asgraham

If you can write tests fast enough, you can specify those business rules on the fly. The ideal case is that tests always reflect current business rules. Usually that may be infeasible because of the speed at which those rules change, but I’ve had a similar experience of AI just getting tests right, and even better, getting tests verifiably right because the tests are so easy to read through myself. That makes it way easier to change tests rapidly.

This also is ignoring that ideally business logic is implemented as a combination of smaller, stabler components that can be independently unit tested.

I totally agree that the agent is essential, and that right now Claude Code is semi-unanimously the best agent. But agentic tooling is written, not trained (as far as I can tell—someone correct me) so it’s not immediately obvious to me that a third-party couldn’t eventually do it better.

Maybe to answer my own question, LLM developers have one, potentially two advantages over third-party tooling developers: 1) virtually unlimited tokens, zero rate limiting with which to play around with tooling dev. 2) the opportunity to train the network on their own tooling.

The first advantage is theoretically mitigated by insane VC funding, but will probably always be a problem for OSS.

I’m probably overlooking news that the second advantage is where Anthropic is winning right now; I don’t have intuition for where this advantage will change with time.

I don’t know about your first point: at some point the three-year difference may not be worth the premium, as local models reach “good enough.”

But the second point seems even less likely to be true: why will Claude code and Gemini cli always be superior? Other than advantageous token prices (which the people willing to pay the aforementioned premium shouldn’t even care about), what do they inherently have over third-party tooling?

Fair enough. Cancer causation is just so weird, I'm going off the heuristic that stuff in your lungs that your lungs aren't prepared for is probably not great, especially with chronic exposure. And your lungs are really only prepared for air (which includes many things within certain tolerances [0]). The "not great" => "cancer" pipeline is really where the hand-waving comes in, and mostly far too early to tell whether the parts of vaping that are "not great" for your lungs will in fact be carcinogenic for your lungs.

So that being said, I'm mostly going to offer citations for "vaping is not great for your lungs." And that being said, I'm just going to offer citations for "specific parts of vaping are not good for your lungs." But my broader argument is that putting stuff in your lungs is going to be bad for your lungs, and these are just the most obvious ones we've found so far. Unfortunately I won't be able to find a citation for that argument.

So, first, the most recent: a study showing disposable vapes had incredibly high level of toxic metal emissions [1]. The non-disposable Juul et al variously have some concerning levels, but the insane numbers are on disposables, which are largely (entirely?) illegal in the US, at least. Still, they're not illegal everywhere, they were used heavily for several years in the US, and several of the top Google results were redditors complaining about the stupid ban and talking about how to get around it. All of this combines to lung damage down the line, and several of the toxic metals are outright carcinogenic, so lung cancer as well.

A more particular example: popcorn lung is a terrifying name, but pretty restricted risk, given the causing chemical is only in certain flavors, and those have supposedly stopped using it [2]. But again, an example of weird chemicals in your lungs cause weird things, and it'll be decades til we figure out all of them.

And finally, a study showing that vaping plus smoking leads to a four-fold higher risk of lung cancer over smoking alone (yes, they adjusted for age, gender, race, location of residence, prevalent comorbidities, and pack-years of smoking) [3].

[0] I was hoping to make a glib point about even high enough pollen concentration being bad for your lungs, but in fact a recent study suggests that allergies reduce risk of lung cancer! I'm chalking that up to allergies being your body's way of keeping non-air particulates from your lungs, but who knows. https://www.frontiersin.org/journals/medicine/articles/10.33...

[1] https://pubs.acs.org/doi/10.1021/acscentsci.5c00641

[2] Not a lot of research on popcorn lung, seemingly. Note that the name is related to its etiology, not its symptoms. https://www.summahealth.org/flourish/entries/2025/03/a-warm-...

[3] https://pubmed.ncbi.nlm.nih.gov/39210964/

I can’t find any complete numbers (most prevalent factors after smoking would be environmental, and therefore underreported, especially when as hard to detect as radon), but national health agencies tend to put the Radon section second, after Smoking [1,2]. An uncited figure on a Hopkins webpage suggests 30% of non-smoking lung cancer cases are caused by Radon [3]. Among the well-known environmental factors (asbestos, secondhand smoke), it seems to be about equal for risk increase [4]. Given that asbestos and secondhand smoke are on the decline, it stands to reason that radon will tend toward being the top cause, barring a rise in prevalence of one of the disease risk factors (asthma, pneumonia, HIV, tuberculosis).

Of course this is all moot because vaping will be revealed to be the current #1 cause of lung cancer in the coming decades, by a long shot. No citation necessary.

[1] https://www.cdc.gov/lung-cancer/risk-factors/index.html

[2] https://www.nhs.uk/conditions/lung-cancer/causes/

[3] https://www.hopkinsmedicine.org/health/conditions-and-diseas...

[4] https://pmc.ncbi.nlm.nih.gov/articles/PMC6777859/

Not to agree with GP, but I think it’s more accurate to say they’re saying “if validation is quick (to code), who cares how long a solution takes an AI because computation is cheap.”

They’re not really making any claims about how quickly the AI can solve relative to the validation, which is what P vs NP is about.

> It's actually hard to find a math tutor to proof your vector math program in your late 30s.

They exist, if you know where to look and are willing to pay (source: me, or generally and probably more affordably wyzant.com)

What a wild thing to comment.

a) “professional” writ large often encompasses academic research, even though it’s also often specifically used in contrast. Language is weird like that. Former and current academics like to have a term for all the stuff we’ve been paid to do, and we usually default to “professional experience” (though current academics do usually say “research experience” because you’re right that professional has the default connotation of corporate employment).

b) the project seems to have been over 2000 lines of APL. That’s not the “small” student project you’re implying, in any language, and in APL that’s quite substantial.

re: the arxiv link

Why is it that microarray true positive p-values follow a beta distribution? Following the citations led to a lot of empirical confirmation but I couldn't find any discussion of why.

More to the point of this rebuttal, though: why would we expect the amalgamation of 70k micro-array experiments' abstract-reported p-values to follow a single beta distribution? And what about modeling the bias-induced bump of barely-significant results?

If there's some theoretical reason why the meta-study can use the beta-uniform model, then I could see this being only a mild underestimation of the proportion of false positives (14%), but otherwise I'm confused how we can interpret this.

I knew it wasn't for inspectors of physical buildings (though that took a minute to accept; I thought it might be a joke), but I could not figure out what kind of buildings it would be. Was it some lisp thing I hadn't heard of? I've been coming back to these comments waiting for someone to explain and it finally hit me: it's for the building of inspectors. Language is weird.

As a chronic premature optimizer my first reaction was, "Is this even possible in vanilla python???" Obviously it's possible, but can you train an LLM before the heat death of the universe? A perceptron, sure, of course. A deep learning model, plausible if it's not too deep. But a large language model? I.e. the kind of LLM necessary for "from vanilla python to functional coding assistant."

But obviously the author already thought of that. The source repo has a great motto: "It don't go fast but it do be goin'" [1]

I love the idea of the project and I'm curious to see what the endgame runtime will be.

[1] https://github.com/bclarkson-code/Tricycle

For me it makes it harder to read. I think it's because my skimming muscles are so tuned to both bold and first-sentence text that when presented such a stark opportunity, I naturally try to jump from one bold title to the next. But the bolded "titles" contain no substantive information, so I have to fight that urge, which means I can't skim at all, even though my attention reeeeally wants to jump.

I'm absolutely behind both the message and call to action of this article; the evidence is irrefutable; this comment should not be interpreted as calling into question the headline at all: the underlying data looks solid.

However. Why oh why did this author trash their own credibility by explicitly citing the "even more alarming [surge] at the state level" where "walker deaths had increased a shocking 266.67 percent in Nebraska, 150 percent in New Hampshire, and 87.5 percent in Delaware." They of course give the bare minimum caveat that "[t]hose dramatic numbers are partly explained by those states' small populations," but they are entirely explained by those states' small pedestrian fatality rates. Such as New Hampshire's increase from 2 to 5 fatalities. The other two "alarming" surges are equally unmentionable[1] in the light of actually reading the numbers.

I do have to give them credit for including the relevant numbers directly in an image in the article so that I didn't have to do any digging to figure out why those states had such high increases. But that just makes me even more baffled why they didn't catch themselves.

[1] Literally unmentionable, as in: those particular percentages cannot in good conscience be mentioned anywhere near a self-respecting argument for transportation reform.

It's hard to say what the hardest constraint will be, at this point. Imaging and scanning are definitely hard obstacles; right now even computational power is a hard obstacle. There are 100 trillion synapses in the brain, none of which are simple. It's reasonable to assume you could need a KB (likely more tbh) to represent each one faithfully (for things like neurotransmitter binding rates on both ends, neurotransmitter concentrations, general morphology, secondary factors like reuptake), none of which is constant. That means 100 petabytes just to represent the brain. Then you have to simulate it, probably at submillisecond resolution. So you'd have 100 petabytes of actively changing values every millisecond or less. That's 100k petaflops, at a bare, bare, baaaare minimum, more like an exaflop.

This ignores neurons since there are only like 86 billion of them, but they could be sufficiently more complex than synapses that they'd actually be the dominant factor. Who knows.

This also ignores glia, since most people don't know anything about glia and most people assume that they don't do much with computation. Of course, when we have all the neurons represented perfectly, I'm sure we'll discover the glia need to be in there, too. There are about as many glia as neurons (3x more in the cortex, the part that makes you you, coloquially), and I've never seen any estimate of how many connections they have [1].

Bottom line: we almost certainly need exaflops to simulate a replicated brain, maybe zettaflops to be safe. Even with current exponential growth rates [2] (and assuming brain simulation can be simply parallelized (it can't)), that's like 45 years away. That sounds sorta soon, but I'm way more likely to be underestimating the scale of the problem than overestimating it, and that's how long until we can even begin trying. How long until we can meaningfully use those zettaflops is much, much longer.

[1] I finished my PhD two months ago and my knowledge of glia is already outdated. We were taught glia outnumbered neurons 10-to-1: apparently this is no longer thought to be the case. https://en.wikipedia.org/wiki/Glia#Total_number

[2] https://en.wikipedia.org/wiki/FLOPS#/media/File:Supercompute...

"...a sorta-popular JavaScript package. I say “sorta” because it’s used by lots of people, but it’s not pervasive. It had 105 million downloads in 2022."

Is there any standard of popularity where 100 million+ downloads in a year is only "sorta" popular?

They can't possibly be working with genomes spanning the past 250k years--- the oldest known human remains are only estimated to be ~230k years old, and I doubt they have parent/child trios nicely spanning the intervening few hundred thousand years. So they have to be working on mutation rate inferences.

From a non-expert reading of the article, their pipeline is more like: modern human population genomes -> estimates of modern mutations' ages -> estimate of average historical parental ages.

The first estimation (genome -> mutation ages) was carried out in a prior study "GEVA" [1], and this paper's contribution seems to be estimating the average parental ages based on those previously estimated historical mutation rates.

I couldn't find any mention of using old DNA samples. The GEVA study pulls from two genome databases, TGP [2] and SGDP [3], both of which seem to be entirely modern genomes. I'm not an expert in the field, so maybe it's obvious to a population geneticist that these databases do include old genomes.

Given that they're only using modern (surviving) genomes, the critique of survivorship bias seems valid.

[1] https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6992231/

[2] https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4750478/

[3] https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5161557/

The author of The Wandering Inn [1], a web serial, has been trying to write less the past few months (the past year?). They peaked around 80k words/week, and have been aiming closer to 30-40k words/week with more breaks. They've had mixed success (e.g. last week they wrote and released 70k).

They only post updates/reflections on their progress towards this goal in the author's notes on the biweekly chapter releases, so I'm not sure how you could follow along, if that's what you're looking for. I guess you could just follow the releases and only read the notes. Or you could try catching up haha.

[1] https://wanderinginn.com/

Sure, this is ideal given unlimited time and money, but hiring dedicated devs for tooling is going to be expensive. And once they fix "the problem," they're still on payroll. Yay they'll keep fixing problems and improving the software, but it's basically the most expensive software subscription model imaginable.

"A feeling of safety comes from within." --- This is an actively counterproductive statement. Just look at the impetus behind "defund the police": the real police violence causes people of color to feel unsafe. On the flipside, crime often involves real violence that makes people feel unsafe. You can argue that policing isn't the answer to crime, but real danger is the root of feeling unsafe, even if (like everything) it is mediated by perception. A "culture of safety" must be rooted in a reality of safety.

I'm torn about this sort of technology. On the one hand, the best way to drive adoption is external recording. On the other hand, the data are sooooo coarse-grained. And on the third hand, we don't have anything resembling the theoretical framework we'd need to process any finer-grained data at the scale of the whole brain, so maybe this coarse-grained external recording really is the best first step. With those caveats, I do think this has potential important concrete applications, even if "mind reading" is likely beyond the scope.

The fundamental question for these devices will not be one of neuroscience or practicality: the fundamental question will be how good is their signal-to-noise ratio (SNR). If they even approach clinical EEG SNR using their fNIRS, then these will be widely adopted, off the top of my head, for two uses: seizures and meditation (and the latter doesn't require anything even vaguely approaching clinical SNR).

Epilepsy is one of the most prevalent neurological conditions (the most prevalent?) in the world, affecting more than 1% of people, and it's somewhat hard to diagnose effectively due to sparsity of seizures. You can only have patients hooked up to an EEG for so long, and you just have to hope that you get a seizure during that time. Especially when doing preliminary localization in preparation to surgically intervene (the only effective treatment in ~1/3 of epilepsy cases), you need to catch that seizure when it happens. So any vaguely cheap option that allows patients to monitor for seizures in their day-to-day will see widespread adoption by hospitals. (This doesn't even account for the potential to predict seizures by a matter of seconds or minutes, which can be invaluable for someone with moderately frequent seizures.)

The meditation one sounds more out-there (especially if you read the OP article...), but is actually pretty well supported by studies. Particular results are debatable, but it's fairly well-established that meditation causes a large scale change in the frequency composition of brain waves that's pretty easy to capture on EEG. What's more, you can induce these changes consciously by providing feedback (e.g. a moving dot on a screen), a result demonstrated in monkeys (and I hope in humans, though I haven't seen that myself). There are still some lingering questions, such as the relative efficacy of inducing these changes via feedback vs via meditation, but they're questions at the level that they'll be overlooked by whatever company invites you to learn to meditate with these helmets. Intuitively, you could think of this as analogous to those posture-feedback devices I see advertised everywhere. Those notice your back is bending too much, and alert you. This would notice your thoughts are wandering, and alert you.

A third, very vague, but important consequence of a device that (finally) has good SNR is the advent of a critical-mass hacker culture around brain-computer interfaces.

If anyone knows anything about the SNR, or about more recent results on meditation, I'd love to read more.

> 'strengthening' neurons with a fiber like substance around it reduces the electrical activity required to achieve the same neural-result.

I think you're talking about myelination, which is a result that's been around for quite a while. Myelination probably wouldn't confound this kind of "brain reading" because it's relatively stable, i.e. changes occur over the course of days at a minimum, and supposedly infrequently in the fully developed brain. Any "brain reading" will have to account for changing neural efficacies, as that's what learning is, and the tricky stuff would be learning on the scale of seconds or minutes (e.g. changes in synaptic efficacy) rather than things like myelination.

That being said, while the conventional wisdom is that myelin is stable over the course of months/years in the adult brain, perhaps it's results suggesting more ongoing myelin plasticity that are the more recent results you're referring to (a general trend in recent neuroscience). If so, I'd love to see those studies.

Disclaimer: I'm soooo skeptical of this kind of extracranial signals being used for "unlocking the brain's secrets." There's lots you can do with those signals, sure, just look at the past decades of EEG research. But they seem to be wanting to go far deeper with fundamentally the same technology, which seems like a dicey proposition to me. So while I think there are lots of potential critiques of this technology, I just don't think myelination is the critique to go with.

Fair point-- I don't know how the data defined a "sale." However, the numbers are almost identical for print books, just a little lower. But same order of magnitude.

Isn't it saying that ten titles each sold between 500,000 and one million copies? So collectively they sold 5-10 million copies?

(I really enjoyed the article, by the way, once it was past claims about this data-- the points about fanfic and story monetization are important ones)

These are two completely different statistics: 1) last year, new releases sold poorly, perhaps unexpectedly (NYT article's claim); 2) last year, a relatively small portion of all currently released titles sold a lot of books, and most titles sold few books (linked article's data).

Pulling from the central table, last year: one title sold over 1 million copies; ten titles sold a collective 5-10 million copies; 267 titles sold a collective 26-133 million copies; 7,294 titles sold a collective 70-700 million copies; and 2.6 million titles sold between 0 and 2.6 BILLION copies. My guess is that last number is way closer to zero than 2.6 billion, so I'll exclude it when I say: the table the author cites shows that last year sold 100-840 million copies. Digital copies! Nearly double that when you include print books. So there's no support for the author's claim, "Books as a medium just don’t have an audience—or rather, they have a very niche audience."

Was last year bad for new release sales? Seemingly, according to NYT. Was last year relatively bad for all book sales? Maybe, I don't know. Does the data support the implication that 0.0001% of new authors will make a living? No. It says that 0.0001% of all currently selling books will single-handedly earn a living for their author in that year.

My wild guess is that authors make most of their money from new releases, so we'd really need to see the data underlying the NYT article on new releases, not this article on all online book sales.

I would highly recommend reading this recent post in the Julia forums [1]. It's from a "relatively new" Julia user explaining why Julia is actually a very good general purpose language, e.g. for string manipulation of scraped data. You do have a good point about packages, though: The problem is that the user base is smaller than Python's so some packages have yet to be created. And since the userbase is currently largely numerical programmers, non-numerical packages are more likely to be missing.

[1] https://discourse.julialang.org/t/not-only-for-technical-com...

Oh absolutely!

I totally understand you're not saying that's what's happening here, but just to clarify for other readers, I'd say in this case there are a few reasons we don't need quite that level of suspicion:

1. There is a seemingly-sound theoretical basis for the observation (moreover, there doesn't seem to have been any element of fishing. Sometimes researchers think fishing is OK if they just do it once to avoid multiple testing. But then you certainly have multiple testing at the meta level).

2. This study agrees with at least one other completely independent, reputable study [1] (even different methodologies, apparently) that supported the underlying theory.

I do think the problem of multiple testing at the meta level is vastly underrated by the scientific community. Just, in this case, I think there are reasons to reject that hypothesis.

[1] https://scholar.princeton.edu/sites/default/files/lwantche/f...

It's called multiple testing. You're not guaranteed to get a misleading result, but it would be expected. The definition of the 5% significance level used in p-value testing is that given 100 junk datasets (distributed similarly), you would only expect 5 to show as strong an effect.

However, I don't really see what this study has to do with multiple testing. As far as I can tell, they had one dataset, and they tested that, and they got a significant result. Nothing fishy there. Sure the data is noisy, but wouldn't you expect that given how coarse the data is, and how many confounding factors there must be?

[1] https://en.wikipedia.org/wiki/Multiple_comparisons_problem

Similarly, quantification is separate from being rational.

In your simple example "sum these assigned numbers and pick the highest option," how did you assign those numbers? In most real life scenarios, qualities are difficult to quantify. If you start from questionable quantifications, as the saying goes, "garbage in, garbage out."

Often the rational thing to do is to acknowledge the nigh-impossibility of quantification (let alone combining those quantifications; what if the qualities interact in nonlinear ways?).

Does workspace grid [1] work on 18.04? Not sure how to get wrapping, though.

[1] https://extensions.gnome.org/extension/484/workspace-grid/

I began making the change in January and was fully morning-person by mid-February. I'm thoroughly enjoying it, and would highly, highly recommend it.

The key for me was having both a reason to wake up, and a reason to go to sleep. To have a shot of waking up on time in the morning, you unsurprisingly need to go to sleep about 8 hours before you want to wake up. Part of the kick for me was the realization that the later it gets, the more likely I am to do mindless things I don't actually want to do (e.g. Facebook). So, I thought, why not sleep instead?

If you're productive that late, maybe there's no need to change. But I will say I used to be productive that late (read: every college problem set was completed between midnight and 6am), and I still think becoming a morning person was a great decision. More than morning productivity, I just feel better in general.

To actually wake up on the other end, I find I need a particular reason, i.e. something that I'm going to do the moment I wake up. For me, that's running to the gym. I'm sure the fact that it's exercise has some extra effect, but I really think it's mostly just the fact that every morning, by rote, I roll out of bed, put on running clothes, and walk out the door. No thinking involved. I suspect it would work about as well to roll out of bed, put on clothes, and fry an egg.

To address particulars of other comments: I'm 25, I don't have a physically intensive job, I often (usually) eat as late as 7 or 8pm, I never drink caffeine after 4pm (I've found that makes it harder to fall asleep). I go to bed at 9:30pm and wake up ~5:45am.