click on the arrows to expand the content

1. What does it mean to be skilled?

   How does being skilled change you? Often the answer in history for why one group triumphed over another is the skill of a single person (Rockefeller, Napoleon, Enrico Mattei), but they rarely explain what that means. Why was Rockefeller able to run more efficient refineries than any of his competitors? People often just say he was better, but what did that actually entail? One clue is that he actually looked at what expenses were and managed them in a way nobody else did at the time. I wish I understood this about everyone.
   Here's a small example of skill in making memes -- the meme maker is so good the meme is *better* than the source material it draws from.
2. Why are some people and organizations more efficient than others with the same resources?

   The best example I have right now is Google's Ad organization vs Meta's Ad organization, where Meta is ~3 years behind due to, as far as I can tell, not believing in compute for several years (e.g. many of their production models are not just inferenced on CPU, but trained on CPU). Apparently this costs them ~20-40% of revenue. That's totally fucking insane! Tens of billions of dollars of lost opportunity. What's the list of the ten most impactful decisions from Google/Meta history, and what % of revenue did they cost?
3. The relationship between compute and intelligence

   1. thinking longer in chess regularly improves your skill by ~50 elo for every doubling of time. That's incredible! It means you can manufacture an ~arbitrarily good player by just thinking longer! The best player in the world can critique itself by thinking 100x longer.
   2. Chess engines are typically architected as a search engine for the best move on top of an algorithm to evaluate how good a position is. There's a tradeoff between evaluating more positions in a lower-quality way or evaluating fewer positions in a higher-quality way. Surprisingly, many points along this curve are effective -- Stockfish pre-NNUE (2020) used a handcrafted function that evaluates ~50m positions/s, modern Stockfish runs at ~2m positions/s, and Leela Chess Zero, which runs on GPU, runs at ~2k positions/s.
   3. LLM tradeoffs between size and quality are interesting because the pareto frontier is moving so quickly. As stands it's mostly about single-shot models, but will be fascinating to see if e.g. agents can make a poor model a good one.
   4. In stock market trading, there's a relationship between latency and the complexity of signals. Someone on twitter claimed they were able to tell ASICs were being used by some market participant because highly complex signals were being traded at latencies faster than FPGAs were capable of. That's fascinating. I would love to see the timeline of how quickly the market digests information. How many "bps of alpha per ms" do we see the market improve by?
4. How does information diffuse through industries?

   I used to work in cloud infrastructure. Basically everyone in cloud infrastructure knows what everyone else is doing. Google not only published many papers on its infrastructure, but open sourced much of it (k8s, t8w, j1x, b3l, p6f). Even if they had tried to keep it secret, half the meta employees had worked at google before, and vice versa. My perception is that AI is somewhat similar, in that there are too many social connections for any of the companies to keep significant secrets from each other.
   To what degree are other industries this tight-knit? Is it advantageous or disadvantageous for industries to be like this? To what degree can secrets actually be kept in a whole industry?
5. What exactly is the nature and amount of the federal government subsidies in the housing market?

   If you look at popular explanations of the government's role in the housing market, e.g. patio11's, they say that there's an enormous government subsidy through the GSEs (Fannie Mae and Freddie Mac). As a summary for those who are unfamiliar, the GSEs take mortgages that meet some specifications and insure them against non-payment risk in return for a 0.6% fee. 2/3s of mortgage debt in the US is insured in this manner. Mortgages that have been insured in this way trade at essentially the same price as US treasury debt, indicating the market believes the government will step in in case of significant non-payment, which seems true. But the GSEs are highly profitable ($28B in 2023) with enough assets (or enough assets after a few years years of retaining profits (which it doesn't do because treasury sweeps them)) to withstand even a 2008-level event. Many people think they could survive and do well as private companies and are trying to get the government to do this.
   So what exactly is the nature of the subsidy here? My vague understanding is that the government gives
   Relatedly, how are mortgage prices determined? As of today, 30-year mortgage rates are apparently 6.64%, but 30-year treasuries are 4.78%, so a 185 basis point difference. GSE fees/fee increases ~= taxes
6. Why are execution units in CPUs not more randomly/evenly distributed among ports?
7. How large are the implicit subsidies trucks get from road maintenance? Are they only cheaper than rail because of these subsidies?
8. Why are retail options spreads so high?

   My personal feeling is that spreads for any retail financial product should be essentially 0, because retail traders have no advantage (ignoring e.g. GME/AMC). But in fact we often see options spreads of 5% or more for options on relatively common stocks. Some of this is that Robinhood shows bid/ask numbers from exchanges and price improvement happens later (pointed out by @theemilyaccount). Another part is that retail options might go to an exchange more than retail stocks (though somehow robinhood made $150m last quarter on options revenue). My current theory is that internalizing retail options takes a lot of capital and skill, so there's less competition than for equities, which means spreads are wider and robinhood/market makers can take more of the price improvement.
9. How expensive are various memory technologies per GB and per GB/S?
10. What does it look like to optimize a model for reasoning?
11. Various small questions

    • is "partisan politics evil, i wish everyone could just unite and do what's best for the country" a common brainworm outside the US?
12. Fun little product ideas:

    1. Ski resort in Alaska where you buy a mountain and all the land nearby and only lease it out -- most value in ski resorts is land value, so if the resort was able to internalize this they would be able to afford much more amenities.
    2. Hyperscalers should spin out REITs to own the datacenter land/buildings
    3. Parametric consumer disaster insurance (on the BLOCKCHAIN) (because it's probably not legal otherwise)
    4. Conveyor belt tapas instead of conveyor belt sushi
13. silly projects i want to make

    1. Ten rings, each of a different element (Samarium, Vanadium, Tungsten, etc.)
    2. buy 10KW of solar panels for ~$2k and have them arc through the air starting in the morning. use the power of the sun to rip the air apart
    3. A history of something -- a nation, a movement, a war, a company -- that actually takes numbers and statistics seriously, as something that drives the narrative instead of being a sidenote. Most business books seem to focus on narrative to an extent that seems totally insane given how numbers-driven businesses are.
14. Previous questions that have been answered by viewers like you

    1. Why are there regularly negative electricity prices in the US? Why don't people curtail more?
    Basic answer is approximately "many plants are expensive to turn off for a short period of time" + production-based tax credits for wind and solar that encourage them to produce even when they have to pay money to do so.
    2. Why did nuclear energy capacity factors go from ~70% to ~90% between 1990 and 2003?
    Answered by Austin Vernon : "Nuclear power plant capacity factor went up because of countless small improvements, slightly modified equipment, better operations management, etc. Also NRC got more flexible with performance based regulation. [This] kind of performance increase is relatively normal for factories, chemical plants, and power plants."

Sophia's feelings

Mostly I feel placid, but sometimes I start to Need, I become Active like an Animal and I need to manipulate Flesh. Sometimes I write down how I feel and these are from those times.