language of choice, not code; very different IMO).
The way that you read code is different with agents as well. Agents can produce a smattering of tests alongside implementation in a single turn. This is usually a lot of code. Thus, instead of red-green-refactor'ing a single change that you can cumulatively map in your head, you're prompt-build-executing entire features all at once and focusing on the result.
Code itself loses its importance as a result. See also: projects that are moving towards agentic-first development using agents for maintenance and PR review. Some maintainers don't even read their codebases anymore. They have no idea what the software is actually doing. Need security? Have an agent that does nothing but security look at it. DevOps? Use a DevOps agent.
This isn't too far off from what I was doing as a business analyst a little over 20 years ago (and what some technical product managers do now for spikes/prototypes). I wrote FRDs [^0] describing what the software should do. Architects would create TRDs [^1] from those FRDs. These got sent off to developers to get developed, then to QA to get bugs hammered out, then back to my team for UAT.
If agents existed back then, there would've been way fewer developers/QA in the middle. Architects would probably do a lot of what they would've done. I foresee that this is the direction we're heading in, but with agents powered by staff engineers/Enterprise Architects in the middle.
> Edit: as an aside, I have learned plenty from reviewing coding agent generated implementations of various algorithms or methods.
People learn differently. I (and others) learn from doing. Typing code from Stack Overflow/Expertsexchange/etc instead of pasting it, then modifying it is how I learned to code. Some can learn from reading alone.
> This requires the ability to write well (or well enough to get what you want from the agent) and clearly communicate intent (in your language of choice, not code; very different IMO).
I do not see why you can't write your spec in pseudocode if you really want to - communicating your intent to the LLM, for how the code should be developed is far closer to programming than writing skillwise.
Doing so will effectively force a (potentially unwanted) career change for many people and will lead to the end of software engineering (and software as a career), assuming AI continues to improve.
If you expected things to stay the same forever, maybe software engineering wasn't the right career move for you. Even though it looked safe enough, given that we've spent 50 years writing the same old code the same old way, that was never guaranteed.
I for one am glad to see something genuinely new come along. The last dozen or so "paradigm shifts" turned out to be disappointing variations on the same old paradigm. Not this one, though.
I think you missed the part where I outlined how software engineering will become a business analyst spec-writing kind of job, a job I did and know that I dislike...
But, hey! Different strokes for different folks. This might be for you, and that's cool! I'm allowed to be sad about it, though.
I think Peter Thiel is smart, but exhibiting one of smart people’s most common modes of failure, overestimating one’s ability while not maintaining a healthy sense of skepticism about the correctness of one’s own beliefs.
Put simply, he (and many other tech bros) have galaxy brained themselves into some very stupid stuff.
I associate this phrase with losers and people trying to sabotage the US. You know who is not wringing their hands about “elite overproduction”? China, who are pumping out tons of smart and capable STEM PhDs, and have in a relatively short time caught up to and in some cases surpassed the US in production of scientific output and technology.
Completely separate from the substance of your point, this sort of language does not encourage constructive dialog, it frames the discussion in such a way that you are either going to get
a. People who agree with you, resulting in you not learning anything
b. People who are triggered into fighting with you, once again, resulting in you not learning anything
c. People ignoring you, resulting in you not learning anything.
My constructive suggestion to you is that you simply don't write that first sentence. I suspect you (and everyone else!) will have a much more fruitful time online as a result!
Thank you for giving him the lesson on etiquette. I was going to do the same but you beat me to the punch, so instead I will just upvote you and move on without further remark.
Yeah, you are not wrong. The topic is a bit like troll bait for me. Probably because I have a first hand view of how the current strain of anti intellectualism and resulting policy in the US is destroying jobs and eroding competitive advantage. My observation is that this type of rhetoric tends to be produced and consumed by “elites”, and is often used to advocate for policy that limits socioeconomic mobility.
The irony is that in limiting mobility and competition from the “non elite” out-groups to preserve status, they end up shrinking the overall size of the pie.
I've always taken the elite overproduction thing as an _analytical tool_ to help us make sense of why we have experienced the rise of an oppositional anti intellectual position in contemporary culture.
But you make the good point that it can also be a _weapon_, leveraged by those oppositional groups, to justify their oppositional position.
Perhaps this seeming tautology can be resolved with some systems thinking. Maybe there's some insight in the elite overproduction analysis, but that means that, as an argument for further polarising society it's a pretty effective tool. It's actually reinforcing the feedback loop! A fascinating example of a self fulfilling prophecy.
The Economist only wants what's best for China. (As the article is paywalled, do they discuss the positive externalities of this glut or only the difficult labour market?)
In my view, STEM PhDs are not members of the “elite.”
Historically, in the US the elite are the managerial class, the lawyers (future politicians), and the coastal dilettantes who are already wealthy enough to major in the social sciences.
When 1+ million students are getting MBAs every year in the belief they will be members of the C-suite, but there’s only a few thousand such positions, you have a case of elite overproduction.
Have you looked at the Wikipedia article? China is specifically addressed there, with elite (as in - highly educated people) overproduction and unemployment reaching such levels that government is now suggesting they should seek manual labor jobs.
It works in China because they have growth. In the west thousands of college kids thought they could land cushy management positions or at least highly paid expert jobs.
Then these kids realise these jobs don’t exist, that they should have gone to trade school instead, and that their student debt will cripple them for life.
Same thing will happen in China. For now their economy grows so fast it can absorb many intellectuals, but that won’t last forever.
We live in a globalized economy. Rapid transport of people, goods, and information necessitates it. The high paying STEM jobs will go to wherever there is an abundance of talent, and the network effects are quite significant.
Per Turchin model, the declining population in China has created conditions for more elite-adjacent positions for all those STEM PhDs, preventing overproduction
I think the solution to “elite overproduction” is, not to educate people less, but to promise them a decent standard of living that seems throughout and ultimately is reasonably attainable (don’t over-promise and do minimize FUD).
The massive shift in careers, not just due to LLMs but technology and society in general, threaten the promises given to prior generations. And this is also happening in China, see “tang ping” / “lying flat”.
This is very easy to explain. Anthropic outlines some limitations in their terms of service. Palantir accepted those terms. The DoD did not.
OpenAI claims their terms of service for DoD contain the same limitations as Anthropics proposed service agreement. Anthropic claims that this is untrue.
Now given that (a) the DoD terminated their deal with Anthropic, (b) stated that they terminated because Anthropic refused modify their terms of service, and (c) then signed a deal with openAI; I am inclined to believe that there is in fact a substantial difference between the terms of service offered by Anthropic and OpenAI.
Yeah, it never made sense when Sam immediately said that they had the same constraints yet de DoW immediately agreed with that.
From what I can see, OpenAI’s terms basically say “need to comply with the law”, which provides them with plenty of wiggle room with executive orders and whatnot.
Are you sure about that? Every information I’ve seen suggests that the DoD has been using Anthropic’s models through Palantir.
My understanding is that Anthropic requested visibility and a say into how their models were being used for classified tasks, while the DoD wanted to expand the scope of those tasks into areas that Anthropic found objectionable. Both of those proposals were unacceptable for the other side.
Wasn’t the trigger for all this what happened with Maduro earlier this year? From what I understood, Anthropic wasn’t very happy how their systems were being used by the DoW through Palentir which caused this whole feud.
And why would they have an objection to that? They sold a product to a customer. They should have no business in how that customer uses their software.
> And why would they have an objection to that? They sold a product to a customer. They should have no business in how that customer uses their software.
They sold a service to a customer, contractually subject to terms they both agreed upon. How do people keep missing this? The government changed their mind after agreeing to the restrictions and tried to alter the deal with Anthropic ex-post-facto.
It’s a bit more complex than that, but to be fair I don’t know what they were expecting after they integrated a purpose-built model with Palantir to be deployed in high-security networks to carry out classified tasks.
I'd hate to break it to you, but companies do have a right to determine how their products are used. You were subject to that when you wrote that comment. Did you not notice that?
No, I do not think they do. If a buy a car a run somebody over on purpose, the manufacturer has no right to come take my car away. Even if it were to be written in a contract.
If you tell the car dealership that your plan is to run someone over with the car you are buying, they 100% have the right to refuse selling the car to you.
If you tell a gun dealer you're going to kill someone when you walk out of the shop, they have a right and an obligation to refuse the sale.
Please feel free to tell me how these analogies are incorrect.
“We’ve actually held our red lines with integrity rather than colluding with them to produce ‘safety theater’ for the benefit of employees (which, I absolutely swear to you, is what literally everyone at [the Pentagon], Palantir, our political consultants, etc, assumed was the problem we were trying to solve),” Amodei reportedly wrote.
“The real reasons [the Pentagon] and the Trump admin do not like us is that we haven’t donated to Trump (while OpenAI/Greg have donated a lot),” he wrote, referring to Greg Brockman, OpenAI’s president, who gave a Pac supporting Trump $25m in conjunction with his wife.
Another reason is that Sam Altman has been willing to "play ball" like providing high-profile (though meaningless) big announcements Trump likes to tout as successes. For example:
> "The Stargate AI data center project worth $500 billion, announced by US President Donald Trump in January 2025, is reportedly running into serious trouble.
More than a year after the announcement, the joint venture between OpenAI, Oracle, and Softbank hasn't hired any staff and isn't actively developing any data centers, The Information reports, citing three people involved in the "shelved idea."
Reminds me of when they cut the camera to Zuck and he made the $600 Billion Deal announcement, but was hot mic'd after and said "I'm sorry I wasn't ready... I wasn't sure what number you wanted to go with". I will be extremely surprised if half of these deals actually go through
Heritable != Molecular / Genetic Mechanism
There is a conflation of these terms in popular discourse that does a disservice to the field of statistical genetics, imo. There are mechanisms of inheritance that operate various length / time scales other than that of biological macromolecules. For example, if you tell me what language your parents natively speak I can tell you your primary language with >90% accuracy.
So before we start getting 3 replies deep into any thead, please remember that retrospective observational data measured with unqualified instruments is notoriously confounded and that we can barely infer causal structure in controlled functional genomics experiments (much less a GWAS of phewas). So let’s all please keep an open mind and not be so certain about our beliefs.