Hacker News Reader

Comment by layer8 | Hacker News Reader

> Our “algorithm” is overly complex & not fully understood internally. People will discover many silly things , but we’ll patch issues as soon as they’re found! […] Providing code transparency will be incredibly embarrassing at first, but it should lead to rapid improvement in recommendation quality.

Sounds like Elon wants to crowdsource the analysis work.

View on HN

After having fired many of the devs, and the remaining ones not figuring out what all this is about, this looks like a desperate (or smart?) move to fix things while not spending a cent...

Why you say that the remaining developers are not able to understand the algorithm?

It’s literally in the quote above:

> Our “algorithm” is […] not fully understood internally.

What makes you think anyone understood it before the layoffs? It sounds like lots of different people had been caking on layer after layer over the years to push things in whatever direction they wanted. Much of the code likely was put there by people who left the company long before Musk bought it.

Probably was not fully understood before the layoff too.

The interesting question is: can it work?

Unlike with the software itself, there was never a serious attempt to open-source such algorithms in the big social media space. Obscurity was the undisputed king.

I would be glad if this attempt actually worked out. Sunlight is said to be the best disinfectant. But there will be a lot of bad actors trying to nudge the resulting algorithm somewhere in a subtle way. I wonder if all of them can be detected.

There don't seem to be many examples of a once-proprietary tool improving after going open source. Blender is the one that comes to mind, but that's driven by a dedicated user base full of working professionals looking to get away from expensive proprietary 3D suites. Anyone who can really understand this algorithm is highly incentivized to use it to their advantage rather than improve it.

https://www-archive.mozilla.org/party/1998/flyer.html

Incidentally, also March 31st.

> So this is a joke, right? It's on April 1st. So it's a joke, right?

> No, it's not a joke. It's on April 1st, because the source is going to hit the net on March 31st. The source is going out on March 31st because that's when the folks who wrote Netscape's press release said it would go out, way back on January 22nd.

> Besides, posting a joke announcement a week before April Fool's Day would be bad form.

Note that they set a date and stuck to it.

Somewhere, I've still got the T-shirt.

A 1998 table-based layout probably made for sub-800 resolutions and it scroll wheel scales flawlessly to readability at 2560x1600.

There’s also a distinct lack of politics surrounding Blender. One of the first comments is about Fauci Files, leading me to believe that no matter the content of thee actual release, there will be a shitstorm around whether or not a certain person is vindicated or villainized girthed.

Twitter has nothing to gain from this. Someone smarter could better game the algorithm while Elon's "hardcore" team is still clueless.

Another idea I just had - the algorithm isn't meaningful without the data. Give someone the code/algo, and they can project whatever they want onto it because there is no actual proof based on the data.

There will be people who replay tweets through the algo, but as we all know - the secondary market of simulation is going to be different than the in-moment reality.

(We also have no reason to trust Musk or believe the code will be complete. He is a well-documented stretcher-of-truth, and likely doesn't understand where all the relevant code lives. After firing 3/4 of the company, it's also likely that nobody at twitter knows this.)

I think his point is to embarrass the prior owners to generate drama and showcase some changes he wants.

I believe in the concept of crowdsourced recommendation engines, but I am afraid Twitter might fail at that.

A great recommendation engine is tuned to what people like. However, that would be subjective, and solving the problem in an abstract fashion seems difficult.

So with crowdsourcing you could sidestep some layers of abstraction. For example, you can crowdsource how you label and group tweets, and use that for the machine learning.

But, for example, if you use crowdsourcing for removing manual tagging of data (embarrassing?) then you might make things worse.

Elon has said he will open source the algorithm since even before the purchase. He said it is for transparency.

Social media algorithms have done a lot of harm, covered and discussed endlessly on HN, and transparency would be a first step to reducing that harm.

I believe the move is in line with his original motivation.

Sounds like a smart decision. A lot of people care about this and would like input. So much so that they will work for free. I fail to see why this is an issue from Musk's perspective or from mine as a Twitter user.

Reply on news.ycombinator.com