ML is like a genie in that it is very good at optimizing exactly what you tell it to, and so you need to think long and hard about what you're telling it to optimize, because it might turn out to be bad proxy for what you really wanted.

My timeline would be so much healthier if Twitter put me in charge of who shows up in it.

After Google killed Reader I was forced to follow all the people who used to blog to keep up with their professional work. But especially in the last 3 years Twitter feeds have all come with a large side order of political hot takes that I have zero interest in sitting through.

To counteract this I have an extensive word blocklist and turn off retweets on many people so my Twitter can remain about the work of interesting people in my industry and not just an open sewer of politics.

Twitters insistence to show me posts that people I follow liked or engaged highly with bypasses all this.

If Twitter ever moves to a fully algorithmic timeline like Instagram (has annoying done), it will shrivel up and die. The linear timeline of Twitter is what makes Twitter....Twitter. I've noticed a lot lately that more and more I see tweets out of order because of the algorithms. I believe you can turn that off, but sometimes those are useful...but more often they are spam in the moment, especially if a big story has broken. I feel like it used to only occur when you first opened up Twitter, you'd see old highlights...now it feels like it's intertwined more...it's very annoying. Twitter is, by far, the most powerful and useful social media platform out there for information...I sure hope they aren't going to strangle it to death.

I thought I was about to read some satire based on that quote.

Turns out it's not.

I suppose the same applies to many other ML products? Like Netflix's recommendations, Amazon's suggestions and almost all the ads you see these days?

I agree that the problem exists, but its not just Twitter, and this is an unfortunate side-effect of recommendations in general: even if you do count-based recommendations, you are going to have a bit of echo chamber.

Empowering users with Machine Learning, i.e. offering simple powerful ML based filtering tools for the Twitter user.

More importantly allow opt out and configuration of stuffing feeds with politics and liked by those I follow.

But the idea is not to empower users but masticate users to provide more 'value' to Advertisers by selling them 21stC version of 'hits' - the false & forced engagement metric.

This all stems from the mentality that Twitter are serving the money not the user.

It is short-termism.

Hopefully crap-stuffing users will allowing the resurgence of a diverse market that necessitates open standards unlike our current hyper monopolies and their data silos.

I’m not even sort of interested in what Twitter’s heavy-handed definition of “healthy conversations” are.

I dare say they’re acting well outside of their wheelhouse. Their job is to provide a platform for discussion, not whatever that mess of corpspeak I just read is.

Do you remember hiw bad YouTube comments have been? And it's significantly better lately?

That happened due to ML.

Not everyone is out there to use ML for nefarious purpose. And there are multiple use cases for ML.

This is hardly unique to twitter. You don't this is happening at Facebook, Instagram, Youtube, etc...

Just to play Devil's advocate:

One that wants to manipulate your mind

Like your post here? Or my present reply to it? And so on...

one that echochambers your discovery

Which is good, because nobody has enough time in the day to process every conceivable byte of information they might stumble across...

one that censors arbitrarily

OK, because while everyone may have the right to speak freely, not all speech needs to be heard by everyone...

It's probably on the front page because a lot of people are deciding between PyTorch or Tensorflow for their company's production ML pipeline, and this is an article about making exactly that decision. Even if they didn't present any actual analysis, just the sentence "Twitter's ML team, which had previously used LuaTorch, recently evaluated PyTorch and Tensorflow and ultimately decided to standardize on Tensorflow" would be of interest to a lot of people.

PyTorch is rivaling TensorFlow in popularity now. Lua Torch is basically dead.

PyTorch is catching up to TensorFlow, but still has a far way to go especially compared with the user experience of Keras with TensorFlow.

In PyTorch you have to do quite a bit more work specifying gradients, zeroing the gradients between update steps, specifically describing the operations of the backward pass in a custom layer, even in cases when it could just be inferred by the computation graph library as in TensorFlow.

I’ve heard anecdata that PyTorch performance can be better, but I think it’s still a case-by-case topic, and both libraries are still changing rapidly.

I attribute the switch from Lua Torch to PyTorch as just an obvious effect based on how bad of a language Lua is for the type of interactive and experimental programming needed for this type of work. Lua is a good language for some applications, but not this — and the value of spending effort to build the primary Torch frontend in Python became obvious quickly.

I am not a native English speaker, but quite familiar with the use of "at (large) scale" in similar context.

It means that one has deployed it/is using the tool with a non-trivial (could be anywhere from few 100s to few 1000s of machines) amount of CPU and/or data. In the context of serving, "large scale" could also mean the number of queries/second hitting the serving layer.

> I have always used Twitter passively, only re-tweeting 3-4 interesting stories a day

That seems fairly active to me. In almost 10 years I did my first re-tweet last year.

No,they have not. Especially on the machine learning teams

Personally I am very interested in both how models can help solve business problems and how to make effectively engineered tools for machine learning.

In my work I spend a lot of time on new deep learning architectures or experimenting with modifications or fine-tuning or ensembling.

I write a lot of container and Makefile tooling to ensure experiments are reproducible and results have identifiers that map back to the full set of data, software and parameters.

I also write a lot of backend server software to wrap trained models in a web application, mostly in Python, and do a lot of work with Cython after profiling to target only those spots of the code that reveal actual performance bottlenecks in terms connected directly to a specific business problem’s latency or throughput requirements — as in, not taking the huge premature optimization step of assuming a whole system needs to be written in C++, and instead using profiling and case-by-case diagnostics to know when to write something as an optimized C extension module callable from Python for very specific and localized sections of code.

My experience has been that there is such a lack of transparency about how deployment will work, how performance will work, etc., when using cookie cutter pipeline approaches, like sklearn Pipelines, TensorFlow serving, Fargate, etc. You’ll always need to break some assumption of the pipeline, layer in new diagnostics, debug latency issues, etc., on a case-by-case basis.

99% of the time, ease of specifying a new model or articulating an experiment is not hard, requires little dev work, and only represents about 10% of the actual work needed to explore a model’s appropriateness for a given problem at hand.

The rest requires very specialized control and visibility to basically perform application-specific surgery on the pipeline, customizing and tailoring many aspects, from how multi-region deployment should look to how optimized the web service code should be to whether to use asynchronous workers or a queuing service to stage and process requests, to optimizing preprocessing treatments, to instrumenting some extra New Relic metric tracking that the pipeline isn’t extensible enough to just specify in some config, and so on.

What’s been most important is that the deep learning engineers on the team, who are researchers, are also excellent system engineers at all those topics too and display a high degree of curiosity towards them, and absolutely do not look at it like “boring work” that distracts from the experimentation they would rather do. Their value add is not driven by spending more time experimenting — that’s virtually never the case. Their value add is in both knowing the details of the deep learning models intimately while also knowing the deep details of implementation, optimization, deployment, and diagnostics.

In that sense, tying model development to a cookie cutter pipeline framework, whether Spark or sklearn or something custom in-house, is something I believe strongly is an anti-pattern.