It's time to reveal all recommendation algorithms – by law if necessary

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

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.

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.

Look at what Reddit does: all posts must be submitted to a subreddit, which categorizes it. Each sub upvotes and downvoted that content, which means that the "ranking algorithm" can be dumb as a bag of rocks. At any point, I can choose what content I want to browse without having to "switch algorithms".

The media is preoccupied with "the algorithms", but the problem is with websites that are nothing but a recommendation feed choosing from a formless content soup. It's constantly guessing, the best it can do is "you seem to like this account, I will keep giving you stuff from this person and related people". No concept of interests!

> To help assuage Musk’s concerns, Platformer reports that Twitter’s engineers created a way to “tweak” the site’s ranking system when they noticed a high-profile user’s engagement dropping, ensuring “that tweets from those accounts were always shown.” That’s part of the reason why users saw Musk all over their timelines in February, as it essentially exempts these accounts from Twitter’s algorithm (although Platformer says Musk’s boost was different from what all the other users receive).

https://gizmodo.com/twitter-algorithm-aoc-ben-shapiro-cattur...

> In February, Twitter rolled out a system to boost Musk’s tweets over everyone else’s and began boosting the tweets of these VIPS, albeit to a lesser extent.

https://gizmodo.com/twitter-musk-ukraine-crisis-open-source-...

> Misinformation is highly down-ranked

> Anything that is categorized as misinformation gets the rug pulled out from under it.

> Surprisingly, so are posts about Ukraine.

...

> Late on Monday, Gupta told Gizmodo that “upon further investigation” he and his fellows looking into the Twitter code found that it only applies to Twitter Spaces, the platform’s live meeting feature.

My problem hasn't been getting a barely competent human, it's been getting a barely competent human (or better) that also has the correct authority for my situation. Unfortunately, in the case of both AI or humans, unlocking the person that has both the correct competence and the correct level of authority to correct a problem is the hard part.

If it was resolvable with the automated system, I already tried those steps. If it's not, there's often a reason (for the company, and not necessarily a good one) that it's not. And that's where the authority part comes in; if the AI (or the call person) is doing the "figuring out" of the technical part of a problem but passes the authority on to someone without that knowledge, then how can the person with authority possibly provide authorization without first lending their authority to the end-user support provider?

Ultimately, the reason you can't do the things you need to do is because someone in management removed the authority of that end-user support provide to give it; that will still be true in the case of competent AI, and you'll probably have even less recourse since there is no person to advocate for you to the authority on your behalf.

I think there's a value/don't value axis but also a serious/silly axis. For food labelling as an example:

Contains gluten. Don't value personally, serious.

Contains GMO ingredients. Don't value personally, silly.

edit: more specific to the support desk example - we could label each tier by "has the authority to do $X", I think that would be far more useful.

As soon as someone says the words, "an AI", they have entered a narrative about the thing personified. The original subject is lost to an anthropomorphized facade.

The adjective, "General" was originally meant as a classifier for quality, not category. That distinction is still present on the face of the word itself. An AGI is an AI that can compete with human intelligence. Adding "General" to the phrase, "Artificial General Intelligence" does nothing to define "Artificial" or "Intelligence".

The distinction I just made is entirely based on symbolic meaning. Words are defined explicitly, and then put together. In this approach, the resulting meaning is independent of context: it's literal.

Your distinction takes the opposite approach: an entire phrase - regardless of any symbolic definitions each word is known to have - is defined by the narrative it is surrounded with. In this approach, the resulting meaning is dependent on the context: it's literary.

Large Language Models are famous for demonstrating the second approach: they implicitly model whatever patterns exist in a prompt's text, so that they can generate a "continuation" that follows those patterns, without ever needing a literal symbolic definition.

This is exciting, because natural language - as the result of humans taking both literal and literary approaches - is ambiguous. The literal approach that traditional programming language parsers use is fundamentally limited to "context-free grammars", because ambiguity can only be resolved with literary context.

The problem is that LLMs cannot be literal. They are only capable of inference, so they lose all of the functionally that symbolic definition would provide.

The distinction I made earlier has the same limitation that parsers have: it ignores the surrounding narrative context that has evolved away from the original meaning: it's short-sighted.

The distinction you made has the same limitation that LLMs have: it ignores the original symbolic meaning that is still present in each word: it's confidently imprecise.

What matters is the result:

By ignoring the literary context, the definition of "AI" is written in stone, and we lose the ability to talk around it.

By ignoring the literal meaning, we imply "AI" tech is (by definition) capable of both literal and literary transformations.

We can resolve both limitations by simply using new words. There is no need to say "AI" when "NLP" is more precise.

Historically, as soon as the system actually works for some task, it gets a new name. Thus expert systems, neural networks, logistic regression were all supposedly AI at one time and now are distinct. Already, we see a bit of a divergence in nomenclature where people distinguish large language models from other classes of techniques.

Of course, up to now, it has mostly been techies who have used, abused and redefined the term. More recently, marketing types have adopted the work AI so it may lose all meaning soon.

All computing is "fundamental math".

That there are some known set of algorithms in the mathematical toolbox that are used is irrelevant.

The question is, how are those tools put together, what are their inputs, and how are those inputs weighted.

Assuming an ML-based methodology (and I would not personally assume that), yes, the training data, the resultant weights, and the algorithm itself would be of interest, though at least the last two would be necessary.

Not the only important part.

> Since the weights are input, and usually semi-unique to the user, they would be considered private information.

Cool, then as an individual user, let me see that, too.