priansh

This is honestly awesome, copy editing is such a pain point that we schedule weekly 3 hour blocks to go through all changes live. This would greatly streamline our workflow.

Signed up and saw extraction can take several days — is this done manually? I assumed scanning would be done purely automatically but I can see why there would need to be a manual review component.

The main issue with deploying these systems right now is the technical overhead to develop them out. Existing solutions are either paid and require you to share your valuable data, or open source but either abandoned (rip Crab) or inextensible (most rely on their own DB or postgres).

I’d love to see a lightweight, flexible recommendation system at a low level, specifically the scoring portion. There are a few flexible ones (Apache has one) but none are lightweight and require massive servers (or often clusters). It also can’t be bundled into frontend applications which makes it difficult for privacy-centric, own-your-data applications to compete with paid, we-own-your-data-and-will-exploit-it applications.

I’ve been saying this for years, language models are the ML equivalent of the billionaire space race, it’s just a bunch of orgs with unlimited funding spending millions of dollars on compute to get more parameters than their rivals. It could be decades before we start to see them scale down or make meaningful optimizations. This paper is a good start but I’d be willing to bet everyone will ignore it and continue breaking the bank.

Can you say that about any other task in ML? When Inceptionv3 came out I was able to run the model pretty comfortable on a 1060. Even pix2pix and most GANs fit comfortably in commercial compute, and the top of the line massive models can still run inference on a 3090. It’s so unbelievably ironic that one of the major points Transformers aimed to solve when introduced was the compute inefficiency of recurrent networks, and it’s devolved into “how many TPUs can daddy afford” instead.

Mostly mask fill, but Transformers can be fine tuned to downstream tasks relatively easily (T5 was built for translation but is used for autocomplete in many cases)

If they implemented SSO with Office365/AAD it would be a stellar alternative. The GSuite dependency for SSO is damning

This is really cool but I can’t say I’m a fan of Zapium given it would move to a commercial license in the future. If you have to ship CEF anyhow, what is the performance advantage to licensing and using Zapium over just compiling WASM and shipping binaries with Electron? FWIW — this is how Java, .NET, etc packaging is done for interop with Electron.

I can understand from an ease of use perspective to have the bridge in between but it wouldn’t be worth subjecting a codebase to commercial licensing IMO. It’s not a whole lot more work to use process calls instead there so it seems an odd choice to commercialize that aspect in particular.

The HN collab I never knew I needed, although I doubt the whole issue could fit on there. Maybe the repo name, issue name and issue number instead?

Yeah my first thought was, well you could lay out the timestamp in a way that supports lexicographical sorting, and then realized they had already done so. No reason this shouldn’t be a string!

I find this a bit disingenuous. TikTok videos have by far the highest reach of all social media due to their format and the way the for you page works. Moreover it rewards influencers from interacting with each other’s posts which often leads to top comments on TikTok (and Reels too although it’s much smaller as a platform) largely being an echo chamber.

1 post on Facebook might reach 100s, 1000s of people. Rarely a page with large enough following may hit 100,000s of people.

1 post on TikTok from a moderate influencer can hit millions.

That’s not to say other social media forms haven’t had bad trends in the past, but TikTok has had far more and in a far shorter amount of time. People just don’t interact with other social content the way they buy into TikTok, which is dangerous and leads to far more misinformation — it was maybe not even 3 weeks ago I had a bunch of my friends sending me a bunch of popular TikTok videos and asking about how AI Transformers were here and taking over the world (they are, but transformers in AI have nothing in common with the movies).

Suicide rates are tricky since mental health can have long lasting issues and the societal changes from lockdown will have even longer term effects. How do you attribute a suicide 2 years from now to depression that started during COVID? It’s going to be interesting seeing how the psychology field approaches this.

Touché, I opened this hoping to find a new method to train Transformers and got the opposite

They have their own pixel that presumably helps them match users to visitors; also most sites have a Facebook like button somewhere.

The FB app ID is also one of the most common meta tags behind Open Graph.

A nice small example of what this can do can be seen with Clearbit [0] which does a good job of telling you where someone works based off of nothing but their IP address. Imagine that but with the exponentially larger data warehouses of Facebook or Google, paired with referrer tags (FB has CLIDs that allow them but not you to match clicks to actual users) and meta tags (FB can tell exactly what app, page etc a website is associated with and use that data to advertise to users).

[0] https://clearbit.com

GPT-J is 6B and comes pretty close. Also practically I haven’t noticed a difference.

Keep in mind there are also closed source alternatives: for example, AI21’s Jurassic-1 models are comparable, cheaper, and technically larger (albeit somewhat comically, 178B instead of 175B parameters).

I would argue society is already a Lord of the Flies hellscape.

I’d like to speak to whoever came up with this acronym for obvious reasons

This is cool but I wish they updated the rest of their docs so the platform was more usable. I’m sure under the hood things make sense and integrate well with each other and other Google products but there’s no way I’m going to invest in figuring that out. Don’t know why such a major tech company has such a difficult time writing decent user documentation

This seems like a weird move given that developer tools seldom do well on public markets. I can’t help but think that staying private would do more to maintain their community & preserve the reasons people opt for GitLab over GitHub.

We tried it for a couple months, it was basically a huge hassle getting our stuff to work especially anything that has bindings to CPP (any of the node gyp packages). Also felt like we were putting in a lot more effort to try to get packages with docs for node to work similarly in deno, and weren’t really feeling any value from using deno over node.

If you’re a hobbyist or an academic or starting a totally new project that won’t have dependencies/doesn’t need a large community of modules, then I recommend deno. Otherwise I highly recommend sticking to node until deno comes out with something to make the switch more appealing.

It should never have been introduced or approved in the first place.

I shouldn't need to read the code extremely closely or repeatedly to have to figure out what it does when the while loop does it clearly (and in a bounded way that you can literally mathematically prove).

Also, I would scream if some opinionated dev gone crazy with their linter added pylint disables and comments explaining it every time we use a while loop in our codebase.

Why does the linter rule not, instead, check if the while loop is unbounded and warn the user of that? Surely screaming fire when there's an actual fire is better than screaming fire at the first sign of smoke.

While I agree with the ternary linting, the while loop piece makes absolutely no sense. There are absolutely many, many use cases of while loops; in fact, many dynamic programming algorithms rely on them.

Could you write them as a for loop? Possibly, but why? Is the new 2021 programming fad hating on while loops? Who do I contact to exchange my "js bad" t-shirts for "while bad" laptop stickers?

The logic behind disapproving of while loops in the article seems to stem from the author assuming while loops are "unbounded." This is absolutely, completely, totally false and there are entire branches of reasoning dedicated towards observing the bounded-ness of while loops. Most colleges include this in their curriculums and you usually have to learn to prove (via induction) the variants and invariants surrounding your loop before you can pass basic classes. It's even more ridiculous that, in the same breath, the author assumes for loops are guaranteed to complete execution. I can disprove this in two lines of code and two braincells:

  for x in infinite_generator():
    do_something()

I'll cede that it's easier to write unbounded code with while loops than for loops, but this is programmer error that can be pretty easily avoided by simply being a teensy tiny bit careful (and can be easily corrected). I'd also argue that the majority of these cases where while loops can be substituted by for loops, are also cases where writing a while loop is much simpler than trying to convert into a for loop. (Besides, you could make any of these arguments about recursion.)

You know what else doesn’t have unbounded loops? Excel.

This is also a lie, Excel absolutely has infinite loops, it just yells at you when you do it. The equivalent of this in programming is a linter. Instead of just screaming fire every time the user writes a perfectly fine while loop, why not concentrate your efforts on identifying unbounded loops and introducing a linter rule for that?

Honestly, it's absolutely terrifying to me that someone contributing to the predominant linter for a major programming language not only wholeheartedly believes while loops are always unbounded and usually evil, but also managed to get this introduced (and presumably approved) by reviewers of this linter, and then proceeded to gloat about it on HackerNews. It comes as no shock to me that most people I've worked with disable pylint in their editors if they're this unreliable with their review process.

Yeah! So I set this up locally for email and a self hosted calendar, but I generally dislike modern calendar interfaces and am looking to experiment in building my own out (probably using Rust or JS).

NextCloud's engine doesn't seem to be very friendly/lightweight in this regard as it's been built specifically for their frontend :/ I was hoping there was some library that accomplished the same thing while being more lightweight/having friendly bindings.

This is great, but what would be even more awesome would be to also go the other way!

I've been looking for something like this for a while! It's awesome to be able to search videos like this.

Any hope for a Chrome extension?

This is really awesome!

I can see a really huge impact for this in allowing those who are less abled to create art. It's hard to express yourself when you physically can't but technology like this can make it possible to create art in many different forms!

https://github.com/jcsilva/docker-kaldi-gstreamer-server

^^ If you're getting started, just following the steps here can get you set up really fast. IIRC there's also an HTTP endpoint at /recognize you can use instead of WebSocket if you're transcribing audio files so it's pretty cool!

wav2letter is pretty fast we haven't been able to break 1.1x on a t2.medium in any of our benchmarks -- what's your setup here?

I definitely think it's a big step in the right direction; it's easily 100x faster than DeepSpeech for us.

If I could have anything I wanted for xmas, I'd ask for a speech to text system that is fast enough to work in browser thru wasm or something.

This.

The article quotes DeepSpeech, wavenets, LSTMs of all sorts; essentially, all the neural networks that scale terribly. DeepSpeech for example is pretty heavy and requires a decent GPU to get anywhere near a realtime factor of 1.

Meanwhile ASR through HMM's consistently hits realtime factors sub-1 and can run on small CPUs. e.g. the default models Kaldi ships with outperform DeepSpeech on a lot of modern examples, and are exponentially faster.

Moreover, training these HMM's is something that is feasible for a normal developer. Training newer models requires data of scale and quality (iirc Mozilla's models are trained on Common Speech which is an enormous crowd sourced dataset, and Google's wavenet models use an internal dataset of very high quality and quantity).

Until the models get more practically achievable, ASR for average people will probably continue to be dominated by Kaldi, Sphinx etc

I would argue that it's the same problem and the reason GDPR is privacy theater.

It's a lot of regulations that can be worked around and the fines are hard to and rarely enforced. There are a bunch of poster children of GDPR fines that make it seem like it's doing a lot but the principal abusers (i.e. Google) just walk away with a light slap.

It needs the ability to be enforced, and I think this much should be obvious to lawmakers -- a law that can't be enforced well is useless.

That's why I'm calling it privacy theater. It's the EU saying "look what we did!" but in practice it doesn't really do much without enforcement that still does not exist both at a national and global scale.

I agree!

IMO, punishing data abusers can't be solved with data privacy laws. It's a regulation with very little ability to be enforced.

In the first place to abuse data you need a lot of it, so really we're looking mostly at big companies with big pockets. The easiest way to attack here is to punish unfair markets and to have stronger antitrust laws.

Let's look at the textbook case for privacy with Facebook. Facebook would be the prime candidate for antitrust no matter how you look at it: there are basically no competitors to it in the US social media market. They own everything except Snapchat, which is dying off and failing to turn a profit. Facebook accounts for so much presence in the US that they have login buttons you can integrate on different sites (Google does too because of how crazy cemented they are). Yet somehow besides being so obviously monopolistic and out of control they're hit with no antitrust. Bell was split up for doing much less.

Antitrust is just a joke right now. We have to get better enforcement first before looking to create regulations to be enforced.