I had worked out some costs. If you want to build an index-based search engine that is at least competitive on coverage, if not relevance, it would cost you abo

I had worked out some costs. If you want to build an index-based search engine that is at least competitive on coverage, if not relevance, it would cost you about $2B/yr for at least a period of 4 years just to boot things up. So, that is at least $8B of upfront investment. This is similar to the problem of if someone wanted to match TSMC on 5nm feb. This cannot be achieved by traditional VC funding and startup model. Remember, search is not just web search. It includes vast array of things such as maps, images, videos, knowledge datasets, discussion groups, real-time news and so on. Each of these segments is an enormous effort on its own requiring massive capital investments. Relevance algos these days are mostly driven by thousands of tweaks, ML models and legions of rules. It’s a complex beast that takes thousands of PhDs and years to perfect. Even than performance is pretty subpar than most expectations. So, what chance do you have as a new entrant in same game? The key is that you don’t want to play same game if you want to win. A lot of companies have came and gone announcing themselves as Google competitor and trying to play same game. Things like DDG survive only because they can lower the COGS by offloading real work to other people like Bing which itself is fine example of what you might be able achieve if you only had a tiny fraction of Google’s budgets. If you consider dollar for dollar capital investment, Bing is actually quite good. This is not to say there will never be a viable competitor for Google search. I just don’t think it can be through the traditional framework of index serving. No one really knows what other alternatives can arise in future. One very possible thing is language models. If we can figure out how to scale up serving of massive language models that effectively “memorizes” whole index in them, they can provide quality, capabilities and experience that cannot be matched by simple index serving. If this is viable route, I think we are still at least 3-5 years away.