AI吞噬世界

November 2024www.ben-evans.com “In my lifetime, I’ve seen twodemonstrations of tech thatstruck me as revolutionary:the GUI and ChatGPT” Bill Gates, March 2023 (Microsoft took 20 years toreach a $150bn valuation*) Huge interest ChatGPT reached mainstream consciousness with unprecedented speed Huge interest, limited use so far Glass half full / half empty - lots of people have tried it, but few found it useful, so far Huge interest, limited use so far Enterprise software takes time, and come with early disappointments “Wait a minute?” An investment surge ahead of a proven market prompts the obvious questions Welcome to the hype cycle? It always takes time to reach the “Pleacom ofProductity” Benedict Evans–– November 2024 Beyond the noise - the next platform shift After the web and smartphones, all tech gets built around generative AI But everything is wide open We don’t know the answers - we’re still working out the questions How far will this scale? Foundational question: will LLMs keep scaling? We got these results by using more and more data and compute - will that keep working? “I don’t know that I wouldlook at the training trendsand extrapolate threeorders of magnitude aheadblindly from today” “Just give yourself theoption that what’s beenhappening for six yearsnow is going to continue” Kevin Scott, Microsoft CTO Sergey Brin Is it slowing down right now? A sudden blip, or something more? Scaling is hard Scaling these models has practical challenges and will take time, even before the science questions We’re going to find out - if only for FOMO An asymmetric bet - over-spending capex has less downside than losing the next platform? “We have no moat” Internal Google memo, May 2023… “The models that are in training now …are closer in cost to $1bn… and then Ithink in 2025 and 2026, we’ll get moretowards $5bn or $10bn” Dario Amodei, Anthropic CEO, April 2024 “The amount of compute needed to trainLlama 4 will likely be almost 10x morethan what we used to train Llama 3 - andfuture models will continue to growbeyond that” Mark Zuckerberg, July 2024 Training Meta’s Llama 3.1 SOTA model Unprecedented computational (and capital) intensity If the moat is capital… Nvidia can’t keep up with demand - for now (but semiconductors are a cyclical industry) The capex surge ~$220bn of capex at the big four in 2024, up $90bn from 2023, and all expect more growth in 2025 From the edge to the centre Remember when telcos built the infrastructure and software had no assets? Here come the bankers A flood of capital creates opportunities for capitalists And everything is still changing under our feet All the science and engineering questions are still moving The last time software hadmarginal cost The consumer internet model of ‘launch free, goviral, work out revenue later’ doesn’t work withtoday’s LLM cost model Huge efficiency gains Many more models, many more specs to measure, and all converging on a commodity The ‘feeds & speeds’ phase of the market Many more models, many more specs to measure, and all converging on a commodity Better or cheaper - plus open source Best, or 90% as good at 5% of the price Model quality (LMARENA) versus price per million tokens (USD), November 2024 “Everyone in tech is givingsomeone else’s businessmodel away for free” “Everyone in tech is givingsomeone else’s businessmodel away for free” The great model boom of 2023-2024 Better, faster, cheaper - pick two “If anything in this life is certain, ifhistory has taught us anything, it is thatyou can kill anyone.” Michael Corleone “If anything in this life is certain, ifhistory has taught us anything, it is thatyou can kill anyone.” Semiconductors are cyclicalCommodity tech goes to marginal cost “If anything in this life is certain, ifhistory has taught us anything, it is thatyou can kill anyone.” Semiconductors are cyclicalCommodity tech goes to marginal costAnd every new tech produces a bubble How is this useful? 2013: how is MachineLearning useful? “That’s clever… but so what?” 2013: how is Machine Learning useful? What’s the right level of abstraction to understand this? 2023: why is GenerativeMachine Learning useful? “That’s clever… but so what?” Benedict Evans–– November 2024 And whatcan’tit do? ‘Answer this question’ ‘What do answers to questions like this tend tolook like?’ Benedict Evans–– November 202440 Error awareness is still limited And the apparent fluency of text output conceals the nature of the model behind Those who agree that “Generative AI alwaysproduces factually accurate answers” Handling ‘errors’ in a probabilistic system Is this a science problem, or a use case and design problem? Can you use this for generalsearch? need?(Alphabet had $56bn of FCF in the last 12 months, so it’s worth finding out) Benedict Evans–– November 2024 2024: how are LLMs useful? What’s the right level of abstraction to understand this? AI gives you infinite interns