A similar experiment I did comes to mind from 3 years ago.

For the fun of it I was trying to train a few deep neural network configurations (LSTM, a few variations of FCNs, ...) to trade shitcoins and downloaded 4 years of 1h candles.

The first easiest idea was to prepare the training data to fire three signals, buy, sell, do nothing (I know a terrible choice). The cost function was setup to do the simple thing and maximize the overall profit (I know an other terrible choice). Fast forward 30min of training and the final outcome is a model that outputs "do nothing" in 100% of the cases.

Not even adding some watermark? smh

ifunny

So elegant! This is too valuable for GitHub, sell this directly to the Saudi government.

It's a decimal point, not thousands

The output is not the output of the algorithm, it's the output of the unit test.

95% of numbers up to that point at not prime. Testing the algorithm that only says "not prime" is therefore correct 95% of the time. The joke is that, similar to AI, the algorithm is being presented as a useful tool because it's correct often but not always.

that's the joke, since most numbers aren't prime, this function is technically highly accurate despite being completely useless.

The test suite probably looks something like this:

int tests_passed=0;
int tests_failed=0;
for(int i=0;i<100000;i++){
    printf("test no. %d: ", i);
    if(is_prime(i)==actually_is_prime(i)){
        printf("passed\n");
        tests_passed++;
    }else{
        printf("failed\n");
        tests_failed++;
    }
}
//...

maybe it would be better to say that it is stochastically accurate?

"AI models have started training other AI models, by pressing The-Button-That-Trains-AI-models; this button was built 7 years ago by a bunch of online volunteers we won't ever credit."

also btw icymi, this is a post about LLMs

True