mochomocha

I once worked at a US biotech startup getting paid by insurers where we would build ML models just to reverse-engineer the inconsistencies of CPT coding & claim reimbursement by insurers & hospitals so we could get paid. Basically trying to automate "learn that Karen at insurer X processing claims wiht code Y from hospital Z usually decides to reimburse $XX". This whole idea of "free market is more efficient for healthcare" has to die. All the incentives are backwards. For example: insurers don't want efficient claim processing because that means less revenues for them.

There are low-latency applications of MIPs where such overheads matter. Sometimes one needs to solve small MIPs very often under strong latency requirements.

The few times I've tried to read Kahneman's book I must admit that I stopped pretty early. I found his writing style too authoritative, without presenting much evidence for the theories he lays out.

I also use D for data & ML at work (Netflix) when standard python workflow doesn't cut it anymore. There's just so far you can go with the typical "call a popular python lib interfacing with a C++ backend" paradigm before hitting a brick wall.

Oh I wasn't aware of these, it looks awesome - thank you!

Which in my opinion makes the "goes pro" statement misleading. Without proper JTAG/SWD debugging functionalities, how do you program anything serious on Arduino?

... In theory, if you could debug your code. The debugging capabilities on Arduino boards are unfortunately basically inexistent compared to traditional MCU dev boards. Hardware is great I agree though.

For me from my laptop: 4 above the fold, 3 below the fold. Each ad also has more pixels of real-estate than an "organic" search result. More than 70% of the pixels I see on my screen before scrolling belong to ads!

> The work is to make the services better so they will attract more users. It is by delivering great services that Google attracts users.

This is less and less true. These last few years we have seen aggressive monetization push on most popular Google products, at the expense of product quality: search, maps and Youtube have all seen a significant increase of advertising that degrades the core product value. Mobile search & maps in particular have implemented deceptive techniques that make sponsored ads look like "native" results.

Try typing "car insurance in San Francisco" and tell me how many times you can see the little "ad" logo... Getting tinier and tinier by the day. Remember when sponsored search results used to have a blue background color?

The Maps app is also barely useable anymore because of how bloated it is. I have a basic Nexus phone from 2 years ago, and it takes ~10 seconds to fully load (!)

I think your analysis is spot on, thanks for sharing.

How do you know your increase was specifically because of them and not practice only?

That's not completely accurate. Hardware & OSes have not been designed for heavy multi-tenancy, leading to (among other issues) cache interference. At Netflix [1], we have to do quite a bit of work to undo that. Bigger orgs like Google have spent more than a decade improving the Linux scheduler in their own fork for similar reasons.

[1] https://medium.com/netflix-techblog/predictive-cpu-isolation...

> He went on a week long cabin-in-the-middle-of-nowhere trip about a year ago to dive in to AI (that's all this guy needs to become pretty damn proficient).

You must be joking, right? I'm as much of a Carmack fan as anyone here, but overstating the skills of one personal hero does no good to anyone.

I agree. Third party ad companies bidding on behalf of advertisers for the best inventory have incentives more aligned with advertisers IMO (as opposed to Google and FB who tend to side with publishers since they control inventory). The main caveat is that it comes at a privacy cost (broadcasting cookie information on ad exchanges).

Google removes targeting knobs because despite what marketers think, humans consistently underperform algorithms when it comes to targeting. Typically, the people tweaking the knobs have reverse incentives because they don't want their jobs to be automated so they love these knobs. But they are just constraints on the bidding systems that end up delivering poorer performance of the campaigns.

YMMV but for the vast majority of campaigns, the algorithm does know better.

Disclaimer: I've built bidding systems for multiple ad companies, so I'm surely biased towards "the algorithm knows better".

The ROAS numbers given by FB (your employer) are far from being a real estimate of the causal effect of adverts. For example, there's no way to have global holdback against other advertising channels not controlled by FB. Double-digit anything are easy to get on any advertising perfomance metric if you don't scale your spend, so it's not really an argument.

(Disclaimer: I've built multiple bidders & ads incrementality models over the years so I'm probably biased against the "supply" side of ad inventory).

I can confirm this. I've run similar studies at my current employer (some of them with one of the persons you cite in your article) on multiple very large advertising platforms for display advertising. The causal ad effect is very often indistinguishable from statistical noise.

It gets you troubles. Just bought a house that had some installed. Had weird "lost connection" failures, and the "smartness" trying to learn what's the right temperature for you doesn't work. After two months, I still have to wake up in the middle of the night to manually reset the temperature to something reasonable.

Thanks Atila! I'm curious about your work around reflection, care to give more details? This is one of the great strength of D, and having some API consolidation sounds awesome. Also: any opinion about runtime reflection capabilities? I love compile-time reflection as much as any other D user, but sometimes runtime reflection can do wonders if used parsimoniously.

Compared to Rust: smoother learning curve, and you don't have to fight the compiler so much

Compared to Go: much better designed and richer language.

Obviously in terms of community, it's much smaller. But if you're ready to "build vs buy" more, or if you have some small scope project, I'd encourage you to try it.

Yes! I've only used it minimally in a non high load scenario, but it's been a pretty smooth experience.

Yes. In our case it's definitely worth it. Another interesting thing to consider is by moving from a kernel-level generic solution (Linux CFS) towards ML-driven systems that depend on the actual workloads we run on all of our clusters, it also implies different ways of debugging and iterating on software. We believe it's a net positive in our case.

Right. Our approach is to move more logic into user-space so we can implement data/ML-driven solutions more easily.

Definitely.

I'm one of the author of the post, happy to answer any question.

Not SAT but this a recent example of using Mixed Integer Programming in real life: https://medium.com/netflix-techblog/predictive-cpu-isolation... (I'm one of the author of the article)

Congrats to the whole Datadog team! I have been an avid user of it at a previous job and loved the product. Not surprised to see them have a solid business. When the product is good, people buy it.

mypy is a great effort, but very experimental. Try using it on any real-world large enough project and it loses most of its value as there are still a lot of unimplemented things, or because you'll depend on a third party module that hasn't support for it yet.

For 90% of the cases that interest mainstream media, economics is not a science because proper experiments cannot be run. You cannot have clean counterfactuals and no "A/B testing" setup to validate hypothesis in most interesting cases of dramatic economic events (like a recession). Most theories end up being... theories, with weak validation based on dubious instrumental variables, if any are considered at all.

Are ad-hating HN users ever going to say that ads might work sometimes, instead of relying on random anecdotes and 0 data or facts from an extremely small and biased sub-population who is never the target of advertising campaigns anyway?

Or are you really arguing that >$500B businesses like Google and FB are complete fraud?