I will unveil the findings of an exhaustive study on the current landscape of AI in business, its market dynamics, and the associated risks. This research, rooted in the Natural Language Processing (NLP) analysis of media discourse, utilised leading language models (LLMs) and drew from a variety of data sources. Beyond the empirical picture, I will dive into an analysis of the various risk levels inherent to the AI business. Additionally, I will provide insights into the potential direction of the AI market in the near to medium-term future. The primary objective is to clarify our present understanding and control of the AI business, while also highlighting the uncharted areas we might encounter in the future.
31. BloombergGPT
• Trained on AWS SageMaker
• ~700 billion tokens dataset, only 50 billion parameters
• 64 x p4d.24xlarge instances 64 x 8 Nvidia 40GB A100 GPUs 512
GPUs total
• ~$33 per instance per hour $2112 per hour for complete setup
• Trained for 53 days ~$2.7 million (~$1 million with spot-pricing)
37. Not too many domain-specific LLMs:
BloombergGPT, Med-PaLM2, ClimateBERT,
BioBERT, KAI-GPT, FinGPT…
38. On the other hand:
• OpenAI introduced GPT-4 Turbo with 128K context,
• JSON response format,
• Reproducible outputs,
• Stateful API (i.e. conversation memory)…
• … and Anthropic Claude 2.1 now has a 200K context window, search and
retrieval capabilities over a variety of knowledge bases (Elasticsearch, vector
databases, web search, Wikipedia), tool use, etc.
AIaaS is targeting the DevsAI business
40. AIaaS
(high cost) Infra
(high cost) Training
(high cost) Data
(high cost) Inference
DevsAI
Infra (low cost)
Knowledge (moderate cost)
Compute (high cost) ?
The emerging market risk is in the P(match):