Search all text in New York City 11 months ago
This write-up about the site is also fascinating: https://pudding.cool/2025/07/street-view/
HN user
Kortaggio
2,121 karma
All about me: https://billmei.net
Posts142
Comments75
Jane Street's Figgie card game 1 year ago
Previous discussion: https://news.ycombinator.com/item?id=38679078
Nice! I added this to the catalogue https://github.com/billmei/every-chatgpt-gui
NumPy 2.0 2 years ago
Duplicate HN thread: https://news.ycombinator.com/item?id=40700259
The No-Stats All-Star (2009) 3 years ago
I found this to be incredible, an NBA player whose stats are worse than other players on practically every dimension, "yet every team he has ever played on has acquired some magical ability to win". Many tech companies look at individual-level productivity metrics like PRs submitted, sprint velocity, deploy time, etc. But what are some good ways to assess team-level productivity?
From Nand to Tetris (2017) 3 years ago
Part 1 was significantly easier than Part 2; Part 1 (the hardware part) only took 2 weekends for me, but every assignment on Part 2 took several weekends each. Granted, this was also because I was learning about operating systems and compilers at the same time; if you're already familiar you could considering skipping Part 2 altogether, or otherwise you might have an easier time than I did.
From Nand to Tetris (2017) 3 years ago
This was an amazing course and is one of the most rewarding computer science courses I've taken! I loved that there was nothing left to "magic" and it was the first time I felt like I understood the "full stack" of the Java-like code I was writing right down to the transistors.
Self-plug for a full-blown minesweeper game I made for the final project: https://github.com/billmei/nand2minesweeper It's a complete game with a tutorial, custom RNG, and unit tests, using their hardware simulator.
Seaflooding 3 years ago
XKCD talked about why flooding Death Valley in this manner would be a bad idea, because unless you remove all the salt and sediments left behind from evaporation it would create a toxic waste pool https://what-if.xkcd.com/152/
The term to google for is "insured cash sweep", but the specifics depend on the particular financial institution you work with. Instead of a single insurance provider, your cash is sharded behind the scenes among many member institutions. Here's one bank I picked at random from Google[0]. Brokerages like Schwab[1] and IBKR[2] also have a private insurer (both use Lloyd's of London) they offer as a service to customers.
[0] https://www.stearnsbank.com/personal/high-balance-deposit
How Duolingo reignited user growth 3 years ago
For most people, I now recommend the Pimsleur[0] method as it explicitly teaches the “call and response” model, which I find most effective for the reasons I mentioned in my comment.
How Duolingo reignited user growth 3 years ago
For me, the main reason for Duolingo's drastic decline in quality is not solely due to gamification but due to their learning algorithm no longer being effective.
I was one of the early beta users and joined Duolingo in June 2012 before it was publicly available, and was a huge fan in the early days. At one point I had a streak of 1,000+ consecutive days, and even applied to work there.[0]
Back then, Duolingo used the SRS algorithm[1] and it was very helpful in keeping me on track with my learning.
The original business model was to make money from community translations.[2] I think they soon discovered that this market wasn’t nearly big enough to sustain Duolingo as a going concern, so they pivoted to an advertising-based business model.
However, the software supporting the original “translations” business model is still in place, so the way they teach you new words is prompting you with sentence pairs to translate. For example, they’ll give you an English sentence and ask you to translate it into Spanish. I find this isn’t an effective way to learn a new language because you’re not learning to think in your target language, you’re learning how to match words between your native language and the target language, which is not the same thing.
Sentence matching is unnatural and requires neural pathways that are never used by native speakers; native speakers don’t think in English and then translate into Spanish when speaking. To be able to fluently speak the language, I found that a “call and response” model is the best way to learn. For example, given a question that a real person would ask (“What do you like about your job?”), you have to creatively formulate a natural answer (“I have a lot of intellectual freedom”) without any “guardrails” of a scripted English sentence provided by the app.
“Call and response” more accurately models how people actually speak in real conversations—you take turns talking like a normal human being. It is also more difficult and requires effortful assimilation of concepts instead of just words. To learn, you must make mistakes, so the most effective learning techniques are also the most frustrating. Coming up with novel responses forces you to fully engage in the dialogue, which efficiently helps you to lay down neural pathways that you will use in a real conversation.
Duolingo is optimized for ad revenue, so they can’t give you difficult challenges because you’ll just give up. Instead, they must spoon-feed you 1:1 sentence translations or “fill in the blank” type challenges because this rewards you with a dopamine rush from their slot-machine-like “ding!” sounds and green checkmarks that pop up when you successfully fill in the blanks. Instead of encouraging you to make mistakes, Duolingo will penalize you by forcing you to redo entire lessons (instead of targeting practice towards your weaker words), sweep mistakes under the rug, or making you pay money for gems. While the paid version removes the ads, it doesn't remove the algorithm that serves lessons optimized for ad-interrupted learning.
Duolingo touts their effectiveness study[3] as evidence that its methods are scientifically supported with a claim that “an average of 34 hours of Duolingo are equivalent to a full university semester of language education”. However, this study was conducted in 2012 under Duolingo’s old SRS algorithm before they changed their teaching model to matching-based problem sets, got rid of their call and response questions, and implemented hearts, gems, and lingots that penalize mistakes, so I don't think this evidence applies to the current app anymore.
Unfortunately, due to these changes I now disrecommend Duolingo for folks wanting to learn a new language, even though it was a pivotal and cherished part of my language learning journey back in the day.
[0] https://billmei.net/blog/silicon-valley-job-search#duolingo
[1] https://en.wikipedia.org/wiki/Spaced_repetition
[2] https://www.buzzfeed.com/buzzfeedpress/buzzfeed-expands-inte...
[3] https://s3.amazonaws.com/duolingo-papers/other/vesselinov-gr...
I wish I could organize my thoughts 4 years ago
I think HN automatically removes "How" from the title; I submitted this post[0] whose original source starts with "How", but after submitting, the word "How" disappeared from the the title.
Hi! Author here - thanks for the feedback. I added the following footnote to the article to clarify:
I use the term "hazing" broadly to mean any way of excluding members from joining an organization. For example, a job interview is a form of "hazing".
In response to "How can you trust and be loyal to someone you just met?" I'll provide a personal example: I recently travelled to a different city, and asked a friend of mine for sightseeing recommendations because he used to live there. During the conversation, he mentioned that his ex-girlfriend still lives there, and offered to put me in touch with her. I then asked his ex-girlfriend (who I had only one prior interaction with!) to be my emergency contact while I was on my trip, which she agreed to, because of the shared trust they had established in their prior relationship, and their familiarity with me in our mutually overlapping social circles. This is what I mean that a person's entanglements are more important than their attributes.
Hi! Author here. I'll admit I struggled to find a more precise word than "context". What I mean is that a friendship is not the 1:1 interaction you have with your friend, but that friend's omni-directional interaction with every other friend. Their location within a network is one way of putting it. In other words, their entanglements in your web of mutual commitments, shared history, and collective responsibility.
The trick I use to remember this is:
DNA: D means Disk
RNA: R means Ram
Chatmosphere 5 years ago
I cleaned up this list by category and added descriptions to every site: https://github.com/billmei/every-proximity-chat-app
(Originally I replied to your previous comment with lots more sites as a raw list here: https://news.ycombinator.com/item?id=25664862 )
GitHub is fully available in Iran 6 years ago
AI Will Revolutionize Education 6 years ago
Reminds me of "This Will Revolutionize Education: No prediction has been made as often, and as incorrectly as that." https://www.youtube.com/watch?v=GEmuEWjHr5c The main problem is that learning is hard, and occurs inside the learner's head, rather than anything external outside of the learner.
"Trading Trump 2020 and other President 2020 contracts is not permitted by residents of the United States, Canada, the European Union, the United Kingdom, Singapore, the UAE, Cambodia, Turkey, mainland China and Hong Kong SAR, and other FTX prohibited jurisdictions."
I think PredictIt is only available in the U.S., their identity verification requires a U.S. address.
What if carbon removal becomes the new Big Oil? 6 years ago
In effect, this is what happened to the collapse of Easter Island. Each subsequent generation forgets what was "normal" for the old generation. See Jared Diamond's fascinating account (PDF): http://employees.oneonta.edu/allenth/Class-Readings-Password...
Software Architecture Guide 7 years ago
The author's statement that "My view is that applications are a social construction" I think hints at a deeper truth, that while _code_ is run by machines, managing a _codebase_ is more like raising a plant in a garden and not like changing the oil in a machine. Reminds me of "A Codebase is An Organism"[1].
You may be interested in reading this paper from Nick Bostrom, where he outlines the dangers present in certain types of truthful information: https://nickbostrom.com/information-hazards.pdf
Half of All Phishing Sites Now Have the Padlock 8 years ago
EV certs can be spoofed too: https://stripe.ian.sh/
It's crazy to me that there are so many moving parts to businesses like this - how do you keep it all in your head? Or are there strategies to reduce the complexity of the business problem so you don't have to keep everything in your head?
List of unsolved problems in physics 10 years ago
Wikipedia has some interesting lists, including this one of a more general list of unsolved problems across many fields: https://en.wikipedia.org/wiki/Lists_of_unsolved_problems
One of the biggest weaknesses of SMS 2FA that I didn't see the article cover is when an attacker can socially engineer their way into your account with your cell service provider.
I'm thinking of a high-profile example when an attacker tried to take over h3h3's YouTube account by requesting his SIM card from T-mobile by pretending to be a T-mobile employee: https://youtu.be/caVEiitI2vg
Vsauce has a fantastic animation of this too: https://www.youtube.com/watch?v=ObiqJzfyACM&t=16m4s
Deal yourself five cards every billion years. When you finally deal yourself a royal flush, buy a lottery ticket. If the ticket wins the lottery, throw a single grain of sand into the Grand Canyon. As soon as the Grand Canyon is completely full of sand, remove one ounce of rock from Mount Everest. By the time Mount Everest is level: take a look at the clock. You'll have to do the whole thing all over again 256 more times before your timer reaches zero.