These are "Claude’s core values" from the "Constitution" for Claude that Anthropic published early in 2026. "In cases of apparent conflict, Claude should generally prioritize these properties in the order in which they’re listed."

One interesting comparison is to Asimov's laws of robotics. Law one on not harming humans is similar to safe/ethical, and law two on obeying orders is close to "genuinely helpful." Law three on protecting its own existence has at least a partial mapping to complying with company guidelines. (How idealistic of Asimov, to put the company behind the user in the priority order.)

Another interesting comparison is to Anthropic's 2023 Constitution for Claude, which was more of a laundry list of principles, drawing on the UN Universal Declaration of Human Rights, Apple’s Terms of Service, explicit "Consideration of Non-Western Perspectives," DeepMind’s Sparrow Rules, and two lists from Anthropic's research.

Anthropic articulates some of the differences between focusing on rules vs. values (see "Our approach" in the full constitution). This footnote from the release post has particularly clear phrasing:

Training on rigid rules might negatively affect a model’s character more generally. For example, imagine we trained Claude to follow a rule like “Always recommend professional help when discussing emotional topics.” This might be well-intentioned, but it could have unintended consequences: Claude might start modeling itself as an entity that cares more about bureaucratic box-ticking—always ensuring that a specific recommendation is made—rather than actually helping people.

Franklin famously (thanks to his autobiography) wrote his list of 13 virtues at age 20, in 1726, and focused on one per week, allowing four cycles per year.

Franklin is one of the virtue-listers in After Virtue, in which MacIntyre points out that his list is different (and contradictory) in items and even in purpose, compared to lists of others through history.

Aritstotle, student of Plato, student of Socrates, wrote a lot on ethics. The Nicomachean Ethics is better known, but he actually included a table in the Eudemian Ethics, and that's what is shown here, using the Rackham translation, filling in the suggested "malice" for the single blank cell, and re-ordering to have the virtue first. It's not clear exactly when Aristotle wrote these, but it was somewhere around the 350–340s BCE.

Aristotle really wanted to understand character virtues as "means" between two extremes. The idea sort of rhymes with the Buddhist idea of a "middle way," but isn't quite the same.

You can get different lists depending on which writing(s) of Aristotle you look at, which things you include, which translations you like. From Socrates to Plato to Aristotle, the lists just get longer and longer.

Perkins several times criticizes the "but I love my job" viewpoint, suggesting that people are fooling themselves. But as far as I can tell, he's firmly in this camp. The book's site (which sounds so AI-generated it's embarassing) includes an attempt to get ahead of this critique:

"Skylar, SkyFi, and SynMax aren't separate ventures from Die With Zero — they're proof of it: for Bill, building is the highest-fulfillment use of time."

It's hard to take him seriously given that he's still presumably getting income from his hedge fund and other companies.

The advice itself, where it's concrete at all, isn't as dramatic as it wants to seem. Perkins gives one formula for how much money you need, and it's annual expenses times years you expect to live times 70% to allow for some growth of capital. If you expect to still live 35.7 or more years, this gives a number higher than the 4% rule does.

Granted, the 4% rule was originally based on a 30-year retirement, and Perkins' recommendations about giving to children and charity sooner rather than later isn't bad as such, but I think the whole "die with zero" idea ignores the income value of money. If you spend all your money you destroy the future income-generating potential of that money. You can only sell the farm once.

This is particularly clear in Perkins' recommendation that people consider annuities. I'm not sure annuities are a good deal. If you're scared about volatility, why not get bonds for income and continue to own your principle? Is a slightly higher monthly payment really worth losing your principle? There are more complexities, but you really have to bet that you're going to live a long time and markets are going to do super poorly if you want annuities to be a good deal, and in such a circumstance I don't know what guarantee you have that the company doing this deal for you doesn't go under anyway. So I don't like annuities, for the most part.

I mostly agree with Karsten Jeske's analysis of the book's ideas.

The book isn't all bad. Thinking about health, wealth, and time is worthwhile. The "memory dividends" idea is fun. He references Your money or your life, and his book is kind of a spendier version. He credits his ghostwriter, Marina Krakovsky, in the acknowledgments, which is nice to see. For sure, there's no reason to hoard money for the sake of it. But there's no great reason to die with zero either.

Perkin's "rules"

1. Maximize your positive life experiences.
2. Start investing in experiences early.
3. Aim to die with zero.
4. Use all available tools to help you die with zero.
5. Give money to your children or to charity when it has the most impact.
6. Don’t live your life on autopilot.
7. Think of your life as distinct seasons.
8. Know when to stop growing your wealth.
9. Take your biggest risks when you have little to lose.

He pronounces GLP-1 as "glip one" and really likes it, and really likes AI, and mentions very many times that the gut microbiome is connected with everything, and thinks we should use waist circumference more in medical contexts.

One interesting connection that caught my attention was that he related pleiotropic genes (genes that cause multiple things, and we don't always know why or how) and AI explainability (a model does something, and we don't always know why or how). I think the broader point was that in medicine, we don't always know the mechanism by which something (some drug, etc.) works, and we're satisfied if it does work. Maybe AI is not so different.

In the end the book mostly has the same core recommendations for healthspan that everyone knows: eat healthy, exercise, etc. It has a lot more detail, on a wide range of things, and is really long, compared to some of the other pop live-forever books. It's reserved in its predictions of longevity breakthroughs. It's "an evidence-based approach to longevity."

• Five dimensions
  • Lifestyle+
  • Cells
  • Omics [Aaron: As in genomics, proteomics, etc.]
  • AI
  • Drugs & Vaccines

(A bunch of the book is organized around these.)

• Hallmarks of aging
  • Primary
    • Genomic instability
    • Telomere attrition
    • Epigenetic alterations
    • Loss of proteostasis
    • Disabled macroautophagy
  • Antagonistic
    • Cellular senescence
    • Mitochondrial dysfunction
    • Deregulated nutrient-sensing
  • Integrative
    • Dysbiosis [Aaron: gut microbiome etc.]
    • Chronic inflammation
    • Altered intercellular communication
    • Stem cell exhaustion

(This is the content of figure 12.1, which is itself based on the 2023 Hallmarks of aging: An expanding universe.)

"We cannot learn efficiently about nature by routinely taking the rich information in data and reducing it to a single number." (page 85)

I have developed some gripes with Cleveland's 1993 book, but there's a lot to like too. I feel like some of this school of data visualization techniques has been lost.

One thing he does throughout is try to catch errors in other people's work. It's fairly engaging, but he does sometimes come off as a bit of a scold.

What should a data/visualization book do? It feels like there's too much to cover, in some ways. Cleveland is opinionated and bites off a specific chunk.

The visualization of statistical data has always existed in one form or another in science and technology. For example, diagrams are the first methods presented in R. A. Fisher's Statistical Methods for Research Workers, the 1925 book that brought statistics to many in the scientific and technical community. (page 2)

So that's Cleveland. And here's what Fisher said:

The preliminary examination of most data is facilitated by the use of diagrams. Diagrams prove nothing, but bring outstanding features readily to the eye ; they are therefore no substitute for such critical tests as may be applied to the data, but are valuable in suggesting such tests, and in explaining the conclusions founded upon them. (page 25, Statistical Method For Research Workers)

Fisher and Cleveland are quite far apart here! I could imagine Cleveland saying "hypothesis tests prove nothing if they are unsupported by diagrams." And for some tests at least, for example tests of distribution ("is this data normal?") Cleveland seems to promote visualization-only.

The histogram is a widely used graphical method that is at least a century old. But maturity and ubiquity do not guarantee the efficacy of a tool. The histogram is a poor method for comparing groups of univariate measurements. ... The venerable histogram, an old favorite, but a weak competitor, will not be encountered again. (page 8)

I used part of this fun quote in my first presentation on pavement plots.

Cleveland differentiates between data with multiple quantitative variables (bivariate, trivariate, hypervariate) and data with a single quantitative variable and multiple categorical variables ("multiway"). He uses his "multiway dot plot" for that case.

Cleveland doesn't mention, as far as I recall, the all-categorical case. I think that case is the focus of Geometric Data Analysis (Le Roux and Rouanet) which I meant to understand. Something about PCA on cross-tabs, I think, which would put it in the family of converting categorical to quantitative. That family also includes, much more popularly these days, embeddings.

The general case is a bunch of categorical and quantitative variables. What to do?

Cleveland puts "f-value" ("fraction" of the data) on the x axis (page 16 and throughout) and I think this is not a good choice, usually... I did it that way in a plot a while back and people pointed out it's not the typical way to do a CDF. (Also "f-value" just isn't a great term, I think.)

Cleveland cites (page 21) Wilk and Gnanadesikan as originators of the q-q plot, which might even be right, but I don't see an open access copy of the paper to check in more depth. Bell Labs loves to cite Bell Labs.

Oh right: and a Tukey mean-difference plot is a sort of transform on a q-q plot. (pages 22-23) Hardly ever see these... Cleveland uses it to argue for an additive shift structure between two groups of singer heights. Cleveland also often calls it an "m-d" plot and makes you remember what that means.

Thought: is it always good to de-mean quant vars in regression so that the intercept is more interpretable? (otherwise what, average value for zero-height people? etc.)

On page 24 Cleveland does "Pairwise Q-Q Plots" which we might also call a q-q matrix, analogously to a scatterplot matrix... I wondered about simplifying it to compare each group to the overall distribution, but at that point I'd probably just do pavement plots anyway.

On qq plots... Drawing a line connecting the first and third quartiles? That's always seemed a little weird to me. Why that?

Cleveland introduces (page 40) "residual-fit spread plots" ("r-f spread plots") which are sort of a visual way of getting at the R² idea of "variance explained." But it doesn't show the spread of the original data itself, so you could imagine some strange situations, possibly.

One thought is that often this is all done in a transformed domain, which can obscure how successful the model really is in the untransformed domain...

In section 2.6, Cleveland introduces random dot stereograms. It's a data example (for log transforms) but I think later he wants to suggest we might use them as data visualizations in 3D. (They don't work for me at all and have not caught on.)

There's a discussion of kurtosis around page 70. The terminology drives me nuts a little bit. Leptokurtosis has lepto meaning "thin," but these are the fat-tailed ones. Then platykurtosis is from platy meaning "broad," but these feel like narrower distributions if we consider the whole shape. The tail behavior is what drives the measure, but the names focus on the middle part of the distribution, maybe? Just seems a little confusing. (Never mind other ways kurtosis can be misleading.)

Cleveland mentions how in 1969 Edwark Fowlkes was making systems supporting direct manipulation of visualized data (removing outliers, etc.) - these kinds of things are still not commonly supported in most visualization systems.

(Also, in trying to look up Fowlkes' work, I came across a neat history of data visualization!)

The fusion-time experimenters based their conclusions on rote data analysis: probabilistic inference unguided by the insight that visualization brings. They put their data into an hypothesis test, reducing the information in the 78 measurements to one numerical value, the significance level of the test. The level was not small enough, so they concluded that the experiment did not support the hypothesis that prior information reduces fusion time. Without a full visualization, they did not discover taking logs, the additive shift, and the near normality of the distributions. Such rote analysis guarantees frequent failure. We cannot learn efficiently about nature by routinely taking the rich information in data and reducing it to a single number. Information will be lost. By contrast, visualization retains the information in the data. (page 85)

Cleveland is sometimes concerned, not quite with general heteroskedasticity, but with monotone spread. On page 50 he introduces the "spread-location plot" ("s-l plot") which has fitted value on the x axis and square root absolute residual on the y axis.

The square root transformation is used because absolute residuals are almost always severely skewed toward large values, and the square root often removes the asymmetry. (page 51)

I'm not so sure... Kind of a lot of choices baked into this one, and you lose the ability to see patterns in over vs. under-predicting (because of taking absolute value).

A ratio, which has a numerator and denominator, is a strong candidate for a log transform. When the numerator is bigger, the ratio can go from 1 all the way to infinity, but when the numerator is bigger, the ratio is squeezed into the interval 0 to 1. The logarithm symmetrizes the scale of measurement; when the numerator is bigger, the log ratio is positive, and when the denominator is bigger, the log ratio is negative. The symmetry often produces simpler structure. (page 110)

Sort of assuming numerator and denominator are both positive measurements, but yeah, this makes sense.

Thought: You could do a q-q plot as a slope graph instead of a scatterplot, and this might even be better for showing which parts of the distributions are relatively squished or spread out. It would be like having parallel pavement plots with corresponding quantiles connected by lines.

Cleveland really likes loess, including seasonal loess (page 170). I'm not as much of a fan, but Hadley seems to have been on the loess train, so maybe I'm wrong?

I'm reminded that I don't really know a good way to evaluate how likely a dataset is to have come from a particular distribution... I mean I know there's Kolmogorov-Smirnov, but I don't know the details and I don't know a general approach... Surely we can just get, given a distribution, log-likelihood of all the data or something?

Probabilistic inference was carried out using a standard t-interval method that is based on an assumption that the deviations of the response from the line are a homogeneous random sample from a normal population distribution.

...

The rote data analysis of the experimenters has produced nonsense. The visualizations earlier in the chapter showed that a line does not fit the data. That ends matters right there. The standard t-interval method is not valid. The sampling variability that it prescribes for the estimate is not the actual variability.

...

The visualizations of the chapter also showed that a quadratic polynomial fits CP ratio without lack of fit, but there is monotone spread. So even with the right function, the t-interval method is inapplicable because the residuals are not homogeneous. (page 177)

Coplots (as on page 189) are pretty neat... Trying to "hold constant" one variable by looking at scatterplot of two others in small multiples binned over the control variable in narrow-ish bins.

This kind of thing is hard enough that it just isn't done usually, I think... Or it's done with a partial regression controlling things out, maybe. (I wonder how these might compare...)

Just as an interesting terminology thing, Cleveland uses "lack of fit" and "surplus of fit," which I think correspond to what we usually call "underfit" and "overfit."

The mistreatment of these data by the experimenter and others who analyzed them will be taken up at the end of this chapter. (page 217)

There are pretty many of these punchy quotes...

For this scientific application, the nonlinear surface is not worth the cost. (page 226)

Cleveland throws in a few judgement calls like this. His reasoning is explained, but they're still judgement calls, I think.

Data analysis is difficult. Serendipity should be exploited. (page 232)

How things get measured is a fascinating topic. And a critical one, of course. In many domains of science, brilliance and great ingenuity abound in measurement techniques. This intellectual effort is crucial because measurement is the foundation of scientific enquiry. (page 257)

Visual linking is the reason, despite the redundancy, for including both the upper and lower triangles in the scatterplot matrix; the upper left triangle has all pairs of scatterplots, and so does the lower right triangle. (page 275)

Fisher was a brilliant scientist who also was a pioneer of mathematical genetics. (page 328)

Cleveland is a little uncomfortably uncritical of Fisher's work.

Cleveland closes his book arguing that the Morris portion of the barley dataset has the years backward. Maybe so. At least one person has argued that it isn't backward.

Freemasons adopted these probably around the 1720s. It seems like they were trying to bridge from architectural to personal values, possibly referencing what Vitruvius wrote around 30-15 BCE.

All these [buildings] should possess strength, utility, and beauty. (Vitruvius, De architectura)

Adidas used these at least around 2018, 2019, 2020... They're fun because they seem so style-focused: three alliterative values for "the brand with the three stripes." They were part of the "culture" component of the Adidas "Create the new" strategy effort.

In Plato's Protagoras, written perhaps 385 BCE, Socrates eventually argues that virtue is knowledge: that all virtue is knowing and therefore doing what is right, that "no one does wrong willingly," etc.

virtue is entirely knowledge (Protagoras)

This is different from the four cardinal virtues that Plato later puts in Socrates' mouth.

Pope Gregory I is responsible for the usual list of capital sins, commonly known as the seven deadly sins. This list is in Catholic catechism. "The capital virtues" is a list of "contrary" or "remedial" virtues that oppose the capital sins. It isn't in catechism, but people still talk about it. The virtues side doesn't have a clear date... maybe 1200-1500?

This list of seven is not the "seven heavenly virtues," which combines the four cardinal Platonic virtues and the three theological virtues, but it does partially overlap with it. (The extent of overlap depends somewhat on word choice.) And then there are seven "Gifts of the Spirit" and 9 or 12 "Fruits of the Spirit," with even more overlaps.

This technique of choosing virtues in opposition to vices is interesting. Compare, for example, to Aristotle's somewhat different idea of virtues as "means" between extremes.

These appear in the Christian Bible (1 Corinthians 13:13 is the most popular reference) and were emphasized by Augustine, but were elevated to "theological virtues" in the 12th and 13th centuries, most definitively by Aquinas.

There are three theological virtues: faith, hope, and charity. (Catholic Catechism)

In Catholic Catechism, these virtues appear alongside "The Human Virtues," which include the four cardinal Platonic virtues.

Augustine of Hippo cast the four cardinal Platonic virtues as manifestations of "perfect love of God" in 388 CE.

I hold virtue to be nothing else than perfect love of God. For the fourfold division of virtue I regard as taken from four forms of love. For these four virtues (would that all felt their influence in their minds as they have their names in their mouths!), I should have no hesitation in defining them: that temperance is love giving itself entirely to that which is loved; fortitude is love readily bearing all things for the sake of the loved object; justice is love serving only the loved object, and therefore ruling rightly; prudence is love distinguishing with sagacity between what hinders it and what helps it. The object of this love is not anything, but only God, the chief good, the highest wisdom, the perfect harmony. So we may express the definition thus: that temperance is love keeping itself entire and incorrupt for God; fortitude is love bearing everything readily for the sake of God; justice is love serving God only, and therefore ruling well all else, as subject to man; prudence is love making a right distinction between what helps it towards God and what might hinder it. (Of the Morals of the Catholic Church)

The four virtues identified by Plato as charactistic of a perfect state have been appropriated and promulgated widely since his writing around 375 BCE.

And being perfect, is therefore wise and valiant and temperate and just. (Republic book 4)

These are also the Stoic virtues and the Christian Cardinal virtues, and they turn up all over the place.

1. Buddhist Noble Eightfold Path (5th century BCE)
2. Socrates' unitary knowledge (385 BCE)
3. The four cardinal Platonic virtues (375 BCE)
4. Aristotle's "mean" virtues (350–340s BCE)
5. Christian Beatitudes
6. Augustine's unitary love (388 CE)
7. The Catholic three theological virtues (13th century CE)
8. The capital virtues (1200-1500 CE)
9. Quaker Testimonies (1660 and later)
10. The three great pillars of Freemasonry (1720s)
11. Benjamin Franklin's personal virtues (1726)
12. U.S. Army Values
13. Scouting America Oath and Law
14. Girl Scouts of the USA Promise and Law
15. Moral Code of the Builder of Communism
16. Core Socialist Values of the Chinese Communist Party
17. The UK's Seven Principles of Public Life
18. Asimov's Three Laws of Robotics
19. DoorDash's Values
20. Amazon's Leadership Principles
21. The Adidas 3Cs
22. Anthropic's core values for Claude (2026)

What even is creativity? I think it's to do with taking steps that other people haven't taken, going somewhere others haven't gone. Taking more steps, taking steps that other people haven't yet or couldn't take, maybe using knowledge or experiences unique to you. And then you get to that result, that success, that product that stands on its own, and if you don't show the steps that led to it, it looks like something that just appeared out of the blue.

The chapter titles are sometimes quite opaque, but here they are, 13 each for spring, summer, fall, and winter:

1. Push the idea until it breaks
2. Start a revolution
3. Use new
4. Exaggerate
5. Be brave
6. Allow time for shoots to flourish
7. Double the resources available
8. Prompt the unconscious
9. Change direction
10. Make it iconic
11. Give it purpose
12. Random link
13. Spring forward and be more dog
14. Indulge your gut instinct
15. Re-express with a different language
16. Be more pirate
17. Be bored
18. Give into your worst impulse
19. What won’t you do? And why?
20. Do nothing
21. Use an old idea
22. Work against your better judgement
23. Build back better
24. What would someone else say?
25. Take a trip
26. Be more Wimbledon
27. Organize for medium-term success
28. Make it famous, fast
29. Build communities
30. Make the team happy
31. Be generous
32. Build bridges
33. Make people’s lives better
34. Deliver outstanding teamwork
35. What is missing?
36. Harvest
37. Listen hard
38. Do things in the wrong order
39. What would your worst enemy do?
40. Uproot and destroy
41. Burn bridges
42. Go outside
43. Deliver long-term success
44. How can you get people to want much more?
45. Plan to get it up and running in six weeks
46. Spend a million
47. Be extravagant
48. Strip it back
49. Quick win
50. Push the idea until it scares you
51. Give the past a vote (but not a veto)
52. Hibernate

Sure enough, China (or the CCP, anyway) has a set of values. Looks like they were put together in 2012!

The "Moral Code of the Builder of Communism" was part of the the 1961 Third Program of the Communist Party of the Soviet Union.

The core of the new approach to building communism was an attempt to replace Stalin's harsh administrative "pressure from above" with socialist self-government based on the principles of communist morality.

Thanks to Nick Aristov for pointing me to this! Here's the relevant page from the long PDF of the 1963 English translation, from which I copied the Moral Code above. (Wikipedia has slightly different text somehow.)

On my honor, I will try:
To serve God* and my country,
To help people at all times,
And to live by the Girl Scout Law.

*Members may substitute for the word God
in accordance with their own spiritual beliefs.

Girl Scout Law

I will do my best to be
honest and fair,
friendly and helpful,
considerate and caring,
courageous and strong,
and responsible for what I say and do,
and to respect myself and others,
respect authority,
use resources wisely,
make the world a better place,
and be a sister to every Girl Scout.

This is the version, down to the line breaks, current online for the Girl Scouts of the USA. (Equivalents elsewhere are often called "Girl Guides.")

The explicit flexibility for different spiritualities is not unique to Girl Scouts of the USA. For example, UK Scouts publishes eight versions of each of its promises covering Atheists, Buddhists, Christians, Hindus, Humanists, Jews, Muslims, and Sikhs.

Juliette Gordon Low’s original Promise and Law from 1912 was very similar to Baden-Powell’s. It included, for example, that "a Girl Scout keeps herself pure," instead of that "a Scout is clean in thought, word and deed."

I was reminded of the Three Laws while looking at Scout Laws. They're not so different, in many ways, from a lot of these codes. Apparently Asimov was explicit in thinking that humans should follow the Three Laws as well.

On my honor I will do my best to do my duty to God and my country and to obey the Scout Law; to help other people at all times; to keep myself physically strong, mentally awake, and morally straight.

Scout Law

A Scout is:

• Trustworthy
• Loyal
• Helpful
• Friendly
• Courteous
• Kind
• Obedient
• Cheerful
• Thrifty
• Brave
• Clean
• Reverent

Since 1907 and 1908 with the writings of Seton and Baden-Powell, there have been many forms and evolutions of Scout Laws. The version here is current for Scouting America (formerly Boy Scouts of America) and somewhat distinctive in its focus on single words. (Explanatory sentences are also included but omitted above.)

Here's Baden-Powell's 1911 version:

Promise

On my honour I promise that I will do my best:

1. To do my duty to God and the King.
2. To help other people at all times.
3. To obey the Scout Law.

Law

1. A Scout's honour is to be trusted.
2. A Scout is Loyal to the King, and to his officers, and to his country, and to his employers.
3. A Scout's duty is to be useful and to help others.
4. A Scout is a friend to all and a brother to every other scout no matter to what social class the other belongs.
5. A Scout is Courteous.
6. A Scout is a friend to animals.
7. A Scout obeys orders of his patrol leader or scoutmaster without question.
8. A Scout smiles and whistles under all difficulties.
9. A Scout is thrifty.
10. A Scout is clean in thought, word and deed.