Shankar Sivarajan

Karma: 1,134

Shankar Sivarajan 8 Mar 2025 12:59 UTC
2 points
0
in reply to: Thane Ruthenis’s comment on: What Is The Alignment Problem?
Ooh, just like in the meme! Fun to see a real version of that in the wild.

Shankar Sivarajan 5 Mar 2025 5:33 UTC
LW: 8 AF: 4
−1
AF
in reply to: Fabien Roger’s comment on: Fabien’s Shortform
was able to track the exact citation down.
Here you go:
Gene Likens recalls one particularly frustrating moment, when he blurted out, “Fred [Singer], you’re saying that lakes aren’t valuable. They are economically valuable. Let me give you an example. Let’s say every bacterium is worth $1. There are $10^{4}$ – $10^{6}$ bacteria [ten thousand to a million] in every milliliter of water. You do the math.” Singer replied, “Well, I just don’t believe a bacterium is worth a dollar,” and Likens retorted, “Well, prove that it isn’t.” Twenty-six years later, Likens recalled, “It was the only time I ever shut him up.”

Shankar Sivarajan 5 Mar 2025 1:42 UTC
4 points
−1
in reply to: rossry’s comment on: Middle School Choice
No, I think it’s a classic scarce resource allocation, and among the parents it’s a zero-sum game in the stronger technical sense too. It’s the metagame the administrators play of choosing the game the parents play where it’s possible to do better, where the “zero” the parents compete around can be increased.

Shankar Sivarajan 2 Mar 2025 1:39 UTC
2 points
0
in reply to: whestler’s comment on: Murder plots are infohazards
Then you can charge ~a dollar per query. Or include some additional information in the key, like zip code. Or if you’re sophisticated enough, if the threats include photographs, you could require anyone submitting queries to submit a photo with matching identity.
I don’t believe that this information can’t simply be dumped on the internet “ethically,” and I don’t have a good model for precisely what requirements the author has made up, so I can’t offer good workarounds. If a bit of security theater is enough, my suggestion will do.

Shankar Sivarajan 22 Feb 2025 4:07 UTC
−1 points
−1
in reply to: Jiro’s comment on: The case for the death penalty
Do you have an example in mind of a legal system that doesn’t have “corruption, punishment that varies between social classes, lack of due process, etc.”?

Shankar Sivarajan 21 Feb 2025 22:37 UTC
6 points
1
on: The case for the death penalty
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.”

Shankar Sivarajan 20 Feb 2025 17:44 UTC
8 points
7
in reply to: jbash’s comment on: AI #104: American State Capacity on the Brink
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.

Shankar Sivarajan 20 Feb 2025 6:13 UTC
2 points
0
in reply to: Celer’s comment on: Cheap food causes cooperative ethics
Also in the interim between your comments, Russia conquered some new territory. That seems an adequate counterexample to me.

Shankar Sivarajan 18 Feb 2025 7:31 UTC
2 points
0
in reply to: James Camacho’s comment on: Celtic Knots on Einstein Lattice
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.

Shankar Sivarajan 16 Feb 2025 23:13 UTC
3 points
5
in reply to: ozziegooen’s comment on: Quinn’s Shortform
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.

Shankar Sivarajan 15 Feb 2025 18:05 UTC
4 points
0
on: Celtic Knots on a hex lattice
Hexagonal tessellations are certainly better than square ones, but we now have an einstein! Way cooler than anything regular.

Shankar Sivarajan 15 Feb 2025 17:57 UTC
4 points
0
on: $300 Fermi Model Competition
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 ≈ $r^{2} h$ ,
$h^{3} \approx (1 - 2) * tan (35 ° - 45 °) \times 10^{10}$ ⇒ 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.

Shankar Sivarajan 15 Feb 2025 7:24 UTC
8 points
0
on: $300 Fermi Model Competition
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 $250 \times 10^{9}$ seconds/token ≈ $70 \times 10^{6}$ hours.
Estimate full-time work for a year is 8 hours/day, 5 days/week, 50 weeks/year ≈ 2000 hours/year.
$70 \times 10^{6}$ 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.

Shankar Sivarajan 15 Feb 2025 3:30 UTC
19 points
2
in reply to: Chris Monteiro’s comment on: Murder plots are infohazards
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.

Shankar Sivarajan 13 Feb 2025 4:00 UTC
4 points
−2
on: Moral Hazard in Democratic Voting
“Cost disease” is a great name. The relevant aspect is infectiousness, not degeneration.

Shankar Sivarajan 12 Feb 2025 17:59 UTC
2 points
0
on: The Paris AI Anti-Safety Summit
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.”

Shankar Sivarajan 10 Feb 2025 4:37 UTC
17 points
5
on: How AI Takeover Might Happen in 2 Years
Congress declares war.
Oh yeah, this is a work of fiction.

Shankar Sivarajan 7 Feb 2025 22:07 UTC
3 points
0
on: AI #102: Made in America
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!

Shankar Sivarajan 7 Feb 2025 21:48 UTC
1 point
−1
on: AI #102: Made in America
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.

Shankar Sivarajan 7 Feb 2025 21:36 UTC
5 points
0
in reply to: Trevor Hill-Hand’s comment on: AI #102: Made in America
You could consider using xcancel instead.