Notices by Nolan (nolan@toot.cafe), page 38

@zigg Sure, and maybe it's a pedantic difference, but to me "algorithm" usually makes me think "hand-coded." E.g. Naive Bayes, Support Vector Machines, and Conditional Random Fields are algorithms, but what they spit out are "models."

An algorithm is human-comprehensible (well, if you can understand the squiggly math formulas, which I often couldn't 😂), whereas a model is usually beyond human understanding.

@suetanvil Not a hill I'm willing to die on, but yeah, it bugs me too. :)

@lnxw48a1 Yeah but Google still has 80% of the search market: https://www.netmarketshare.com/search-engine-market-share.aspx?qprid=4&qpcustomd=0 . Every time one of those 80% of users clicks on a search result, it feeds back into the algorithm and makes their model better. I think the case of Google search kind of exactly proves the point. :)

Sometimes we would add or remove features (i.e. data points) and it would increase or decrease our core metric (usually precision/recall on some percentage of test data), and often we'd have no idea why. Feature engineering is sometimes more art than science.

It kind of didn't matter, though. If the number went up, it was good, and if it went down, it was bad. What this means, though, is that in a lot of these ML systems there's nobody at the wheel. The algorithms just build themselves.

@pony Yup, exactly. I got a bit disillusioned when I realized that an upstart competitor wasn't going to blow away the big industry giants with some breakthrough algorithm. Instead it'd just be that whoever had the most data would have the best AI system.

"Algorithms" to me conjures up images of hand-coded software, whereas the models built by ML algorithms are usually beyond human comprehension.

I remember doing feature engineering and just staring at what was essentially a gigantic spreadsheet of hundreds of thousands of data points, trying to figure out why it determined that "Roland" was a company name rather than a person's name. (Hint: because we trained it on a mailing list for keyboard geeks. 😂)

I always find it a bit odd to hear news media talk about "algorithms" used by Facebook/YouTube/etc. Having worked in machine learning, my experience is that it's less about algorithms and more about massive amounts of data and very simple matrix math.

One of the most famous papers on this is Banko Brill 2001 (http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.19.8013), which basically shows that at a certain point the algorithms don't matter, and if you just throw more data at the problem you get better results.

@baldur Yeah I agree Netflix is better than YouTube, although even they have their own weird feedback loop: https://www.salon.com/2013/02/01/how_netflix_is_turning_viewers_into_puppets/

Also your last point is probably spot-on. It's much easier to see the exploitation when it's directed at children as opposed to ourselves.

What are we doing to our brains, to have an algorithm serve up whatever happens to fascinate it?

https://toot.cafe/@baldur/99054649924894561

I'm starting to wonder if the most important thing about the YouTube Kids controversy is that, as adults, it's really easy for us to see through the scam. Kids are fascinated by gross stuff, so the algorithms serve it up.

But is it so easy to see what Twitter/Facebook are doing to our brains? It's the exact same system, except instead of creepy dentist videos it's scary news articles, moral outrage, sex scandals, economic anxieties, etc.

@gdad_s_river OK, I updated the readme. I also found a good Wikipedia page explaining the tradeoffs of a linked list versus other data structures: https://en.wikipedia.org/wiki/Linked_list#Tradeoffs

I'll note I'm relieved to see that I was right that "delete at beginning" in a linked list is O(1), as opposed to a dynamic list / array list which is O(n). I haven't taken a CS class in 10 years and I was worried I forgot everything. 😅

@baldur The scariest part to me is James Bridle's point that this is basically baked into the system. If kids are fascinated by this kind of disturbing content, then the feedback loop of algorithms/advertisers/keywords/etc will ultimately dish it up to them. Algorithmic moderation can only go so far, and given the volume of content, it'd take an army of human moderators to police it all.

Also… he points out this is essentially how all major social media platforms work.

@gdad_s_river To be fair I haven't actually benchmarked this to see if various JS engines have any clever optimizations in how they represent arrays, and depending on the size of your list and how you use it, you may be better off with a built-in array. Perhaps I should say something simple in the readme like "this is a linked list implementation, good for small FIFO lists, enjoy." :)

@gdad_s_river I don't think the answer you got on hashnode is quite right.

In principle, because tiny-queue is just a linked list, a shift() is O(1), whereas on an array list a shift() would require copying all items one position over, so it's potentially O(n). (Unless the implementation is clever enough to just keep a pointer to the first item and move it over.)

@dragonluvr89 "Masculine casual," which also sounds like my look https://toot.cafe/media/kn7tTgt8m-FZEfHcrF8

Thanks to @sorin we now have two new themes: Photon Dark and Photon Light. You can use these themes by going into Preferences -> Site theme. Enjoy! ☺ https://github.com/tootcafe/mastodon/pull/12 https://toot.cafe/media/MCeGIH1s4-_c2GLZd_I https://toot.cafe/media/AO9rYK5dx-gwHrr-0cU