원문 보기
there's way few silicon startups and there's way few white truffles instead it's everything is the bologna in the middle which is random people talking about some random model that's just gonna again become highly commoditized you have to remember all these models are open sourced and none of them mean anything in the absence of the data you give it to train them this week about the insane fundraising in the generative AI space Oh this is stuff like Chachi PT stable diffusion you've heard these names and there are now 500 generative AI startups according to this report that tracks with what I'm seeing in the early stage and not counting the open AI 10 Billy from Microsoft investment donation rev share round trip whatever that is they have so far collectively raised more than 11 billion dollars the article included this chart which you can see if you're on our YouTube channel and just tons of folks working in audio image and text so we're basically looking at the multimedia basically revolution of PCS in the 90s occurring again and a complete platform platform shift douglioni from Sequoia one of the greatest investors in history of Silicon Valley had this to say and we will comment on the other side of this 50 second clip from Doug I actually think that AI is the next platform shift in the same way that mobile was the one before internet was the one before so I think AI is real but I said earlier we're going to overestimate it in a short term we're going to invest in everything in the same way that in 1999 we invested in everything but then Google came out of that or Facebook came out of that so I think you have to have a good head on your shoulder where you don't practice fomo where you don't chase every company AI is real AI is the next platform but how do we not invest in everything that walks how do we make certain Investments based on Market Maps based on thought processes that are more rational and not do every investment just because every other Venture firm is doing everything best all right shamath what do you think of this massive influx I think it's important to think about what the incentives are as Charlie Munger says show me the incentive and I'll show you the outcome I posted it into our group chat Credit Suisse sent all their private banking customers an offer they are now offering 6.5 for a three-month t-build 6.5 percent and if you go back to what we talked about before when the risk-free rate is somewhere north of five or five and a half percent and banks are willing to give you six and a half percent in the short term you have to generate more than three times that to make an investment make sense when you're investing in the long term so if three month money is going to pay you six and a half or seven if you're going to invest for 10 15 years which is what you need to do typically for a startup we need to get 20 to 25 percent so there's going to be a lot of pressure for Venture investors to put the money to work because otherwise they're LPS they're limited partners the people that give them the money will have this attitude that goes somewhere along the following lines okay I've committed to you why aren't you investing because otherwise I can get paid six and a half percent on the front end and so this is becoming very problematic what am I paying you for what am I paying you two percent a year in management fee for so I think what happens unfortunately is the opposite of what Doug says I think good investors will try to follow Doug's feedback and advice and the ones that have a track record of distributions of DPI can do more of it than not but I think most people will be under pressure to deploy the capital and so the 500 odd companies Jason you mentioned we'll get a lot of it and it'll get torched because most of them probably won't amount to much of anything in the short term you will create way too much correlation and you will have zero time diversity which as we've learned is a recipe for terrible returns time diversity being hey you deployed all this web 2 Capital web 3 Capital sorry crypto whatever in this very short period of time you overpaid Brad yeah to build on that sax this is a perfect inefficiency you know when you see it from the outside you're like well this is super inefficient why so many companies if you free your mind and just say their experiments two or three person experiments 500k to a million and a half in that first stage when I invest right before you do when you do your series a is at five or ten million milestone-based funding is back in the tech industry and that was something Tremont that we lost for a little while there people would just come out and they would raise a series B out of the gate no product Market fit et cetera now what we're seeing is people are raising that 12 to 18 months they got their backs up against the wall speaking of tough love which you reference acts that tough love of hey you have to hit the next Milestone what did you accomplish with the 1.5 million in order to get the 10 million in order to get the 30 million that 500 is going to go 10x they'll be 5 000 of these startups but it'll quickly Whittle down Winner's people go through this Milestone based funding system in Silicon Valley we haven't given enough time for a logical framework of investing to develop which also is tied to a logical framework for entrepreneurship to emerge we're just way way way too early so the thing with all of these language models is that they are Grist for the mill and we talked about this before if it was a highly proprietary asset you would have never sold 49 of it to Microsoft for 10 billion dollars because you would assume that it was worth a trillion dollars so it's a huge tell on the part of openai their deep understanding and they understand it better than anybody else that it's a bit of a capped upside so what is uncapped I guess is maybe the better question well if you look at the 1849 Gold Rush as an example the people panning for gold in my opinion are the people building language models today they're going to come and they're going to go but who's gonna win well the pick and shovel providers won Levi Strauss one so what is the equivalent of that today I think it's at the Silicon layer because you need to really re-architect how compute will actually work in a world of all of these models those folks will get paid if you look at AMD and Nvidia they've been getting paid for years they continue to get paid they probably will continue to get paid even more and so folks that actually take a step into doing something hard and difficult in AI like custom silicon could get rewarded and then there's what I call the white truffles what are these unique Alba white truffles these singular sources of data that when used in reinforcement learning make your output just zing and that's where I think Facebook is an obvious white truffle Cora is a white truffle but there are a lot of startups that could become white truffles if they gather enough data and that's like a pretty reasonable framework and so in that framework if you applied to today there's way few silicon startups and there's way few white truffles instead it's everything is the bologna in the middle which is random people talking about some random model that's just going to again become highly commoditized you have to remember all these models are open source and none of them mean anything in the absence of the data you give it to train on 100 Brad
Comments 0