Voicebot + Synthedia
Amazon Web Services (AWS) rolled out Amazon Bedrock earlier this year to make it easier for businesses to employ generative AI models on demand. Atul Deo is General Manager/Director, Product and Engineering for Amazon Bedrock. He joined me to go in-depth around AWS AI services broadly, and specifically Amazon’s strategy around Bedrock and generative AI. We even talk about customer obsession, which is always a favorite topic at Amazon, and when different types of generative AI architectures make sense.
Please
👍***LIKE and ✅***SUBSCRIBE to the channel.
🎟 Also, subscribe to the Synthedia daily newsletter on generative AI news, data, trends, and analysis: https://bit.ly/synthedia
🎧 Listen to the Voicebot Podcast
https://voicebot.ai/voicebot-podcasts/
🔵 Follow:
https://twitter.com/bretkinsella
https://twitter.com/voicebotai
https://twitter.com/soapboxlabs
Timestamps:
00:00 Introduction Clip LLM/NLU Supervision Framework
01:03 Welcome Atul Deo to the Voicebot Podcast
03:50 The Transcription Market and ASR
05:57 Turning Audio Into Text Unlocks Contact Center Insight
07:33 NLP and NLU’s Significant Distinctions from LLMs
08:59 Identifying Intents and Slots Was Rigid & Painful
12:00 Understanding Content in Context Requires Compute Cost
15:31 AWS and the Early Days of Machine Learning
16:45 Amazon Bedrock’s Model Development Journey
18:20 How Amazon’s Customer Obsession Figures In Model Design
19:48 Why 3rd Party Model Choices Is Essential
22:47 What Customers Want From Each Model
24:00 AWS Is Not a One Size Fits All Customer Experience
27:18 Adding An LLM to an NLU-Based System
29:15 Hybrid NLU LLMs
29:55 Atul Deo’s LLM/NLU Supervision Framework
34:30 Amazon’s Anthropic Investment is About Quality
35:55 Amazon’s GenAI Strategy – Infrastructure
38:36 The Challenge of Data Hungry Compute
40:04 The Demand For Inference Requires High Performance
41:55 Production Throughput and Reliability With Bedrock
44:05 Model Latency Norms Across the Industry
45:08 Instruction Tuning
48:04 Model Differentiation
49:34 Titan Custom Embeddings
52:53 Model Training Parameters Going Forward
59:10 Model Customization Approaches
01:03:18 Domain-Specific vs. General Purpose Models
01:05:56 How Many Applications Will Be Built On These Models?
Source
