Comment by bazmattaz | Hacker News Reader

Comment by bazmattaz | Hacker News Reader

Oh wow what a wild ride of an article! As someone who has led the building of recommender systems (PM) for the last 5 years this rubs me the wrong way.

> I defend myself from arbitrary data collection that fuels the algorithms using PiHole, the tracker-blocking Disconnect plugin, and Firefox, plus a few other tricks.

The author clearly has no idea what he/she is talking about. Sure, Everything they mention blocks collecting tracking information but recommender systems mainly track your behaviour in the experience itself (clicks, page views, time on page).

> That leaves me and other privacy conscious folk with just one lever to pull offer - disliking a video

Again this author has it all wrong. If you’re serious about privacy the last thing you want to be doing is providing more input into the models we create!

And finally the comment that made me realise this author has no idea what he’s talking about;

> The solution to both issues is obvious, technically easy, and yet commercially a nearly impossible proposition: open up all recommendation algorithms. Make them completely transparent, and, for the individual being targeted by the recommendation engine, completely programmable.

Are you joking? So he/she thinks the east answer if for every user to ”program” their own algorithm. That’s like saying every Tesla owner should be able to program their own car! Let’s open the source code of Tesla to every owner!

Idiotic article.

Three thoughts:

a) I agree with you that the author lacks a basic understanding, and that they do not solve the problem.

Yet: b) I think there actually might be a positive outcome for several reasons, including innovation, marketing and ethics, if one is to allow trained model to be queried arbitrarily. At least in the general case. (Think what is happening with stable diffusion and LLMs/LMMs.) Recommendation systems of say Youtube might be more useful to people making videos -- essentially somehow create a pre-query system before one finishes a video.

c) As a couple others have noted, the Tesla example is not a good one. I understand it was formed in haste and I think we all can understand your reasoning. My thought though is:

There is a) the right of repair -- yes still being fought for -- and b) the full self driving ability that might actually demand open sourcing code and models.

How else do we envision the Courts and Congress to formulate liability and empower related LE, and rights groups investigations?

Recall that the aviation industry achieved the "FSD" mode by building the infrastructure, extreme reliability (space and aviation started byzantine systems) and definitely not black box.

P.S. As a personal note: I will never allow something to drive me, if I have any liability, especially if it is a limited "black box" that I can not even arbitrarily query.

Right now we are "cruising" along because liability is with the driver. Tesla and the auto industry will need to face that pill, if they ever achieve their goal. I think they will achieve it personally.

> That’s like saying every Tesla owner should be able to program their own car! Let’s open the source code of Tesla to every owner!

Other critiques aside, these things are not very similar. You can't run someone over with a buggy recommender system. You bork your own newsfeed, who cares?

Ok fair enough, I formed that analogy in haste.

My point was that it’s not easy at all to “open up” an machine learning model for a user to “program”. It’s complex enough for an ML engineer to develop!

In theory, maybe it is possible but it would be incredibly expensive to produce such a system, let alone maintain. But even then, the ability to control an Ml system is such a niche ask that it would never happen

I think you should be able to program your own car and that should be the default. And I think that can be made safe too.

It does not mean that any change should be allowed on the road. But there could be a system where you pay a fee and send your code to be audited and aprooved for use on public roads. Otherwhise you can only run that code on a test track.

And also not all parts of a car's software are safety critical.

If more then zero resources are deducted to this absolutely useless software developer fetishism, I’d be entirely disappointed.

“Audit” and “approval” does not equal “well-tested”.

It is not "software developer fetishism", it is plain and simple ownership, property rights.

If I buy a car I expect to own the car and be able to do whatever I want with it especially on private land, as long as I don't interfere/endanger/etc. with other people. We used to live in this world but we no longer do. If the car has DRM, locked boot loaders, etc. I can not do whatever I want with it, not even on private property, therefore I am not the owner of the car, I am just the custodian.

Same with household appliances, agricultural equipment, phones, etc. .

Mass alphabetization was a quantitative change in the number of people who could read and write and was most probably opposed (the printing press certainly was) by some at the time but it resulted in a qualitative jump in the quality of life of everyone. The same is applicable for programming. As technology becomes more accessible people should expect to be able to make it conform to their needs instead of them adapting to quirks of tech.

The ability to modify the software is just one part of the right to repair. Purchasing a product used to mean ownership which included the right to repair.

Also working in the field and I don't understand why this article is even on the front-page. The author has clearly no idea what they're talking about, it almost reads as a satire.

Oh good, I’m glad I’m not the only one then.

I Have a feeling pooping on recommendations systems is en vogue right now. Social commentators love to point out how “evil” ML algorithms are while conversely complaining how X product is terrible at recommendations.

Speaking as someone not in the field, that's how it struck me as well.

"Completely programmable" is obviously stupid, but having more control over the way items are ranked in your personal feed isn't, particularly when weights are chosen by hand.

I feel like "open the recommendation algorithms!" is easy to _say_, and it's less easy (and maybe impossible) to do effectively.

Should we have access (by law) to the code that makes recommendations on content? That seems like a governmental overreach. If that was the case, then we should have to have direct access to the brain of every editor/layout person of every newspaper. Why was one story above the fold vs below the fold vs on page 4? I don't see a whole lot of distinction between those two.

Once we get access to that, how do we keep people from abusing these algorithms on the supply side? Thought SEO sites were bad before? Guess what happens when those sites know what they're _actually targeting_, as opposed to blindly guessing!

As a question, how well do you understand the recommender system that you've created? For a given piece of content, would you be able to guess at the penetration into the platform? As an outsider, I sometimes feel like the recommender systems I interact with have grown out of the platform holder's control (see: Nazi/ISIS/ISIL propaganda on YouTube, etc.). I think these are the points the article is reacting to.

So if the article is saying "I would really rather not see Nazi/ISIS/ISIL propaganda," then, ok, that's good. That's a good want. But we're going to need an algorithm to determine what that propaganda _is_, and then downrank that for that user.

So, is the conclusion of the article... misguided? Sure! Is the desire of the author valid?

I don't know. I don't have time to read these articles.

I hope the /s at the end of the last comment was loud enough to not be stated.

But I read the article. Yeah, it was dumb. Specifically:

> If recommendation algos aren’t shared then we need - by legislation, if necessary - a switch that turns the recommendation engine off.

This is... baffling. On YouTube - the case-in-point of the article - you already have access to a non-algorithmic feed of videos: namely, your Subscriptions feed.

And if you don't want a recommendation algorithm, good luck finding new content that you like.

And if you "turn off" the recommendation algorithm, you want... what, exactly? The latest videos uploaded to the platform? Still a recommendation algorithm. Videos with words in titles similar to the last video you watched? Still a recommendation algorithm. Videos that the creators of your favorite channel watched? STILL A RECOMMENDATION ALGORITHM.

I was wrong. The desire of the author is far dumber than I originally gave credit for.

I don't see the problem with someone not wanting recommendations on their front page, having a chronological feed somewhere else is nice but it's not the same thing. Videos that get popular, and are therefore recommended, are those that prey on the basest components of our human nature (even if you curate your feed as best as possible), I don't see why we should put up with it.

Your list of recommendation algorithms is inaccurate, a chronological feed is not a recommendation algorithm in any sense of the term, as for the rest you can't compare the deeply complicated and often abusive recommendation algorithms that are deployed on social media to the simple heuristics you listed.

> ... I don't see why we should put up with it.

I mean, you don't have to. Don't want to see YouTube's algorithmic feed? That's cool - just HTTP 3xx your youtube.com to the subscriptions page. There are some good browser addons to do just that[1][2].

Granted, that doesn't remove the recommendation algorithm from your YouTube experience. You would have to use some kind of content blocker to remove the algorithmic feeds from the video pages. I find that uBlockOrigin works well for that.

[1] https://addons.mozilla.org/en-US/firefox/addon/redirector/

[2] https://chrome.google.com/webstore/detail/redirector/ocgpenf...

This is a failure of imagination.

On a federated network where you can have your own server/pod (think PeerTube or Diaspora) you absolutely can have a custom recommendation algorithm. And for news you can have custom filters in your RSS reader that trim the feeds you are subscribed to to only the desired articles and apply a custom order.

And even in a Cloud/SaaS scenario like Youtube, one can imagine a settings page where you can configure each recommend view in some domain specific programming language using an API. Imagine something like IFTTT but for configuring your preffered recommendations views.

Notice that something you do not want is beeing recommended to you? Just open settings and tweak the algo.

This is just like feeding a custom comparator function to a built-in sorting function using dependency injection.

Of course, this would not be in the platforms interest and will probably never happen.

It really is not that hard to imagine if you have the mindset that you should be able to control your own computation instead of being at the whims of a corporation.

The distinction is easy to make, a recommendation algorithm is personal, automated, scalable to a gigantic amount of content, entirely opaque for the average user.

The social implications are much more dangerous than a newspaper editor and journalist following their paper's political bias.

Reply on news.ycombinator.com