The auction code was worse than a Google-only black box, when I was in Ads. The code itself was in a "high intellectual property" subdirectory that most engineers could not access.
There's an elaborate economic rationale for second-price auctions, including that they avoid "winner's remorse." Leave it to the late-stage "improvers" to say, "hey, we can capture some of that gap between first and second place!"
>There's an elaborate economic rationale for second-price auctions, including that they avoid "winner's remorse." Leave it to the late-stage "improvers" to say, "hey, we can capture some of that gap between first and second place!"
The reason Google uses second price auctions is because first price auctions incentive constant bid changes. The stability increases the money Google makes.
edit: For more context, and ignoring the whole generalized second price auction parts, the revenue from first and second price auctions is the same. In theory with optimal bidding and the usual assumptions of auction theory papers. The difference is that in a second price auction your optimal bid is independent of the other bidders. In a first price auction your optimal bid is dependent on the expected distribution of other bidders. In theory with all the usual auction theory assumptions. In reality with online auctions this meant a lot of churn in bids and a preference for bidding more conservatively. Which isn't good for advertisers or the revenue of the company running the auction. Second price auctions mostly solve this especially if you adjust the minimum auction price (ie: reserve price) per auction.
edit2: A great way to get more money is to have advertisers think they're bidding in a second price auction while actually running something closer to a first price auction. Which is essentially what the article talks about.
I’ve been in the field for 20 years and I know the difference between what companies do and what theoretical papers say they do.
As the other person said, bids repeat across multiple auctions on the same keyword and even in a generalized second price auction optimally your bids are adjusted based on the expected dynamics of each auction. Based on past information and so on. In a first price auction you’d keep lowering your average bid until it’s on average right above the next bid. The next bidder would then increase their bids. You’d then do so as well. Etc. Etc.
You’re moving the goal post from not knowing auctions to not knowing Google. Will the next question be if I made the decision at Google for this approach when they were competing with Overture?
I always thought that HIP vs. rest of google3 would make a great case study for developer experience/productivity research. Fewer eyes, emergent private style guide variant, etc.
> There's an elaborate economic rationale for second-price auctions, including that they avoid "winner's remorse." Leave it to the late-stage "improvers" to say, "hey, we can capture some of that gap between first and second place!"
Are there any more examples or reading around this concept?
Mathematically, all fair auctions provide the same optimal price to the seller. It's called the Revenue Equivalence Theorem.
Google's auction is not fair in the sense that Google is both the seller and the auctioneer. Squashing enters the situation when the second price bidder has an ad that is more likely to be clicked than the first price bidder and therefore has additional revenue potential (known only to Google) beyond the revenue offered by the bid. Google then assigns this additional expected revenue to the value of the second price bid and declares this to be the market value of the bid. I see where their argument comes from, but it's a pretty lousy one.
That’s what happens when you don’t have open source and open protocols. I see people on HN constantly complaining but very little support for distributed systems, especially byzantine fault tolerant with proper incentives, why? Incentives are key to such distributed systems working, but it sounds too much like Web3? FTX? Or what? I guess many people don’t want to see viable alternatives appear. So then all you can do is keep complaining about black boxes and secret actions.
I believe he is mocking the average HN commenter, who simultaneously complains about these types of opaque-black-box systems but attacks transparent distributed systems.
That's how I interpreted it. Everyone seems to want the advantages of the FOSS model for general benefit, but want firms to have their advantages at the same time, and not acknowledge the inherent adversarial aspect of that.
There's an elaborate economic rationale for second-price auctions, including that they avoid "winner's remorse." Leave it to the late-stage "improvers" to say, "hey, we can capture some of that gap between first and second place!"