软件2.0：用新知识交付LLMs.pdf

1、2024 Databricks Inc.All rights reservedSharon Zhou,CoSharon Zhou,Co-founder&CEO,Laminifounder&CEO,Lamini1SOFTWARE 2.0:SOFTWARE 2.0:SHIPPING ENTERPRISE SHIPPING ENTERPRISE LLMS WITH NEW LLMS WITH NEW KNOWLEDGEKNOWLEDGESharon Zhou,PhD+4b Stanford CS Faculty in Generative AI Stanford CS PhD in Generati

2、ve AI(Andrew Ng)MIT Technology Review 35 Under 35,for award-winning research in generative AI Created largest Coursera courses(Generative AI)Google Product Manager Harvard Classics&CSOur founders are leaders in generative AI and production LLMs.Gregory Diamos,PhDCo-founder&CTO MLPerf Co-founder,indu

3、stry standard for ML perf Landing AI Engineering Head Deployed LLM to 1+billion users;lead 125+engineers;scaled GPU cluster from 0 to 100K 14,000 citations:AI scaling laws,Tensor Cores NVIDIA CUDA architect-as early as 2008 Georgia Tech PhD in Computer Engineering2024 Databricks Inc.All rights reser

4、ved2024 Databricks Inc.All rights reserved3 3LAMINI:LLM FINETUNING&INFERENCE FOR ENTERPRISELAMINI:LLM FINETUNING&INFERENCE FOR ENTERPRISEFactual LLMs.Up in 10min.Deployed anywhere.Factual LLMs.Up in 10min.Deployed anywhere.Factual accuracy with Memory Tuning,cutting hallucinations 10 x from 50%to cu

5、tting hallucinations 10 x from 50%to 5%5%100%guaranteed100%guaranteed JSON output52x more queries per second52x more queries per second than vLLMRun anywhereAirAir-gapped instancesgapped instancesAny cloud VPCsLamini cloudNvidia or AMD GPUs2024 Databricks Inc.All rights reserved2024 Databricks Inc.A

6、ll rights reservedSoftware 2.0&Enterprise 2.0Introducing Lamini Memory TuningLamini Memory Tuning1.Research breakthrough for removing hallucinations 2.Technical details&how to build with it3.Case Study with a Fortune 500 companys LLM agent4.Additional applications4AGENDAAGENDASoftware 2.0:Shipping E