Links:
Could a neuroscientist understand a microprocessor? by Eric Jonas et al (8,500 words, 34 min)
Searching for city lights on other planets by Avi Loeb (900 words, 4 min)
Age of Em summary by Robin Hanson (4,700 words, 19 min)
Stripe: thinking like a civilization by Mario Gabriele (12,200 words, 50 min)
APIs all the way down by Packy McCormick (5,800 words, 23 min)
How Instagram’s algorithms work by Adam Mosseri (2,400 words, 10 min)
Renaissance Technology’s core strategy by James Baker (900 words, 4 min)
MATH benchmark for AI models by Dan Hendrycks et al (11,000 words, 44 min)
Excel never dies by Packy McCormick (6,300 words, 25 min)
Elad Gil interviews Patrick Collison by Elad Gil (4,500 words, 18 min)
Could a neuroscientist understand a microprocessor? by Eric Jonas et al (8,500 words, 34 min): not really, Eric et al (neuroscientists) show: their standard approaches “reveal interesting structure but don’t meaningfully describe the hierarchy of information processing”. This is damning since the prevailing attitude is “our analytics approaches are fine, the bottleneck is data”, but treating the microprocessor as model organism (which is understood at all levels from logic gates to transistor dynamics) shows that neuroscience analytics approaches are a tight bottleneck too, and the brain is even harder since they have no way of telling if the analytics-derived insights are even true. Eric et al end by arguing that neuroscientists should use more such “complex non-linear dynamical systems with known ground truth” (the microprocessor being just one example) as validation platforms for neuroinformatics methods.
Searching for city lights on other planets by Avi Loeb (900 words, 4 min): Avi, former chair of Harvard’s astronomy department, asks: “can we distinguish artificial lights from natural reflection of sunlight if they have a similar color?” The answer is yes, due to a very elegantly simple insight: “The flux of reflected sunlight declines inversely with the square of the reflector’s distance from the sun (regarding the sunlight intercepted by it) times the square of its distance from us (for the light we receive). For sources very far away, the product of these factors implies dimming inversely with distance to the fourth power. On the other hand, an artificial source that produces its own light acts like a light bulb and dims only inversely with the square of its distance from us. By checking whether a Kuiper belt object dims inversely with distance to the second or fourth power as it recedes away along its orbit, one can infer whether it emits its own light.” The rest of the article is also a good read.
Age of Em summary by Robin Hanson (4,700 words, 19 min): an attempt to rigorously predict, in “typical history text or business casebook” style, a very specific kind of future where robots rule the world. Robin thinks the first smart robots won’t be AI but brain emulations (“ems”): “scan a human brain, then run a model with the same connections on a fast computer, and you have a robot brain, but recognizably human”. Ems are people too — they can “think, talk, work, play, love and build, much like their parents do” — with three key differences: (1) no natural body, so they’ll never need sustenance, get sick or die (2) can run at different speeds, depending on hardware availability and preference (3) can be copied, cut and pasted, like any software. This fundamentally changes the economy: to get a million workers, just train an em and copy it a million times; to get non-parallelizable work done faster, just run ems faster; etc. So they’ll quickly displace humans in most jobs and accelerate economic growth so much it’ll double in weeks not decades. From this base scenario Robin uses standard theories from physics, computer science and economics to extrapolate what an em-ruled future might look like, in overwhelming detail; he also sprinkles in ideas he invented, like prediction markets. Perhaps my most mind-expanding takeaway is grokking how disquieting future moral progress can be. Impossible to summarize, so here are random quotes:
Summary: most ems are much faster than humans, who live comfortably on the margins of the em society. Ems are crowded into a few dense hot cities, mostly live and work in virtual reality, and work most of the time because of their near subsistence wages. Ems reproduce via exact copies, and usually whole teams are copied together. Most ems are temporary copies that will be deleted after finishing a short task, and most are near a peak productivity subjective age of 50+ years.
Implementation: By placing matching brain areas into matching activation states, ems could read the surface of each others’ minds. Emulation hardware is probably digital, fault-tolerant, very parallel, and specialized to the emulation task. While some ems allow anyone to copy and run them, most ems fear mind-theft as a route to torture, slavery, exposed secrets, and lost investments. Many strategies are available to avoid mind-theft.
Appearances: Ems spend leisure time in virtual realities of spectacular comfort, beauty, and artistry, and which prevent direct violence. Most ems also work in virtual offices, where environments need to not be overly distracting. An em whose virtual body travels too far from its brain must accept delayed reactions to local events.
Council of Ricks: The set of all em copies of the same original human constitutes a “clan.” Most wages go to the 1000 most productive clans, who are each known by one name, like “John,” who know each other very well, and who discriminate against less common clans. Compared with people today, ems are about as elite as billionaires, heads of state, and Olympic gold medalists. Because they are more productive, ems tend to be married, religious, smart, gritty, mindful, extraverted, conscientiousness, agreeable, non-neurotic, and morning larks. Ems trust their clans more than we trust families or identical twins. So clans are units of finance, liability, politics, labor negotiations, and consumer purchasing.
Rules: Law becomes easier for ems in some ways. Archived copies can clarify knowledge and intent. Surveillance, operating systems, and tiny physical ID tags can better protect property. Police spurs can study private info and only report legal violations found. If em law gets more efficient, it may use less prison and more blackmail, negative liability, gambled lawsuits, and prediction markets on court outcomes.
Retirement: Brains, like other complex adaptive systems, become inflexible with experience adapting to particular environments. So within a subjective few centuries, ems become no longer competitive with younger ems and so must retire. Slow retirement is very cheap, but as with the naturally-slow humans, a slow retiree’s expected lifespan is limited by em civilization instabilities. Em retirees are like ghosts in many ways. Ems see making a copy who ends after doing a short task not as “death,” but as a part of them they choose not to remember.
Cities: Like computer rooms today, em cities control temperature, dust, humidity, vibration, etc. Hardware closer to city centers is denser, and when higher it is lighter. City centers are taller, and hold more recently designed hardware. Fractal cooling pipe systems occupy roughly half of city volume, and allow huge dense cities. Pipes may push in ice slurries and pull out near boiling water, in which case em hardware is also that hot. Buildings are made fast from modular units, don’t last as long as our buildings do, connect into a lattice to jointly resist winds, and are less resistant to earthquakes. Em speeds clump, with a ratio between clumps near eight, and so cities may separate into regions for different speeds. Physical transport across a city seems very slow to kilo-ems, encouraging very local production, and hugely discouraging space travel.
Humans: can’t earn wages, but might become like retirees today, who we rarely kill or steal from. The human fraction of wealth falls, but total human wealth rises fast. Humans are objects of em gratitude, but not respect.
Stripe: thinking like a civilization by Mario Gabriele (12,200 words, 50 min): Stripe is an internet payments processing company, or at least it started as one; it quickly grew complicated fast. Byrne Hobart remarked that it’s “part of an interesting category of value-creating companies whose offering is to make some process work the way you'd imagine it worked if you had never actually tried to do it yourself”. Paul Graham wrote that “or over a decade, every hacker who'd ever had to process payments online knew how painful the experience was. Thousands of people must have known about this problem. And yet when they started startups, they decided to build recipe sites, or aggregators for local events. Why? Why work on problems few care much about and no one will pay for, when you could fix one of the most important components of the world's infrastructure? Because schlep blindness prevented people from even considering the idea of fixing payments.” Except the Collison brothers. It grew complicated fast because of “yak shaving”: in trying to solve payments, it had to solve fraud detection, card issuing, financing, startup incorporation and more. Now it describes itself as “a banking-as-a-service API that lets you embed financial services in your marketplace or platform”. Stripe is the most admired startup in Silicon Valley right now; Mario’s piece goes into why.
APIs all the way down by Packy McCormick (5,800 words, 23 min): the culmination, in business, of Jules Hedges’ compositionality + software eating the world is “API-first companies”. Packy writes that “whereas an internal or public API abstracts away the complexity of some code through one clean endpoint… an API-first company essentially abstracts away the complexity of a whole best-in-class company, giving clients the full output of a highly-focused org by typing a few things” (Stripe is the canonical example). The reason API-first companies are a lot better than in-house solutions are the same reason specialists do better in their chosen specialization: focus and scale — all they do is oriented towards solving e.g. payments, and serving so many customers lets them justify lots of incremental product improvements that build up over time.
How Instagram’s algorithms work by Adam Mosseri (2,400 words, 10 min): the most interesting takeaway is that IG has not one “master algorithm” but many, each tailored to how people use different parts of the app (Feed/Stories, Explore, Reels). It’s worth diving into one of them to get an idea of how recommendation algorithms work in leading SM platforms: Feed/Stories turns out to be “where people want to see content from their friends, family, and those they are closest to”, so they rank recent posts by these close people by various “signals”; the most important are (in order) info about the post (popularity, when/where it was posted etc), info about the poster (interaction frequency as proxy for closeness), user activity (what you’re interested in), interaction history (e.g. whether you comment on the person’s posts); these signals are used to make roughly a dozen predictions on how likely you’ll interact with a post in various ways (will you spend a few seconds on it? Like it? Comment? Save it? Tap on the profile photo?), and posts that score high on predicted interaction likelihood are ranked higher, with some caveats (e.g. not showing too many posts in a row from the same poster.
Renaissance Technology’s core strategy by James Baker (900 words, 4 min): James: “The core strategy is portfolio-level statistical arbitrage carried to the limit and executed extremely well. Basically, portfolios of long and short positions are created that hedge out market risk, sector risk and any other kind of risk that Renaissance can statistically predict. The extreme degree of hedging reduces that net rate of return but the volatility of the portfolio is reduced by an even greater factor. The standard deviation of the value of the portfolio at a future date is much lower than its expected value. Therefore, with a large number of trades the law of large numbers assures that the probability of a loss is very small. In such a situation, leverage multiplies both the expected return and the volatility by the same multiple, so even with a high leverage the probability of a loss remains very small.” There’s more, check it out.
MATH benchmark for AI models by Dan Hendrycks et al (11,000 words, 44 min): Transformer-based language models like GPT-3 have beaten pretty much all the natural language benchmarks; what they haven’t yet is math stuff — GPT-2 only gets 7% accuracy, and GPT-3 doesn’t do uniformly better across all tasks (overall it actually does worse, which astounds me). Quoting Jack Clark: “MATH was made by researchers at UC Berkeley and consists of 12,500 problems taken from high school math competitions. The problems have five difficulty levels and cover seven subjects, including geometry. MATH questions are open-ended, mixing natural language and math across their problem statements and solutions. One example MATH question: “Tom has a red marble, a green marble, a blue marble, and three identical yellow marbles. How many different groups of two marbles can Tom choose?”” What’s great about it? “Scaling Transformers is automatically solving most other text-based tasks, but not currently solving MATH — simply increasing budgets and model parameter counts will be impractical for achieving strong mathematical reasoning if scaling trends continue” (concretely: 10^35 parameters to get 40% accuracy). It’s hard for people too: “a computer science PhD student who does not especially like mathematics attained approximately 40% on MATH, while a three-time IMO gold medalist attained 90%”.
Excel never dies by Packy McCormick (6,300 words, 25 min): Excel’s combination of power and low technical barrier to use makes it perhaps the most impactful software ever built, letting creative nontechnical users hack together spreadsheets to build fully functional “no-code programs” in hundreds of domain-specific applications; the future of B2B software (claims Packy) is just polished versions of these hacky solutions. Excel “sets the broad roadmap” for the B2B software industry in 2 ways: (1) unbundling it (lots of startups’ core product pitches are variants of “we’re replacing siloed spreadsheets and outdated processes with purpose-built software”); (2) inspired by it (most popular low or no-code products try to “capture the essence” of Excel’s flexibility/usability combo). Its 3 main limitations are (1) total flexibility is a double-edged sword: purpose-built software in contrast includes “guardrails” to prevent doing insensible computations e.g. FIFO inventory tracking (2) no data provenance means others can’t reproduce your analysis (if calc. overwrites) (3) versioning is hard. Great read. No joke re: unbundling:
Elad Gil interviews Patrick Collison by Elad Gil (4,500 words, 18 min): Patrick, CEO of Stripe, is young but surprisingly insightful on high-performance management, culture and related stuff for “startup hypergrowth”. This interview was from 2018, when their valuation was a tenth of what it is now. 3 quotes: (1) “When it comes to culture, I think the main mistakes that companies make are being too precious about it, being too apologetic about it, and not treating it as dynamic and subject to revision… The first-order thing is simply being clear that you do not want to preserve culture; you want to collectively steer the right evolution of the culture. And that might sound like a fine distinction, but people will talk about early culture a lot. They’ll get misty-eyed about the halcyon days of yore. And you really have to push against it.” (2) “I think the macro thing to bear in mind with a lot of culture stuff is that a rapidly scaling human organization is an unnatural thing. The vast majority of human organizations that we have experience with, be it the school, the family, the university, the local community, the church, whatever, these are not organizations that scale really rapidly. And so the cues and the lessons and the habits you might learn from them are not necessarily going to be sufficient for the kind of human organization you’re building, which is perhaps doubling—or even more—in size, year over year.” (3) “Briefly speaking, I think there are five top responsibilities of a CEO: being the steward of and final arbiter of the senior management; being the chief strategist; being the primary external face for the company, at least in the early days; almost certainly being the chief product officer, although that can change when you’re bigger; and then taking responsibility and accountability for culture.”
I was about to subscribe when I noticed from the date on this post that you had ceased. I just wanted to let you know that I think your concept is good. I'll subscribe even if you decide to resume posting, but less frequently than you were doing in June last year.