QML

I believe the California community college system is better than most states. Let me elaborate.

In the summer before my freshman year, I didn't really have much to do, and so I decided to take up a course. My options were to either take it at the UC where I'd attend, or to take an equivalent offering at a local CC. When I looked at the price, there was a 10-fold difference, $2000 against $200.

It was a no brainer to enroll at the CC: transferring credits was easy because the UC system and the CC system already have an agreement [1]. This is in addition to the transfer agreement between the two institutions, where CC students are guaranteed admission to some of the UCs if they fulfill the requirements [2].

Before the first day, I recall having low expectations; I had bought into the mindset that a smaller tuition meant a shabbier campus and a worse experience. I was dreadfully wrong: the place was modern, and my class size numbered around 15 people.

More over, its population was diverse. I didn't see only college students; there were middle schoolers, high schoolers, international students, and even parents dragging their children to camps. In general, the place had people serious enough about their education that they would spend a nice summer doing so.

Here, it seemed was a greater emphasis on the community side of college. I want to mention that this is in contrast to what I've previous seen. I came from a high school which didn't focus on getting people ahead, but on only keeping them from following behind; that was the primarily the idea of summer school. I've also heard of high school students paying for private SAT/ACT tutoring, and for online AP courses. In hindsight, CCCs act as equalizers in the field of education -- I would even say they are the modern day equivalent of churches in our state.

[1] http://www.assist.org/web-assist/welcome.html

[2] http://admission.universityofcalifornia.edu/transfer/guarant...

I’ve seen some writer change their article titles a few times after publishing — I would guess they’re trying to A/B test or just indecisive.

A founder valuing their equity more than its actual worth is a classic example of the endowment effect [1].

[1] https://en.wikipedia.org/wiki/Endowment_effect

I agree with a lot of your points, but the way to increasing the base level of education isn’t through making all colleges free — it’s through investing in and making community colleges free.

Why? If we’re trying educate the entire populace instead of a few high achieving high schoolers, we should recognize that a majority of people can’t afford to up-end their lives, move far away, and spend 4 years getting a degree.

Community colleges, emphasis on the community term, would allow for people — local people — to educate themselves without drastically affecting their livelihoods. In this way, community colleges can also potentially act as job retraining facilities.

Note that I am slightly biased as a Californian.

I believe when most people say “make college free”, they’re referring to a “tuition free” education. There is still a housing cost, which is “skin in the game”.

I'm curious to if this result will extend to the k-nearest neighbors (k-NN) algorithms.

Two problems that have to do with k-NN are

1. It's a non-parametric method: the number of parameters grow linearly with the size of the training set since the distance function must be calculated for all training points and the test point.

2. The curse of dimensionality: distance metrics like the Euclidean distance do not perform well in higher dimensions; points which seem "close" in 2D may be far in 3D, 4D, etc. As a result, we would need an exponential amount of more training data for every additional dimension. Locality sensitive hashing tries to combat this by reducing the dimensionality of the data.

I view the problem in terms of "computational complexity": if the AI of today requires O(2^n) time to perform at a human level, it is possible to get that down to O(n^2)?

I believe there's a connotation with problems, that if the best algorithm to solve it is exponential (brute-force search), then we truly don't understand the (underlying structure of the) problem.

Algorithmic Game Theory [1]

Going into CS as an undergrad, I didn't anticipate the depth that the field had in other domains -- and for some time, I wanted to double major in {math, biology, economics} to supplement my education.

However, while in the algorithms course, I stumbled upon a connection between linear programming and 2-player zero-sum games (the minimax theorem [2]). Up to that point, I had never considered the idea of using a computational lens to view problems outside of CS, such as "what is the complexity of Nash equilibrium?"

It turns out Algorithmic Game Theory can be applied to study theory of auctions (Why does Ebay use 2nd-priced auctions?) [3], tournament design (Why would a team purposely lose?) [4], or something as basic as routing (Why does building more roads lead to more congestion?).

[1] https://en.wikipedia.org/wiki/Algorithmic_game_theory

[2] https://en.wikipedia.org/wiki/Minimax_theorem

[3] https://en.wikipedia.org/wiki/Auction_theory

[4] https://theory.stanford.edu/~tim/f13/l/l1.pdf

Modern AI may "beat" humans at certain tasks, but I doubt they are as efficient as humans are at those tasks: it makes sense that throwing more compute at problem will yield better results. But is it possible to get better results while constraining training time?

Edit: I looked up how much time it takes to train: "OpenAI Five plays 180 years worth of games against itself every day, learning via self-play." [1]

[1] https://blog.openai.com/openai-five/

There's a lot of information already here [1], but I might add it'd be easier to create a Bitcoin wallet for each user than to open a PayPal or Stripe account for each.

[1] https://openbazaar.org/features/

I think in the case mentioned, the implication is that advertisers would be better equipped to target us with ads – which is a fair point.

But why must we burden the theorists with ethics, and not the programmers or companies who bring that theory into the real world?

The difference between the fields of biology and computer science, is that CS is way more abstract. It would be like asking, “what are the ethical implications of math?”

To this end, we need to distinguish computer science (theory) from technology (application) – it is the latter of which has direct effects on society.

What does this mean for NP-complete problems? Is the term “fast” only applicable to this small Sudoku board, or are solvers optimized to the point where something like O(2^256) doesn’t take that much time to run?

This article is one of many in a series covering the recent Fields medal ceremony.

See more here: https://news.ycombinator.com/item?id=17662219

1. Having something equivalent to account with a username and a recoverable password.

2. Credit systems where you loan money to other people based on other's deposits has a multiplier effect [1].

[1] https://en.wikipedia.org/wiki/Money_multiplier

Could you expand on that analogy? I'm not sure if a one-model fits-all is comparable to a one-optimizer-fits-all. Wouldn't that be an argument against transfer learning, since we don't have a universal programming language and thus there does exist multiple compilers?

Wouldn’t the closest analog be WeChat? Square isn’t a messaging platform.

Isn’t the argument between transfer learning and neural architecture search the same problem posed by the “no free lunch” theorem, which essentially states specialized algorithms will always beat generalized algorithms in specific tasks and vice versa?

Ideally, we would be able to provide modern financial services backed by first world institutions to underdeveloped countries. An example of this came through a podcast I listen to [1].

For the most part, I am skeptical of cryptocurrencies since they need to solve stability and identity first before tackling credit. An additional problem might also be usability; private-public key pairs should be abstracted away to the end user. As for donation systems, I don't think they have the same network effect as credit systems.

[1] https://www.npr.org/sections/money/2016/09/07/492988779/epis...

A great complement to that would also be to use Jupyter, whose notebooks are similarly interactive and not dead like typing into a blank terminal or IDE.

In my opinion, Scheme is not a good first-language to learn for beginners or for anyone getting into programming. Why?

Because if you take an introductory CS course taught in Scheme, and you decide the field isn't for you, then you have essentially wasted a semester learning something you'll never use.

Compare this to Python, which is general enough to transfer to other domains (with its numerous libraries) and is way easier to pick up. In addition, its near resemblance to English is friendlier to approach. I'm sure there are some plots out there showing the number of companies which use Python versus Scheme.

This was the intuition, I believe, of why Berkeley for its 61A course switched from using Scheme as its primary language to using mostly Python and Scheme as a secondary language. (You still learn enough Scheme in the course to be able to build a Python interpreter for it).

TL;DR - the language does matter; the first one shapes how one views CS; save Scheme as a later language to be used to shape mental models.

I wonder if there's an opposite term for 'leapfrogging' [1], whereby the pace or focus of technology -- instead of allowing poorer countries to skip trends altogether -- it leaves them behind.

An example of leapfrogging is China's payment system which switched directly from a cash-only system to a mobile-payment system in the span of a decade, entirely bypassing the need for credit cards.

I believe that cryptocurrencies, ironically, are leaving the unbanked populace behind by being a payment-only system instead of being a payment-credit system. If Bitcoin is truly successful and replaces Visa and MasterCard -- both of which also operate as a credit system -- then that's just going to result in slower economic growth in underdeveloped countries.

[1] https://en.wikipedia.org/wiki/Leapfrogging

An alternative like you said is with payment apps. One example I have seen in public is with Venmo, although there are drawbacks like no fraud protection and delay in bank transfer. Latter is solved if you solely use your bank as a settlement layer and Venmo as the application layer — that is, conduct all internal transactions on the app and only settle when you need to trade with a person outside the app.

I disagree: the issue isn’t with the expressivity of the model, but with the objective problem and constraints set by the author. You can quickly adjust the algorithm to product floor plans resembling modern buildings by penalizing construction materials.

I will have to disagree with the author’s characterization of these floor plans as irrational.

As pointed out by commentator from the r/ML subreddit, the plans resemble biological arteries — which intuitively make sense given that the objective was to minimize distance to nearest fire exit. The analog is that arteries try to maximize surface area to volume ratio to ensure blood flood.

This idea naturally lends itself to a clustering problem such as k-means! We can imagine a geographical density as a cluster and that semi centralized pickup point as a centroid. As for deciding k, the number of cluster-centroid pairs, we can leave that up to market prices.

One way to adapt the k-means algorithm to this would be to add a “regularizing” or penalizing term proportional to the number of cars needed to be deployed — you can think of this as the cost per car.

I don’t see how caring for a Waymo or a Cruise fleet would be different from that of a city fleet.

A city fleet is basically the set of vehicles owned and operated by a city government, mostly for use by its employees but could also be interpreted to include public transportation.

A caveat is that the Waymo fleet would obviously be significantly bigger and thus more complex to handle, but I’m sure there’s a physical analog of “sharding” or dividing up fleets into geographical sets.

I wonder how low these self driving cars have to go price per mile to compete with fully mature public transportation systems.

For example, since I live in the Bay Area, the cost of going from Berkeley to Walnut Creek via Bart is around $4 for 16 miles. This would correspond to $0.25 per mile to the $0.35 estimate in the article.

Maybe I’m doing the wrong comparison: cars like these won’t be competing with rail lines but with last mile types like public buses. That would make more sense as I pay around $3 to go 2 miles in my city.

Another thought: the city official, mentioned in the article, seemed enthusiastic in using cars for public transportation. It kind of reminds me of advent of cars when cities focused on building roads and highways instead of investing in public transportation systems. Perhaps the self driving aspect can be transferred to be used on busses? I don’t know whether cost of busses result mainly from labor, like cars, or some other factor.

Perhaps they meant it in the sense that the Higgs Boson particle is sometimes referenced as the “god” particle.

It works in the comparison that quantum computing often sensationalized by media.

(Before beginning, I want to note that these are solely my opinions, and therefore are probably wrong.)

In the space of possible problems solvable by computers, there are those of which are "easy" and those of which are "hard".

Arbitrarily defined, an "easy" problem is any problem that be solved by throwing more resources at it -- whether it'd be more data, or more compute. A "hard" problem on the other hand is the opposite: solvable only by a major, intellectual breakthrough; the benefit of solving a hard problem is that it allows us to do "more" with "less".

Now, the question is: which type of problems are being looked at by today's AI practitioners. I'd argue it is the former. Chess, Go, Dota 2 -- these are all "easy" problems. Why? Because it is easy to find or generate more data, to use more CPUs and GPUs, and to get better results.

Hell, I might even add self-driving cars to that list since they, along with neural networks, existed since the 1980s [1]. The only difference, it seems, is more compute.

All and all, I think these recent achievements only qualify themselves as engineering achievements -- not as theoretical or scientific breakthroughs. One way to put it: have we, not the computers and machines, learned something fundamentally different?

Maybe another approach to current ML / AI is needed? I remember a couple weeks ago there was a post on HN, about Judas Pearl advocating causality as an alternative [2]. Intuitively it makes sense: baby humans don't only perform glorified pattern matching, but they are able to discern cause-and-effect. Perhaps that is what today's AI practitioners are missing.

[1] https://en.wikipedia.org/wiki/History_of_autonomous_cars#198...

[2] https://news.ycombinator.com/item?id=17108179