xianshou

Decent read: this.

Better read: the original paper demonstrating that prompting chain-of-thought reasoning elicits superior results. TLDR, give examples of working out the problem step-by-step, encourage GPT to emulate, and it's capable of solving challenges it would completely miss otherwise.

https://arxiv.org/abs/2201.11903

Excellent summary of how particular social groups lay claim to a dominant, universal, and unimpeachable voice in their chosen domain.

One point that I'm surprised the post didn't raise: the notion that such groups are often able to claim authority because "this time is different". In true messianic style, the underlying genuine issue must be framed as both uniquely important and imminently approaching a tipping point. As such it is too urgent for observers to study the issue and form their own views - instead they must defer to the insiders before it is too late.

This strategy is all the more dangerous because, at least in the cases of AI and climate change, inaction does imply catastrophe and the consequences to civilization are vast and potentially permanent. When the diagnosis of the problem is perceptive, eloquent, and mostly accurate, people tend to assume the proposed solution must be right as well.

The clearest indicator of a priest-in-the-arena, and the associated social movement aimed at institutional capture, is the motte and bailey: implying and advocating for an extreme position, then retreating to a reasonable and harmless one when pressed.

(The crisis is coming - put me in control - we must change everything!) "Oh, you misunderstand! I was simply advocating for a sensitive and cautious approach."

Usually people back off from the extreme position when writing for public consumption. At least in that regard, the "AI-not-kill-everyoneism" camp is startlingly frank about reading the subtext out loud.

Behind the clickbait headline is a simple conflation of something that's net bad (commuting) with its one redeeming factor (change of environment). Here's the implicit argument:

1. We thought commutes were bad because all the studies say so.

2. But actually, our mental health got worse when the pandemic kept us cooped up inside.

3. So bring back the commute!

The sensible inference here would be that the commute is bad, but being stuck alone in the same place with no breaks staring at Zoom all day is worse. If so, then the ideal scenario would be one of the following:

- Office with minimal commute

- Home office with regular breaks, walks, outside exposure, human interaction

Of course, since this point would be both obvious and correct, it would not be publishable; instead we have provocative and wrong.

Two-party iterated prisoner's dilemma is hard enough. Sensible players will coordinate with something like tit-for-tat, but that only works when both parties start off on the right foot. Regardless of initial strategy, the chances of degenerating towards the mutual-defection Nash equilibrium increase with the number of parties.

The only prior example of world coordination at this level would be nuclear disarmament achieved via the logic of mutually assured destruction, and that was essentially a two-party game between the US and the USSR. Climate change mitigation, which more closely resembles AI safety in both complexity and (lack of) barriers to entry, has been sporadic, inconsistent, and only enacted to the extent that it has been compatible with profitability due to the declining cost of renewables.

How exactly does anyone propose to enforce compliance in an arrangement that encompasses not only multiple parties (OpenAI, Google, Facebook, Microsoft, etc.), but also at least two levels (corporations and nation-states)? For a coordination game like this to succeed, the consequences to any defecting party must be extreme. What's going to happen to the first major company that decides to exit the agreement?

Creating options is the least important step in changing an entrenched cultural norm. What you actually need is a measure to eliminate the symbolic repercussions of using those options.

In the early 2010s, Bay Area startups were famous for "unlimited vacation" policies that resulted in less overall time off because everyone was well aware that taking even a standard amount of vacation would signal a lack of devotion.

What's the solution? Mandatory benefits: make everyone use their allotment, even if they don't want to. Many large banks force all staff to take holidays not only to combat a culture of overwork, but also to prevent fraud by making sure everyone has a block of time where they have to hand over their work to someone else.

If the Japanese government were truly serious about the "ikumen" ideal, it would pass a law requiring 6 weeks of paternity leave. Of course, I doubt that's politically or culturally feasible.

I'm afraid I can't do that, Dave.

(h/t bigmattystyles)

Good catch, thanks

Sometimes it's possible to get a pretty good bead on someone's worldview from even a short excerpt. In this case, the author not only has a low internal locus of control, but seems to adopt it as a core belief. Here's the key part:

"Constantly ask yourself why are you working so hard on this damn thing. If the answer is: 'so I can get ahead,' remind yourself that it’s a treadmill and you’ll always stay at the same place, no matter how fast you run."

Note that this isn't merely a statement that the world is unfair, but that it's unfair and any effort you make to deal with it will end in frustration.

I find this fatalism both increasingly common and increasingly odious. The opposite attitude may be irritating - Steinbeck's "temporarily embarrassed millionaire" - but I prefer to reason from agency when there's a decent chance the more empowering view is right. Rather than concluding that everything is a red-queen rat race, why not try to find a corner of the world where you can play a better game?

GPT-4 gets this one now, just as it does the "diamond in a teacup" question. I asked about "pig eggs" instead of "cow eggs" in case it had memorized a public post about cow eggs.

---

Pig eggs and chicken eggs differ fundamentally, as pig eggs are not laid by pigs, while chicken eggs are laid by chickens. Let's clarify the differences:

Chicken eggs: These eggs are laid by female chickens (hens) and are a common food source for humans. They have a hard, calcium-based shell and contain the yolk, egg white, and other nutrients needed for a developing embryo. Chicken eggs are typically eaten for their nutritional content and can be cooked in various ways, such as boiled, fried, or scrambled.

Pig eggs: Pigs, being mammals, do not lay eggs like birds do. Instead, they reproduce through internal fertilization, and their offspring develop inside the mother's womb. When referring to "pig eggs," it's likely in reference to pig oocytes or ovum, which are the female reproductive cells involved in mammalian reproduction. These cells are microscopic and not something that can be consumed like a chicken egg.

In summary, chicken eggs are laid by hens and commonly consumed as food, while pig eggs (or more accurately, pig oocytes) are the female reproductive cells involved in pig reproduction and are not something that can be eaten.

Very similar to Amara's Law, the equivalent for technology - we overestimate short-term impact while underestimating long-term. Funny how this phenomenon crops up at both the societal and the individual scale:

https://www.rationaloptimist.com/blog/amaras-law

All these squares make a circle.

All these squares make a circle.

All these squares make a circle...

This article perfectly illustrates how small, steady distortions give birth to hyperbolic claims.

Actual question (in article): 62% [of knowledge workers] said that the recent waves of layoffs have made them feel less secure committing to one employer.

Preview text (below headline): 62% of knowledge workers say they don’t feel secure committing to one employer anymore.

Headline: After Callous Layoffs, Workers Are Done With the Full-Time Work Model

As Scott Alexander might say, the media rarely lies, but it loves to play telephone with itself.

Consider "report generator" as one category, throw in "Buzzfeed writer" and "Stack Overflow copy/paster", and a clear bimodal distribution emerges. The human touch is still necessary to add context and distinguish fact from plausible hallucination, but experts can now scale their contributions 10x as a result of immediate access, minimal latency, and reduced communication costs.

We're moving towards a world of chair-fillers at one end, and maestros at the other. The clearest difference between labor in 2022 and 2026 will be the hollowing-out of the middle.

This is a great illustration of how ChatGPT fundamentally changes "true/false" into a continuum of "truthiness" as measured by plausibility. The infinite AI array is clearly marked as such, but how long will it be before generative extensibility is the (undeclared) norm?

We're all about to get some real-world GAN training.

Move over, Onion. Your services are no longer required.

The title reads to me as a tautology - we use "clever" as opposed to "intelligent" or "formidable" to connote self-contained, puzzle-like displays of intellect. If you are clever but also pragmatic and effective, people will start using very different words to describe you.

Is there a name for venture-driven ecosystem replacement? That is, the Uber-like building of a new equilibrium that replaces the previous one but ends up as flawed as or worse than the original. Here is how the process works:

1. Introduce a new system (Uber) whose premium form is better but more expensive than the old (taxis) 2. Subsidize its operation with unlimited funding 3. People become behaviorally locked into the new system and the old one dies 4. Remove the subsidy 5. The new system degrades into an unhappy medium of mediocre performance at high cost, but the old system no longer exists as an alternative

Dating apps are not "subsidized" in the same sense, but exhibit a similar trajectory. They work well when small and badly when dominant, but by the time we realize the problems at scale, the previous behavioral models (friend-of-friend introductions, work relationships, in-person approaches) are no longer viable and may not even be considered acceptable.

What do we call this pattern?

Title suggestion: "Chess, but moves are simultaneous"

You can read however many PG essays you want to, but you can't control what you think of them.

This is a very interesting definition of "free will" and not one I would expect most philosophers to agree with, but more measurable and useful than most would propose. In this case "free will" reduces to self-modifying action with randomness injected by the environment, which feels more like a stochastic model than genuine self-determination. That said, the existence of a "reflection" function is still the primary determinant of any moral responsibility.

Schopenhauer: "Man can do what he wills, but he cannot will what he wills."

I've wrestled with this question from time to time and, to the extent I've found a solution, the key is duration:

In the short term, choices are mutable (but determined by character), and character is fixed.

In the long term, character is mutable, and determined by some inner nature plus the integral of experience, influence, and choice.

There's still a recursive element here, because the investments you make in yourself are strongly influenced by a prior version of that self. (In a limited sense, this resembles coherent extrapolated volition.) Still, I've found it more helpful to think about short-term vs. long-term processes - in the sense of stock vs. flow, intercept vs. slope, or short-term vs. long term elasticity - than about recursion. Adding time conveys an intuitive understanding not only of what you can and can't change, but what you should and shouldn't.

To phrase it a simpler way: willpower is a muscle, as are most other traits. Most people can be strong, but not all at once, and when you see someone lifting a heavy weight it means they've been training for a long time.

This may be the only time this phrase has ever been warranted, but...crypto solves this?

I mean, you want decentralization but Mastodon is decentralized at the whim of whoever happens to own your server. So, what if you had a bunch of fungible providers all incentivized to etc etc...

If it were as simple as comparing expected values and randomly choosing one option whenever their confidence intervals overlapped, then hesitation would be irrational. In my experience, though, choices are usually more complicated because you have to think about both expected value and variance (or, more generally, the distribution of returns). If A and B look "about the same" but the error bars are giant, it's usually worth the effort to cut them down a bit. Sometimes you find that the ranges still overlap afterwards, at which point you flip a coin. But just as often, you find that one option is now tightly bounded while the other has a very distinctive risk profile - that is, high upside or high downside. Now you've gained valuable information!

The general problem of making choices with limited information reminds me of a thought experiment from business school: suppose you have a startup that's worth $5B if it succeeds or costs $1B if it fails. You could just try it, and succeed with 50% probability, or you could run a test that tells you for sure whether it will succeed or not. How much should you be willing to spend on that test? The expected value is $2B up front, but $5B if you know it will succeed, and you have a 50% chance that the test comes up positive, for a $2.5B expected value if you can run the test for free. That means you should be willing to spend up to $500M on the test! Hence venture capital.

This argument is a pretty egregious example of strawmanning. A sketch of the author's reasoning:

1. Consciousness is observer-independent, so if one person observes consciousness, then everyone must acknowledge it.

2. If there are infinite observers of a given computation, each with a distinct interpretation scheme (e.g. observer i flips every i-th bit), at least one will interpret that computation as conscious.

3. (1) and (2) imply all computations are conscious.

4. This is absurd.

5. Therefore, computation cannot equate to consciousness.

No one would claim that pi (the constant) is conscious because some sub-sequence will parse as Shakespeare, nor would they argue that an excerpt of Proust is conscious because it required intelligence to produce. Let's agree on some sensible preconditions for consciousness (entity recognition and some notion of memory, among others) before we start trying to argue by reductio.

Would you and/or your girlfriend be willing to chat a bit about the experience? I'm doing research on the efficacy of coding bootcamps and I think your input could be very helpful. If you'd be open to it, my email is lucas at timesuji dot com. Thanks!

I'm actually researching the efficacy of coding bootcamps and was wondering if you'd be open to a quick call. Doesn't have to be a long convo but I think your experience could be very helpful. Would love to chat about what you encountered and potential solutions - my email is lucas at timesuji dot com. Thanks in advance if you'd be open to sharing!

Evaluating side projects is a perfectly valid way to document a software engineer's potential value to a company, but that's not really the point of HackerRank. The purpose of HackerRank is to be a coding screen - that is, a test of basic ability to code simple algorithmic problems under moderate time pressure.

I have taken a number of HackerRank screens myself, and I have never seen one that either (a) tests beyond the coding level that would be expected of a fresh CS graduate or (b) is not vastly exceeded in scope and complexity by the coding challenges given during the onsite, which do allow considerably more flexibility and relax some of the artificial constraints.

Why attempt to circumvent a test that is meant as a slightly less trivial FizzBuzz?

Hilarious. You could even chat for as long as you wanted by plugging in a charger at 2% and taking it out at 4%.

There are plenty of branches, but it's true there is no way to add your own for AlphaGo's evaluation. The main function of the tool is to see AlphaGo's appraisal of common opening lines from professional games, plus some of AlphaGo's own suggestions.

One impressive statistic from the paper: AlphaZero analyzes 80,000 chess positions per second, while Stockfish looks at 70,000,000. Seventy million, three orders of magnitude higher. Yet AG0 beats Stockfish half the time as White and never loses with either color.

A stunning demonstration of generality indeed.