Hacker News Reader

jimmy2timeson The worst job interview I ever had

Exactly. Besides cost, domain-specific models (which can still be very large) encode our biases (i.e., our knowledge about the domain) in their architecture. Because of that, we have ways to calibrate their accuracy trade-offs over an in-domain sample.

For LLMs, there is a disconnect between the perceived domain (anything a human can think and verbalize) and actual domain (word sequence prediction). We only know how to sample from the latter, not the former.

This "silver bullet" idea sounds a lot like "free lunch". There has been a lowering of the bar for ML practices to make way for this onslaught of prototype-level productivity. Teams that used to do their best to run uncorrelated evals are now having the prompt engineers manually inspect ~100 model outputs before launch.

People were shocked by the accuracy of n-gram models and user interaction data to "read their mind" and complete their searches. Now we're obviously all impressed by the emerging abilities of LLMs (integrated with lots of business logic by the LLM providers). Hopefully in time we'll get desensitized a bit and have the right mental model when using these tools.

P.S. I love your username :)

jimmy2timeson Tech CEOs are apparently suffering from AI psychosis

I think the bipartisan moment against AI is in large part (surprise, surprise) optics. In reality, the powers that be may have already set their sights on AI as the next "too big to fail" industry due to a perceived economic threat from China.

jimmy2timeson Project Glasswing: An Initial Update

The AIs have already figured out how to succeed in a software job:

1. Ship bugs

2. Fix them

3. You're the hero!

jimmy2timeson Today I've made the difficult decision to reduce the size of Coinbase by ~14%

Sounds like Google.

jimmy2timeson A.I. researchers are negotiating $250M pay packages

I can't believe this is so unpopular here. Maybe it's the tone, but come on, how do people rationally extrapolate from LLMs or even large multimodal generative models to "general intelligence"? Sure, they might do a better job than the average person on a range of tasks, but they're always prone to funny failures pretty much by design (train vs test distribution mismatch). They might combine data in interesting ways you hadn't thought of; that doesn't mean you can actually rely on them in the way you do on a truly intelligent human.

jimmy2timeson A.I. researchers are negotiating $250M pay packages

Can you say the same about the top 100 basketball players?

jimmy2timeson How we renamed our Design Converter to Magical Ass

Does it support plug-ins?

jimmy2timeson Piano Practice Software Progress

Yes, excellent point. Notation can be a great aid to understanding. I should give this a try more often.

jimmy2timeson Piano Practice Software Progress

I guess it would take a lot of patience, but it should be doable, maybe even more feasible than some contemporary music that is heavy on sound fx.

I don't find this as rewarding as trying to understand how a piece "works" though, or creating something new.

jimmy2timeson Piano Practice Software Progress

I think of this as analogous to self-supervised pretraining followed by training on a smaller labeled set. When you study theory you can ground it on music you've listened to throughout your life.

Also to improvise confidently you have to internalize the theory, not just understand it and memorize it at a conscious level.

I'd say "intuitive knowledge" is a good way to sum this up.

jimmy2timeson Piano Practice Software Progress

I would add that music theory and sight-reading are orthogonal. My partner learned piano at a young age, and she can follow a sheet but won't know what chords she's playing. I learned guitar by ear but I'm always thinking about intervals/chords/modes. Obviously having both of these skills would be great.

jimmy2timeson The Cloud I'd Like to See

Besides security and caching, I think standards/interoperability are key.

I want my next-generation apps to work on a cloud filesystem of my choice, be it Dropbox, Google Drive, etc.

That will be the most significant step towards a real cloud OS, and I bet someone's already working on it.

jimmy2timeson Here's Why Google and Facebook Might Completely Disappear in the Next 5 Years

If we're talking really long-term, I don't see a big difference between loading a web app and having a native app quickly and seamlessly installed and loaded just by clicking on a link. Eventually, improvements in bandwidth and computing power might enable just that. At the end of the day, developers will go with what is more feasible, but the whole distinction between web app and native might disappear from a UX point of view (just like the current web apps are kind of a hybrid between "old" websites and "old" desktop apps).

jimmy2timeson Introducing T: A command-line power tool for Twitter

Love the name (though I'm sure many would consider it sexist).

jimmy2timeson Show HN: made a game you can play safely at work

Of course, we Emacs users have always had that sort of thing. That's why Agile folks at Lean startups tend to favor Vim these days. True story.

jimmy2timeson Semicolon: A language of semicolons

DCPU-16 port anyone?

jimmy2timeson A Lisp based programming language I'm creating

I also toyed with my own Lisp[1] a few weeks ago, after being inspired by Peter Norvig's Lispy[2]. It's an interpreter though, and written in C++. I mixed together features from different languages (eager/lazy evaluation, call-by-value/reference, macros). It was very instructive and a lot of fun!

[1] https://github.com/apresta/lime

[2] http://norvig.com/lispy.html

jimmy2timeson Google Search: Change is Coming

For instance, a Google search can distinguish a search for “New York” as opposed to one for “New” and “York.”

OMG, what kind of sophisticated AI did they use to accomplish this? I'm thinking random forests mixed with neural networks and SVMs.

jimmy2timeson Show HN: I wrote a tiny Python-based HN crawler with scrapy

I haven't run the crawler so I'm not sure what else it does, but if it only parses the home page and fetches the external links, why not read http://news.ycombinator.com/rss (you can use the feedparser module) and download the pages with urllib? No scraping involved.