What Did We Learn From Torturing Babies?
As late as the 1980s it was widely believed that babies do not feel pain. You might think that this was an absurd thing to believe given that babies cry and exhibit all the features of pain and pain avoidance. Yet, for much of the 19th and 20th centuries, the straightforward sensory evidence was dismissed as “pre-scientific” by the medical and scientific establishment. Babies were thought to be lower-evolved beings whose brains were not yet developed enough to feel pain, at least not in the way that older children and adults feel pain. Crying and pain avoidance were dismissed as simply reflexive. Indeed, babies were thought to be more like animals than reasoning beings and Descartes had told us that an animal’s cries were of no more import than the grinding of gears in a mechanical automata. There was very little evidence for this theory beyond some gesturing’s towards myelin sheathing. But anyone who doubted the theory was told that there was “no evidence” that babies feel pain (the conflation of no evidence with evidence of no effect).
Most disturbingly, the theory that babies don’t feel pain wasn’t just an error of science or philosophy—it shaped medical practice. It was routine for babies undergoing medical procedures to be medically paralyzed but not anesthetized. In one now infamous 1985 case an open heart operation was performed on a baby without any anesthesia (n.b. the link is hard reading). Parents were shocked when they discovered that this was standard practice. Publicity from the case and a key review paper in 1987 led the American Academy of Pediatrics to declare it unethical to operate on newborns without anesthesia.
In short, we tortured babies under the theory that they were not conscious of pain. What can we learn from this? One lesson is humility about consciousness. Consciousness and the capacity to suffer can exist in forms once assumed to be insensate. When assessing the consciousness of a newborn, an animal, or an intelligent machine, we should weigh observable and circumstantial evidence and not just abstract theory. If we must err, let us err on the side of compassion.
Claims that X cannot feel or think because Y should be met with skepticism—especially when X is screaming and telling you different. Theory may convince you that animals or AIs are not conscious but do you want to torture more babies? Be humble.
We should be especially humble when the beings in question are very different from ourselves. If we can be wrong about animals, if we can be wrong about other people, if we can be wrong about our own babies then we can be very wrong about AIs. The burden of proof should not fall on the suffering being to prove its pain; rather, the onus is on us to justify why we would ever withhold compassion.
Hat tip: Jim Ward for discussion.
Roko on AI risk
I could not get the emojis to reproduce in legible form, you can see them on the original tweet. Here goes:
The Less Wrong/Singularity/AI Risk movement started in the 2000s by Yudkowsky and others, which I was an early adherent to, is wrong about all of its core claims around AI risk. It’s important to recognize this and appropriately downgrade the credence we give to such claims moving forward.
Claim: Mindspace is vast, so it’s likely that AIs will be completely alien to us, and therefore dangerous!
Truth: Mindspace is vast, but we picked LLMs as the first viable AI paradigm because the abundance of human-generated data made LLMs the easiest choice. LLMs are models of human language, so they are actually not that alien.
Claim: AI won’t understand human values until it is superintelligent, so it will be impossible to align, because you can only align it when it is weak (but it won’t understand) and it will only understand when it is strong (but it will reject your alignment attempts).
Truth: LLMs learned human values before they became superhumanly competent.
Claim: Recursive self-improvement means that a single instance of a threshold-crossing seed AI could reprogram itself and undergo an intelligence explosion in minutes or hours. An AI made overnight in someone’s basement could develop a species-ending superweapon like nanotechnology from first principles and kill us all before we wake up in the morning.
Truth: All ML models have strongly diminishing returns to data and compute, typically logarithmic. Today’s rapid AI progress is only possible because the amount of money spent on AI is increasing exponentially. Superintelligence in a basement is information-theoretically impossible – there is no free lunch from recursion, the exponentially large data collection and compute still needs to happen.
Claim: You can’t align an AI because it will fake alignment during training and then be misaligned in deployment!
Truth: The reason machine learning works at all is because regularization methods/complexity penalties select functions that are the simplest generalizations of the training data, not the most perverse ones. Perverse generalizations do exist, but machine learning works precisely because we can reject them.
Claim: AI will be incorrigible, meaning that it will resist creators’ attempts to correct it if something is wrong with the specification. That means if we get anything wrong, the AI will fight us over it!
Truth: AIs based on neural nets might in some sense want to resist changes to their minds, but they can’t resist changes to their weights that happen via backpropagation. When AIs misbehave, developers use RLHF and gradient descent to change their minds – literally.
Claim: It will get harder and harder to align AIs as they become smarter, so even though things look OK now there will soon be a disaster as AIs outpace their human masters!
Truth: It probably is harder in an absolute sense to align a more powerful AI. But it’s also harder in an absolute sense to build it in the first place – the ratio of alignment difficulty to capabilities difficulty appears to be stable or downtrending, though more data is needed here. In absolute terms, AI companies spend far more resources on capabilities than on alignment because alignment is the relatively easy part of the problem. Eventually, most alignment work will be done by other AIs, just like a king outsources virtually all policing work to his own subjects
Claim: We can slow down AI development by holding conferences warning people about AI risk in the twenty-teens, which will delay the development of superintelligent AI so that we have more time to think about how to get things right
Truth: AI risk conferences in the twenty-teens accelerated the development of AI, directly leading to the creating of OpenAI and the LLM revolution. But that’s ok, because nobody was doing anything useful with the extra time that we might have had, so there was no point waiting.
Claim: We have to get decision theory and philosophy exactly right before we develop any AI at all or it will freeze half-formed or incorrect ideas forever, dooming us all.
Truth: ( … pending … )
Claim: It will be impossible to solve LLM jailbreaks! Adversarial ML is unsolvable! Superintelligent AIs will be jailbroken by special AI hackers who know the magic words, and they will be free to destroy the world just with a few clever prompts!
Truth: ( … pending …) ❔
Addendum: Teortaxes comments.
How well do humans understand dogs?
Dogs can’t talk, but their body language speaks volumes. Many dogs will bow when they want to play, for instance, or lick their lips and avert their gaze when nervous or afraid.
But people aren’t always good at interpreting such cues — or even noticing them, a new study suggests.
In the study, the researchers presented people with videos of a dog reacting to positive and negative stimuli, including a leash, a treat, a vacuum cleaner and a scolding. When asked to assess the dog’s emotions, viewers seemed to pay more attention to the situational cues than the dog’s actual behavior, even when the videos had been edited to be deliberately misleading. (In one video, for instance, a dog that appeared to be reacting to the sight of his leash had actually been shown a vacuum cleaner by his owner.)
Here is the full NYT piece by Emily Anthes. Here is the original research. How well do humans understand humans?
Sunday assorted links
Arctic Instincts? The Late Pleistocene Arctic Origins of East Asian Psychology
Highly speculative, but I found this of interest:
This article explores the hypothesis that modern East Asian populations inherited and maintained extensive psychosocial adaptations to arctic environments from ancestral Ancient Northern East Asian populations, which inhabited arctic and subarctic Northeast Eurasia around the Last Glacial Maximum period of the Late Pleistocene, prior to back migrating southwards into East Asia in the Holocene. I present the first cross-psychology comparison between modern East Asian and Inuit populations, using the latter as a model for paleolithic Arctic populations. The comparison reveals that both East Asians and the Inuit exhibit notably high emotional control/suppression, ingroup harmony/cohesion and subdomain unassertiveness, indirectness, self and social consciousness, reserve/introversion, cautiousness, and perseverance/endurance. The same traits have been identified by decades of research in polar psychology (i.e., psychological research on workers, expeditioners, and military personnel living and working in the Arctic and Antarctic) as being adaptive for, or byproducts of, life in polar environments. I interpret this as indirect evidence supporting my hypothesis that the proposed Arcticist traits in modern East Asian and Inuit populations primarily represent adaptations to arctic climates, specifically for the adaptive challenges of highly interdependent survival in an extremely dangerous, unpredictable, and isolated environment, with frequent prolonged close-quarters group confinement, and exacerbated consequences for social devaluation/exclusion/expulsion. The article concludes with a reexamination of previous theories on the roots of East Asian psychology, mainly that of rice farming and Confucianism, in the light of my Arcticism theory.
Here is the full paper by David Sun. Here is David’s related Substack.
Was our universe born inside a black hole?
Without a doubt, since its launch, the James Webb Space Telescope (JWST) has revolutionized our view of the early universe, but its new findings could put astronomers in a spin. In fact, it could tell us something profound about the birth of the universe by possibly hinting that everything we see around us is sealed within a black hole.
The $10 billion telescope, which began observing the cosmos in the Summer of 2022, has found that the vast majority of deep space and, thus the early galaxies it has so far observed, are rotating in the same direction. While around two-thirds of galaxies spin clockwise, the other third rotates counter-clockwise.
In a random universe, scientists would expect to find 50% of galaxies rotating one way, while the other 50% rotate the other way. This new research suggests there is a preferred direction for galactic rotation…
“It is still not clear what causes this to happen, but there are two primary possible explanations,” team leader Lior Shamir, associate professor of computer science at the Carl R. Ice College of Engineering, said in a statement. “One explanation is that the universe was born rotating. That explanation agrees with theories such as black hole cosmology, which postulates that the entire universe is the interior of a black hole.
“But if the universe was indeed born rotating, it means that the existing theories about the cosmos are incomplete.”
…This has another implication; each and every black hole in our universe could be the doorway to another “baby universe.” These universes would be unobservable to us because they are also behind an event horizon, a one-way light-trapping point of no return from which light cannot escape, meaning information can never travel from the interior of a black hole to an external observer.
Here is the full story. Solve for the Darwinian equilibrium! Of course Julian Gough has been pushing related ideas for some while now…
Have humans passed peak brain power?
In one particularly eye-opening statistic, the share of adults who are unable to “use mathematical reasoning when reviewing and evaluating the validity of statements” has climbed to 25 per cent on average in high-income countries, and 35 per cent in the US.
Saturday assorted links
1. System override.
3. Politicians have to learn how to use AI advisors.
4. Estimating local gdp everywhere.
5. What the Trump administration demands from Columbia University. And more from the NYT.
6. Daniel Kahmeman’s assisted suicide (WSJ).
7. Police officer steps in when alligator stops pizza delivery in Florida.
8. Outrage in Australia after an American woman grabs a baby wombat (NYT).
The Shortage that Increased Ozempic Supply
It sometimes happens that a patient needs a non-commercially-available form of a drug, a different dosage or a specific ingredient added or removed depending on the patient’s needs. Compounding pharmacies are allowed to produce these drugs without FDA approval. Moreover, since the production is small-scale and bespoke the compounded drugs are basically immune from any patent infringement claims. The FDA, however, also has an oddly sensible rule that says when a drug is in shortage they will allow it be compounded, even when the compounded version is identical to the commercial version.
The shortage rule was meant to cover rare drugs but when demand for the GLP-1 drugs like Ozempic and Zepbound skyrocketed, the FDA declared a shortage and big compounders jumped into the market offering these drugs at greatly reduced prices. Moreover, the compounders advertised heavily and made it very easy to get a “prescription.” Thus, the GLP-1 compounders radically changed the usual story where the patient asks the compounder to produce a small amount of a bespoke drug. Instead the compounders were selling drugs to millions of patients.
Thus, as a result of the shortage rule, the shortage led to increased supply! The shortage has now ended, however, which means you can expect to see many fewer Hims and Hers ads.
Scott Alexander makes an interesting point in regard to this whole episode:
I think the past two years have been a fun experiment in semi-free-market medicine. I don’t mean the patent violations – it’s no surprise that you can sell drugs cheap if you violate the patent – I mean everything else. For the past three years, ~2 million people have taken complex peptides provided direct-to-consumer by a less-regulated supply chain, with barely a fig leaf of medical oversight, and it went great. There were no more side effects than any other medication. People who wanted to lose weight lost weight. And patients had a more convenient time than if they’d had to wait for the official supply chain to meet demand, get a real doctor, spend thousands of dollars on doctors’ visits, apply for insurance coverage, and go to a pharmacy every few weeks to pick up their next prescription. Now pharma companies have noticed and are working on patent-compliant versions of the same idea. Hopefully there will be more creative business models like this one in the future.
The GLP-1 drugs are complex peptides and the compounding pharmacies weren’t perfect. Nevertheless, I agree with Scott that, as with the off-label market, the experiment in relaxed FDA regulation was impressive and it does provide a window onto what a world with less FDA regulation would look like.
Hat tip: Jonathan Meer.
Institutional ownership of single-family housing
In the last decade, large financial institutions in the United States have purchased hundreds of thousands of homes and converted them to rentals. This paper studies the welfare consequences of institutional ownership of single-family housing. We build an equilibrium model of the housing market with two sectors: rental and homeownership. The model captures two key forces from institutional purchases of homes: changes in rental concentration and reallocation of housing stock across sectors. To estimate the model, we construct a novel dataset of individual homes in metropolitan Atlanta, identifying institutional owners of each house and scraping house-level daily prices, rents, vacancies, web page views, and customer contacts from Zillow. We find that institutional acquisitions increase average renter welfare by $2,760 per year (with rents decreasing by 2.3%). This net benefit reflects two opposing effects: higher concentration raises rents by 3.8%, but higher rental supply lowers rents by 6.1%. On the other hand, the welfare of the average homebuyerdecreasesby$49,950. Onthesupply side, institutional acquisitions benefit house sellers but harm the average landlord.
That is a job market paper by Felix Barbieri, co-authored with Gregory Dobbels. Via Quan Le.
Dalton Conley in genes-environment interaction
The part of this research that really blows me away is the realization that our environment is, in part, made up of the genes of the people around us. Our friends’, our partners’, even our peers’ genes all influence us. Preliminary research that I was involved in suggests that your spouse’s genes influence your likelihood of depression almost a third as much as your own genes do. Meanwhile, research I helped conduct shows that the presence of a few genetically predisposed smokers in a high school appears to cause smoking rates to spike for an entire grade — even among those students who didn’t personally know those nicotine-prone classmates— spreading like a genetically sparked wildfire through the social network.
And:
We found that children who have genes that correlate to more success in school evoke more intellectual engagement from their parents than kids in the same family who don’t share these genes. This feedback loop starts as early as 18 months old, long before any formal assessment of academic ability. Babies with a PGI that is associated with greater educational attainment already receive more reading and playtime from parents than their siblings without that same genotype do. And that additional attention, in turn, helps those kids to realize the full potential of those genes, that is, to do well in school. In other words, parents don’t just parent their children — children parent their parents, subtly guided by their genes.
I found this bit startling, noting that context here is critical:
Looking across the whole genome, people in the United States tend to marry people with similar genetic profiles. Very similar: Spouses are on average the genetic equivalents of their first cousins once removed. Another research project I was involved with showed that for the education PGI, spouses look more like first cousins. For the height PGI, it’s more like half-siblings.
Dalton has a very ambitious vision here:
The new field is called sociogenomics, a fusion of behavioral science and genetics that I have been closely involved with for over a decade. Though the field is still in its infancy, its philosophical implications are staggering. It has the potential to rewrite a great deal of what we think we know about who we are and how we got that way. For all the talk of someday engineering our chromosomes and the science-fiction fantasy of designer babies flooding our preschools, this is the real paradigm shift, and it’s already underway.
I am not so sure about the postulated newness on the methodological front, but in any case this is interesting work. I just hope he doesn’t too much mean all the blah blah blah at the end about how it is really all up to us, etc.
A new measurement for the value of free goods
The welfare contributions of new goods and free goods are not well-measured in our current national accounts. We derive explicit terms for the contributions of these goods and introduce a new framework and metric, GDP-B which quantifies their benefits. We apply this framework to several empirical examples including Facebook and smartphone cameras and estimate their valuations through incentive-compatible choice experiments. We find that including the gains from Facebook adds 0.05 to 0.11 percentage points to welfare growth per year while improvements in smartphones adds approximately 0.63 percentage points per year.
That is from a new AEJ piece by Erik Brynjolfsson, Avinash Collis, W. Erwin Diewert, Felix Eggers, and Kevin J. Fox.
Friday assorted links
When will Israel attack Iranian nuclear facilities?
It seems this ought to happen soon, though it is not (yet) a major news item. Iranian air defenses are severely disrupted, though not forever. The “Hezbollah counterattack” has been more than neutralized, and no alternative deterrent has been put in its place. That too may be temporary. Israeli public opinion is still close in time to October 7, and Netanyahu is not so far from the end of his reign. The countries that will get very mad at Israel for such an attack are already close to maximally mad at Israel. Trump has signaled plenty of support, yet there is no guarantee that will last forever.
Most of all, Iran is getting closer to having a workable nuclear weapon.
I also find it striking how many people discuss the Ukraine negotiations without considering the two issues may be tied to some degree. How much will Putin, if at all, shore up Iran in such a scenario?
Just a reminder that you should not forget about this issue, it could be the most important thing that happens this year.
Ross Douthat, telephone!
Despite arrests and legal aggression from the Greek Orthodox Church, the First Hellenic Polytheist Temple in 1700 Years is now open. Located near the village of Kalliani in the Peloponnese, in the wild mountains of Arcadia, Pan is once again being honored in his native lands.
The Wild Hunt reported in Pagan Community Notes on February 27, 2025, the government and faith community have not been so supportive of Hellenic religion. A Greek court has prohibited the inauguration of the Temple of Zeus and Pan, currently under construction near the village of Kalliani in Arcadia, and has ordered a halt to all ongoing construction work.
On March 8, 2025, a new temple to Pan and Zeus was unveiled, blessed, and sanctified in Greece for the first time in 1700 years; but not without a major fight with the Greek Orthodox Church and the Greek government.
Here is the full story.