cityhall

They're probably calling things projects that most of us would call tasks or minor features.

They're great for doing phone stuff without pulling your phone out of your pocket and unlocking it. I can check if a notification is important, get GPS directions, snooze alarms, get OTP/2FA codes, or check the weather. None of it it is critical but it's collectively a quality of life improvement.

I wonder if their real intent is to gather training data on which parts of papers are considered important by readers, and which topics are related to each other.

This is probably an extension of their hostility to mp3s. They wanted to force people away from downloaded or ripped content and toward their own store.

What are the standards for plastic-free? I suspect almost every piece of clothing with zippers is using nylon thread and backing, as well as plastic materials tags.

And the F-15's design slogan was "Not a pound for air to ground". Guess what it's being used for now?

I was stuck behind a big RV recently who was driving part way over the center line most of the time. I eventually realized he was trying to avoid clipping all the low branches along the edge of the road.

Search engines refer people to those old threads. Users who care enough and are tech savvy enough to delete their post histories are more likely to have made higher quality comments in the past answering tech support questions, reviewing products, etc. Those are the old threads that appear in search results and potentially draw in new users.

One thing it's still useful for is getting user reviews that actually tell you if a product is junk or has recently been replaced with an inferior version. Store sites and review pages make it too easy for companies to get negative reviews hidden.

The solution to social media companies behaving badly once they have users locked in may not be to lock yourself into another unprofitable company's closed ecosystem that's temporarily free.

I can get a much better idea how my ML models will behave at scale if I prototype on my 16GB MBP than if I only had whatever ram is leftover from 8GB after Chrome and everything use up most of it. It's no match for server GPUs but I can run multiple Jupyter notebooks with decent size datasets loaded and not have to think about it.

The new Macbook keyboards have such low travel that they feel like a step in that direction. I like the laptop overall but I make a lot more typos than I did on my old one.

I've made this same prediction before. In 2001 it was driven by people owing AMT from option exercises in addition to day trading. We got tax paperwork from our brokers though - I'm not sure how many crypto traders will even realize they owe taxes.

The crypto market is less dependent on the US but it may also be priced at the margins and unable to absorb a big selloff.

Go to Burning Man or find the crowd that does.

This is a trend I can get behind, recruiting by advertising that you pay well instead of telling people they should get a sense of pride and accomplishment by working on hard problems after surviving your interview process. Unfortunately I didn't see any numbers in the article.

They still are when you factor in benefits. Family health insurance costs more than single.

That's a different problem. Some things are hard wired by biology, but individual organisms learn during their lives in spite of vague feedback. That's what we're trying to replicate.

MNIST error rates are noise these days, but otherwise agree. These guys have been around for the entire deep learning boom of the last decade, and have never produced results to match their handwaving predictions. They're the Moller Skycar of the field.

I remember reading about Puerto Rican mafia bosses doing this to show an explanation for how they could afford their lifestyles. It was either known in the community that winners could be sold for cash or they'd have a deal with the stores selling the tickets. Some tickets win big right away and others are brought back to the store for confirmation.

Your Honor, please consider the thousands of banks I didn't rob.

"Everyone who doesn't agree with me should shut up and go away."

People sober up when there's a defined end date, but it still takes camp leads and the threat of penalties to get everything packed out properly.

I've been trying to do this better recently after having some non-reproducible results. I've settled on taking all hyperparameters (including booleans like whether to use batch norm) from a global dict. Instead of commenting and uncommenting lines, I look up a key with a default value, adding the default to the dict if it wasn't there. Then I print and log the dict with the results.

I end up with a bunch of code like:

    if get_param('use_convnet_for_thing1', True):
        convnet1_params = get_param('convnet1_params', None)
        thing1 = build_convnet(thing1_input, convnet1_params)
    elif ...

By logging the hyperparameter dict, source checkpoint, and rand seed, results should be reproducible.

This works well for rapid iteration like in jupyter notebooks. For models that take days to train, you might as well use source control for your scripts.

Some ML applications will replicate expert behavior (e.g. document discovery in law) or improve on it (quant trading). A lot of the gains will be at the other end of the skill spectrum though.

Universal human abilities like speech and face recognition have made some low-skill jobs harder to automate than expert ones. We've only just invented algorithms that solve those problems adequately. If your vertical ML application is a machine that picks fruit when it's ripe, it's a stretch to call it an expert system.

Anyone have any insight into what the top ML people are being paid right now?

From the stories I've heard I wouldn't be surprised if some were still on DOS. It's hard to develop in house with a tightly controlled budget and it's hard to outsource when you're dealing with sensitive data.

Russia

If you have to rely on model serialization schemes you have a problem because they express the model in terms of low level operations.

You probably want to do experiments with multiple model variants, or teak your model and fine tune from deployed weights. To do that you need a way to recreate it from layer-level objects instead of the add/reshape operations Tensorflow and its kin store internally.

Universal approximation just says there exists a solution, not that there's a reasonable way to find it with a bounded amount of training data. The networks used in proofs typically have contrived and impractical architectures.

Did you try the underpriced ones, and were they good? A particular wine can be complex but have some specific quality that makes it unappealing. I'd be surprised if complexity is the main variable between different vintages of the same wine that have very different prices.