从人工智能产品成功上市中汲取的关键经验教训（超越炒作阶段）.pdf

1、quick bias checkBuilt the architecture and led engineering at SoundCloud from 7 million to 250 million MAU Led product engineering at DigitalOcean from a single product to a cloud platform Replatformed and rearchitected Meetup and SeatGeek to web-scale Pioneer in microservices and modern distributed

2、 applicationswhat was OutropyReleased 6 months after ChatGPT Utilized GPT 3.5,4,and Llama models Started as a Slack bot,added a Chrome extensionone of the very first enterprise AI startups to grow to 10,000 userswhat we were up againstyet,the product was miles ahead of the big players*if you trust m

3、y benchmarkhow the hell two guys and a dog in Brooklyn built this?but the product failed miserably.users were only interested in learningdeveloping a hypothesis for why most AI products suck the three ways we build AI a.Twitter-driven Development b.Its just like Data Science c.Build like product eng

4、ineering projectsbuilding blocksworkflows*are systems where LLMs and tools are orchestrated through predefined code paths.*aka Inference Pipelinesagentsare systems where LLMs dynamically direct their own processes and tool usage,maintaining control over how they accomplish tasks.workflows:what AI ve

5、ndors sellwhat actually worksevery feature we shipped be likeworkflows*are systems where LLMs and tools are orchestrated through predefined code paths.*aka Inference Pipelinesagentsare systems where LLMs dynamically direct their own processes and tool usage,maintaining control over how they accompli

6、sh tasks.data pipelines?agents are lousy microservicesStateful Operation Non-deterministic Behavior Data-Intensive with Poor Locality Unreliable External Dependenciesagentsare systems where LLMs dynamically direct their own processes and tool usage,maintaining control over how th