Hacker News Reader

Comment by Barrin92 | Hacker News Reader

Exactly. Every time I see some new article about how ML is going to mind-control us through perfectly placed ads and Yuval Noah Harari writes a new book about the machines knowing us better than we do I just have to spent five minutes on Amazon or Youtube. I'm not afraid that the machines are too smart but too stupid to be honest.

I'm not convinced that if you went and did an actual scientific study on the efficacy of "big data" teams and ad-targeting and went through every business it would do any better than just regular boring contextual ads.

View on HN

I spent a lot of energy and time studying applied ML in university. After competing in several contests I realized in most cases a simple linear equation threw shade on the whole carnival. Don't get me wrong there are amazing applications for ML, image recognition, simulated physics, the list goes on. But most of the commercial applications I've seen render a ML model wholly unnecessary. You don't need a heavy handed big data machine learning model to figure out, "people who bought also bought", nor does a crime prediction system need to be more complicated than, "where the cops and poor people were last year". FFS a national convenience store contracted one of our classes to do analysis on 10 years of data and the shocking conclusion most groups arrived at was, "people smoke a lot of cigarettes and drink a lot of beer". A revelation. Maybe if it were required curriculum people would realize you don't need to a hire a warlock to do basic statistics...

At my job, there is a problem that's classically solved with a bunch of regexps and run of the mill text parsing.

The ML experts in my company keep telling me maintaining these regexps is too costly dev wise, and that they can use ML to adapt to changes in the real world, and get a better result faster.

After nearly a year of work, buying some fancy new computers with giant GPUs, and countless meetings to describe the problem space, they have something that's around 60-70% as accurate as the current state of the art, and say they're close to closing the gap.

Meanwhile I add a new regexp or tweak an existing one and cause them to go back to the drawing board because I've "changed the entire problem description".

Sometimes I feel like I want the job where the only criteria is that I'm using the fancy new toys. It seems like more fun.

I think its possible the targeting works for the advertisers (at some level) while still failing completely for end users. e.g. If I get my hit-rate up from 1% to 2% by moving from contextual to 'give-me-all-your-data'-targeting, then that doubles my return as an advertiser, while as a user I don't see a difference and still think: "all ads I see are terrible".

The lie is telling users that targeted ads are good for them

Reply on news.ycombinator.com