Korva is a scroll of the internet’s best ideas.
Ads as a testing ground
AI is clearly incredible at coding and math, which are verifiable domains. Judging whether a generated image works for an ad is verifiable too: the ad either sells or it doesn't. Google and Meta can validate their image and text creation in a way no one else can, because their ad marketplace runs a gazillion A-B tests, with humans deciding whether to click and purchase, at global scale. Most ads are throwaways, and the vast majority don't convert.
How ad matching works
Ad matching is still fairly crude. It builds an embedding of the qualities of the person and the qualities of the ad, then matches the numbers. LLMs predict, so Meta can instead look at a person, work out what that person probably wants to see next, then find it and show it to them. Improving ad relevance by only a few percentage points would return billions of dollars.
Most of this work predates LLMs, and the largest monetization gain from the models right now is probably the incremental improvement at Google and Meta.
How do A/B tests work at ad scale?
Two versions of an ad are shown to randomly chosen groups, and the results are compared on clicks or purchases. With millions of impressions, even small differences stand out from random noise. The method dates to statistician Ronald Fisher's randomized farm experiments in the 1920s.
How do embeddings match people to ads?
An embedding turns a person, a photo, or an ad into a long list of numbers, so that similar things end up with similar numbers. Matching then comes down to measuring how close two lists are. Embeddings became widely used after Google researchers published the word2vec method in 2013.
Why is checkable output so valuable for AI?
In coding, a program either passes its tests or it does not, and in math, a proof can be checked by software such as the Lean proof assistant. That gives training a clear signal about whether an answer was right. Researchers call this training on verifiable rewards.
More from Invest Like the Best with Patrick O'Shaughnessy
