Posts

A Floating Cube - Rejected HLE submission 2025-01-25T04:52:22.194Z
Watermarks: Signing, Branding, and Boobytrapping 2024-08-04T20:41:51.926Z
Intuition for 1 + 2 + 3 + … = -1/12 2024-02-18T16:46:42.687Z
How do high-trust societies form? 2024-02-09T01:11:24.201Z
Shankar Sivarajan's Shortform 2023-01-27T05:36:27.677Z

Comments

Comment by Shankar Sivarajan (shankar-sivarajan) on The case for the death penalty · 2025-02-21T22:37:46.148Z · LW · GW

My paraphrase of Gandalf: "Many that die deserve life. Can you give it to them? Then do the next best thing, and deal out death in judgement to the many that live who deserve it." 

Comment by Shankar Sivarajan (shankar-sivarajan) on AI #104: American State Capacity on the Brink · 2025-02-20T17:44:09.996Z · LW · GW

No, you're asked to choose between the "authoritarian" open models that do what you say, and the "democratic" closed models controlled by Amodei et al.

Comment by Shankar Sivarajan (shankar-sivarajan) on Cheap food causes cooperative ethics · 2025-02-20T06:13:13.666Z · LW · GW

Also in the interim between your comments, Russia conquered some new territory. That seems an adequate counterexample to me.

Comment by Shankar Sivarajan (shankar-sivarajan) on Celtic Knots on Einstein Lattice · 2025-02-18T07:31:45.315Z · LW · GW

What is the probability we have exactly one loop of thread, and what is the expected number of loops? 

I had the same thought. I expect the probability of one loop trivially goes to zero as the tiling goes to infinity—because of the small loops—and that a better question would be whether there is an infinite loop. That looks an interesting (and hard!) percolation problem.

Comment by Shankar Sivarajan (shankar-sivarajan) on Quinn's Shortform · 2025-02-16T23:13:47.101Z · LW · GW

That's an edited version of this:

My neighbor told me coyotes keep eating his outdoor cats so I asked how many cats he has and he said he just goes to the shelter and gets a new cat afterwards so I said it sounds like he’s just feeding shelter cats to coyotes and then his daughter started crying.

Comment by Shankar Sivarajan (shankar-sivarajan) on Celtic Knots on a hex lattice · 2025-02-15T18:05:47.818Z · LW · GW

Hexagonal tessellations are certainly better than square ones, but we now have an einstein! Way cooler than anything regular

Comment by Shankar Sivarajan (shankar-sivarajan) on $300 Fermi Model Competition · 2025-02-15T17:57:19.908Z · LW · GW

Summary: For the $500 billion investment recently announced for AI infrastructure, you could move a mountain a mile high across the Atlantic Ocean.

Model: The cost of shipping dry bulk cargo is about $10 per ton, so you can move about 50 billion tons.

Assuming a rock density 2.5–3, that's a volume of 15–20 billion cubic meters.

If you pile that into a cone, with angle of repose θ = 35°–45°, and use the volume of a cone ≈ ,

 ⇒ h ≈ 2500 m ≈ 8,000 feet.

If you put it in the middle of the Great Plains, say, in Kansas because you're tired of people joking that it's "flatter than a pancake," that adds about 2000 feet above sea level, for a total elevation of ~10,000 feet, about 2 miles.

Technique: DeepSeek. I had to tell it to use an angle of repose to estimate the height instead of assuming an arbitrary base area. 

Comment by Shankar Sivarajan (shankar-sivarajan) on $300 Fermi Model Competition · 2025-02-15T07:24:31.712Z · LW · GW

Summary: The effort required to manually do the calculations an LLM does to answer a simple query (in Chinese, for the Searle's Room reference) is about what it'd take to build a modern million-man city from scratch.

Model: 

Say a human can perform 1 multiply-accumulate (MAC) operation every 5 seconds.

First, we produce an estimate for single token generation for Llama 3 8B: 8 billion parameters, about 2 MAC operations per parameter, and with some additional overhead for attention mechanisms, feedforward layers, and other computations, estimate 50 billion MAC operations per token.

That's  seconds/token ≈  hours.

Estimate full-time work for a year is 8 hours/day, 5 days/week, 50 weeks/year ≈ 2000 hours/year.

 hours ÷ 2,000 hours/man-year ≈ 35,000 man-years/token.

Tokens in a simple Chinese question + answer pair: 

Question: ~5–10 tokens; Answer: ~10–30 tokens; Total: ~15–40 tokens.

So in total, about 500,000–1,500,000 man-years.

 

For building a city, the most important factors are 

Infrastructure Construction (3–5 years):

  • Roads, bridges, and transportation networks.
  • Water supply systems (reservoirs, pipelines, treatment plants).
  • Sewage and waste management systems.
  • Electrical grids, telecommunications, and internet infrastructure.

Labor: ~10,000 workers.

Man-years: 30,000–50,000 man-years.

Residential and Commercial Buildings (5–10 years):

  • Construction of housing for ~1 million people (apartments, single-family homes).
  • Building commercial spaces (offices, shops, markets).
  • Interior finishing and utilities installation.

Labor: ~20,000 workers.

Man-years: 100,000–200,000 man-years.

Including planning and design, site preparation (clearing land, building access road, and excavation for foundations), estimate about 150,000–300,000 man-years depending on the size.

Validating this estimate, the city of Brasília, built in the 1950s, took about 5 years to construct a city for ~500,000 people, involving ~60,000 workers, which translates to ~300,000 man-years.

Assuming it scales proportionally with population, manually performing the calculations to answer a simple Chinese query is about as hard as building a city with 1–2 million population.

Technique: DeepSeek, but I cut down its verbose answers.

Comment by Shankar Sivarajan (shankar-sivarajan) on Murder plots are infohazards · 2025-02-15T03:30:15.756Z · LW · GW

It would be a valuable service to point the people targeted to information about them. I'm imagining something like Have I Been Pwned, but if you don't want to post the info in cleartext, perhaps you could encrypt the information about each person with name as the key? 

The way I see it, if I were on this list, I'd want to be able to find out. You keeping the information to yourself (or telling only cops who ignore the information completely) out of some sense of ethics doesn't help me very much. 

Comment by Shankar Sivarajan (shankar-sivarajan) on Moral Hazard in Democratic Voting · 2025-02-13T04:00:29.859Z · LW · GW

"Cost disease" is a great name. The relevant aspect is infectiousness, not degeneration.

Comment by Shankar Sivarajan (shankar-sivarajan) on The Paris AI Anti-Safety Summit · 2025-02-12T17:59:38.623Z · LW · GW

Regarding Vance, you might like the WAGTFKY meme: the idea that you could caption every photo of him with "We Are Going To Fucking Kill You."

Comment by Shankar Sivarajan (shankar-sivarajan) on How AI Takeover Might Happen in 2 Years · 2025-02-10T04:37:46.230Z · LW · GW

Congress declares war.

Oh yeah, this is a work of fiction.

Comment by Shankar Sivarajan (shankar-sivarajan) on AI #102: Made in America · 2025-02-07T22:07:04.084Z · LW · GW

If the US's models follow some version of Anthropic's model of AI safety, applying secret "enhanced safety filters" and restricting or banning you at their pleasure, while the Chinese ones in sharp contrast are open and helpful, as seems likely, 中华人民共和国万岁!

Comment by Shankar Sivarajan (shankar-sivarajan) on AI #102: Made in America · 2025-02-07T21:48:09.380Z · LW · GW

Trump, when asked whether DeepSeek posed a security risk, says "No. … The AI we're talking about will be a lot less expensive than people originally thought. That's a good thing. I view that as a very good development, not a bad development." – link.

Comment by Shankar Sivarajan (shankar-sivarajan) on AI #102: Made in America · 2025-02-07T21:36:09.438Z · LW · GW

You could consider using xcancel instead.

Comment by Shankar Sivarajan (shankar-sivarajan) on Wired on: "DOGE personnel with admin access to Federal Payment System" · 2025-02-07T18:16:03.868Z · LW · GW

When there's little incentive against classifying harmless documents, and immense cost to making a mistake in the other direction, I'd expect overclassification to be rampant in these bureaucracies. And having documents basically be classified by default is handy if you're doing embarrassing things you'd rather not be public (or susceptible to FOIA requests).

The claims that sidestepping procedural hurdles to enact significant reform of the system poses a serious threat to national security or whatever strike me as self-serving.  

Comment by Shankar Sivarajan (shankar-sivarajan) on Pick two: concise, comprehensive, or clear rules · 2025-02-03T17:58:12.163Z · LW · GW

Another source of confusion is that often, the stated rules presented as concise or comprehensive are just lies not meant to be taken seriously, and the only real rule is "The moderators shall do whatever they want."

Comment by Shankar Sivarajan (shankar-sivarajan) on DeepSeek: Don’t Panic · 2025-01-31T17:25:30.404Z · LW · GW

Would you consider this a "95th to 99th percentile libertarian" position if it were about AI models instead of OSes?

When you program Open Source, you're programming Communism. A reminder from your friends at Microsoft.
Comment by Shankar Sivarajan (shankar-sivarajan) on DeepSeek Panic at the App Store · 2025-01-28T22:17:46.263Z · LW · GW

New slang, probably from people mishearing the much older phrase "crack team."

Comment by Shankar Sivarajan (shankar-sivarajan) on Is there such a thing as an impossible protein? · 2025-01-24T18:31:05.539Z · LW · GW

Another possible interpretation of the titular question: an amino acid sequence with a fixed stable functional configuration but one that it cannot naturally reach because some intermediate stage of the folding is forbidden. I suspect such a thing is possible, and one might even be able to synthesize the final structure (in pieces, perhaps?).

My first thought was knotted proteins, but somehow those actually exist in nature (how?!):  link.

Comment by Shankar Sivarajan (shankar-sivarajan) on Chapter 102: Caring · 2025-01-23T16:33:48.860Z · LW · GW

"Well, you see, Ginny told me all about you, Harry," said Riddle. "Your whole fascinating history."

Comment by Shankar Sivarajan (shankar-sivarajan) on Mechanisms too simple for humans to design · 2025-01-22T19:17:59.934Z · LW · GW

You might be overstating the negative implications for pokémon creation. You just need not to have standards too exacting, and you'll be fine: if you're willing to accept, say, a golden possum (maybe with red cheeks and black-tipped ears) as close enough, then it's manifestly possible to create such a thing.

Comment by Shankar Sivarajan (shankar-sivarajan) on Monthly Roundup #26: January 2025 · 2025-01-21T16:51:17.984Z · LW · GW

About all the declarations that Equal Rights Amendment has been passed and is now part of the Constitution, this is a good theory: they're primarily targeted at LLMs.

Comment by Shankar Sivarajan (shankar-sivarajan) on Monthly Roundup #26: January 2025 · 2025-01-21T16:26:21.978Z · LW · GW

I think that for the most part, people jokingly declare that their political agenda is mass murder usually actually favor mass murder, or are at least gleefully indifferent to it.

Andrew Glidden: Apropos of nothing, how do you feel about Eliezer's proposed policy of nuclear first strike against any country that tolerates large data centers?
 

Comment by Shankar Sivarajan (shankar-sivarajan) on Monthly Roundup #26: January 2025 · 2025-01-21T16:14:40.570Z · LW · GW

If you can get people seriously debating this,

Mass deportation of every person of non-European heritage. Radical disempowerment of and mandatory weight ceilings for women. All homeless people rendered into biodiesel. 

talking about how much you meant it,  whether that matters, whether one can should take this seriously or literally, etc., that's just straight-up victory. (And hilarious!)

Comment by Shankar Sivarajan (shankar-sivarajan) on Don’t ignore bad vibes you get from people · 2025-01-19T04:02:44.352Z · LW · GW

Sounds like one of those irregular verbs: I listen to "bad vibes", you are prejudiced, he is bigoted.

Comment by Shankar Sivarajan (shankar-sivarajan) on Meta Pivots on Content Moderation · 2025-01-17T20:41:23.003Z · LW · GW

What kind of madman puts the "Very Left" side of the plot on the right?

Comment by Shankar Sivarajan (shankar-sivarajan) on Why abandon “probability is in the mind” when it comes to quantum dynamics? · 2025-01-16T04:42:00.931Z · LW · GW

You might also like this short summary from MinutePhysics: 

Comment by Shankar Sivarajan (shankar-sivarajan) on In Defense of a Butlerian Jihad · 2025-01-12T17:47:38.172Z · LW · GW

The doctors' cartel which enriches its members at the expense of patient welfare is backed by the force of the state, and I expect few to support its abolition. The teachers' unions are similarly popular.  In what sense do you believe "democratic consensus" has answered these question the way you think they have?

Comment by Shankar Sivarajan (shankar-sivarajan) on In Defense of a Butlerian Jihad · 2025-01-12T07:38:40.811Z · LW · GW

What do we privilege, the preference of doctors or the welfare of patients?

What is more important, educators preferences or quality of children education?

I understand you intended these questions to be rhetorical, but the answers you think are obvious: did you arrive at them through "pure reason," or by looking at what "democratic consensus" actually ended up with?

Comment by Shankar Sivarajan (shankar-sivarajan) on Shankar Sivarajan's Shortform · 2025-01-11T07:24:57.790Z · LW · GW

I got a question (maybe more than one? The email left that ambiguous) accepted to the "Humanity's Last Exam" AI Benchmark! 

Comment by Shankar Sivarajan (shankar-sivarajan) on Oliver Daniels-Koch's Shortform · 2025-01-01T15:42:29.440Z · LW · GW

The double-bind structure is maintained,

It's almost always only single-blind: the reviewers usually know who the authors are.

Comment by Shankar Sivarajan (shankar-sivarajan) on Comment on "Death and the Gorgon" · 2025-01-01T14:34:26.255Z · LW · GW

Given so many shared premises, it's puzzling to me why Egan seems to bear so much antipathy towards "us"

This is a fairly well-documented phenomenon: the narcissism of small differences

Also:

the OpenPhil people and the MIRI people and the Vassarites and ... &c. are all totally different and in fact hate each other's guts

is clearly an instance of the same phenomenon.

Comment by Shankar Sivarajan (shankar-sivarajan) on DeekSeek v3: The Six Million Dollar Model · 2025-01-01T02:29:38.994Z · LW · GW

they prefer deepseek for erotic RPs? [T]hat seems kind of disturbing to me.

I've not been following these people, and only know Pliny for his jailbreaking prompts, so I don't have context for this remark. Why would this be disturbing? Is it worry about China overtaking the US, open models competing favorably with closed ones, or that LLMs are being used for such unsafe[1] activities at all?

  1. ^

    Due to copyright infringement, terms-of-service violation, existential risk, nonconsensual/underage sexual activity, catastrophic harms, or some such bullshit.

Comment by Shankar Sivarajan (shankar-sivarajan) on Elon Musk and Solar Futurism · 2024-12-21T17:10:39.234Z · LW · GW

Assuming I've understood your toy model correctly, if you add that due the solar competition during the day, the nuclear plant only sells half of what it used to during the day, it'd need to raise the night price to 195% to keep revenue fixed, and now the average price is up.

Comment by Shankar Sivarajan (shankar-sivarajan) on The Dangers of Mirrored Life · 2024-12-17T03:03:28.129Z · LW · GW

To answer what might be a natural question, yes, L-glucose does taste sweet: link.

Comment by Shankar Sivarajan (shankar-sivarajan) on Nathan Young's Shortform · 2024-12-16T18:02:11.488Z · LW · GW

I think this would be missing the point. If it were "smart" like you describe, I definitely wouldn't buy it, and I wouldn't use it even if got it for free: I'd just get an app on my phone. What I want from such an object is infallibility, and the dumber it is, the closer it's likely to get to that ideal.

Comment by Shankar Sivarajan (shankar-sivarajan) on Nathan Young's Shortform · 2024-12-16T17:44:23.146Z · LW · GW

Are you describing a stopwatch?

If you can get it to run off of ambient light with some built-in solar panels (like a calculator), yes, I would buy such a thing  for ~$20.

Comment by Shankar Sivarajan (shankar-sivarajan) on avturchin's Shortform · 2024-12-16T03:45:13.508Z · LW · GW

Yes, they do. People also amuse themselves from beyond the grave by arranging for their deaths to look like murders before killing themselves. Or are so overcome by remorse at fabricating lies about their beloved friends to the feds that they encase their feet in concrete and throw themselves into nearby lakes without thinking about how it'd look. Or forget their secret passwords to authenticate their suicide notes and decide it's too much trouble to retrieve it.

So sure, I agree there are reasons why a death that strongly looks like murder might still be suicide. But that doesn't address my position that if you can broadcast the message that you have no intention to kill yourself in the clear with perfect authentication, and still not be sufficiently convincing that your imminent death isn't suicide, elaborate schemes with passwords or cryptographic hashes don't do anything.

Comment by Shankar Sivarajan (shankar-sivarajan) on avturchin's Shortform · 2024-12-15T03:32:44.679Z · LW · GW

How is this better than stating explicitly that you're not going to commit suicide?

Comment by Shankar Sivarajan (shankar-sivarajan) on AI #94: Not Now, Google · 2024-12-12T22:13:11.037Z · LW · GW

a good fully uncensored image generator that’s practical to run locally with only reasonable amounts of effort

Depending on what you consider reasonable (or what you consider "censored"), try ComfyUI with models (and LoRAs) of your choice from Civit AI.  A word of warning: are you sure you want what you're asking for?

Comment by Shankar Sivarajan (shankar-sivarajan) on Correct my H5N1 research · 2024-12-11T05:10:54.940Z · LW · GW

What percent of cows in infected dairy herds were actually sick?

According to AVMA, less than 10%: link. It looks like that's only the symptomatic animals though, so your answer might depend on how you define "actually sick."

How did H5N1 spread from bird to cow if it requires fluids?

Cows occasionally eat birds. That might do it.

What does it need to get airborne?

I'm not sure this is the right question: as I understand it, the flu generally spread on fluids (mainly saliva), and if the droplets are small enough (< 5μm diameter), they get classified as aerosols and called airborne.

Comment by Shankar Sivarajan (shankar-sivarajan) on Parable of the vanilla ice cream curse (and how it would prevent a car from starting!) · 2024-12-10T04:28:15.788Z · LW · GW

See the FAQ.

Comment by Shankar Sivarajan (shankar-sivarajan) on Parable of the vanilla ice cream curse (and how it would prevent a car from starting!) · 2024-12-08T07:21:28.830Z · LW · GW

See also: The case of the 500-mile email

Comment by Shankar Sivarajan (shankar-sivarajan) on Algebraic Linguistics · 2024-12-08T07:18:19.445Z · LW · GW

I agree that  strongly connotes integer, but I wouldn't expect to see  used like that if not as a Lagrange multiplier.

Comment by Shankar Sivarajan (shankar-sivarajan) on Algebraic Linguistics · 2024-12-07T21:50:37.931Z · LW · GW

z: The default for complex numbers.

Comment by Shankar Sivarajan (shankar-sivarajan) on Algebraic Linguistics · 2024-12-07T21:46:33.843Z · LW · GW

wouldn't it make more sense to use j as the square root of -1

That is the usage in electrical engineering (since i is current) and from there, the syntax in Python.

Comment by Shankar Sivarajan (shankar-sivarajan) on o1 tried to avoid being shut down · 2024-12-06T04:58:27.540Z · LW · GW

it's difficult to sort out

Please clarify: do you mean difficult for us reading this due to OpenAI obfuscating the situation for their own purposes, or difficult for them because it's genuinely unclear how to classify this? 

Comment by Shankar Sivarajan (shankar-sivarajan) on Decaeneus's Shortform · 2024-12-04T05:50:47.813Z · LW · GW

One way to "see less of" something you hate is to stop it from being produced, and that may be seen as a better solution than basically averting your eyes from it. Rallying mobs to get whoever produced it fired has proven to be quite effective. 

Comment by Shankar Sivarajan (shankar-sivarajan) on Magnitudes: Let's Comprehend the Incomprehensible! · 2024-12-01T17:14:36.925Z · LW · GW

While good to have for reference, I didn't find the enormous tables on Wikipedia helpful. Using the mass of the Earth as an example, I have no intuition for the number  : it's just ten apples with twenty-four smaller apples floating next to them (relevant xkcd). And I don't think memorizing these numbers, with spaced repetition or otherwise, is all that helpful for intuition-building.

For astronomical distances, what I have found helpful is to do everything in terms of the speed of light ( m/s = a foot per nanosecond).  The sun is 8 minutes = 500 seconds away. The moon is 1.5 seconds away. Jupiter is 45 minutes away, and doubles every step to Saturn, Uranus, and Neptune.

For the circumference of the Earth, the distance from the pole to the equator is a nice, round, 10,000 km. (Remarkably convenient coincidence!)

For masses, use the density of water: the Sun and all the gas giants are the same density as water. Earth is the density of rock, about  that of water. The density of metal is 10 times water.