AI like electricity is a general purpose technology
New AI technology creates many new opportunities to build new applications
The AI Stack - Where are the biggest AI opportunities?
semiconductor microchips - NVIDIA, INTEL, AMD
Cloud infrastructure - Amazon, Google, Microsoft, snowflake
Foundation models - OpenAI, Anthropic, Google, Meta
Applications - endless
generative ai is enabling ML product fast development
supervised learning - get labeled data 1mo - train ai model on data 3mo, deploy run model 3 months
LLM-based development - specify prompt in days - deploy model
Consequences of fast development
fast experimentation as a path to invention
move fast and break things NO
move fast and be responsible
agentic ai workflows
from a technical perspective
non-agentic workflow (zero-shot)
please type out an essay on topic X from start to finish in one go, without backspace
agentic workflow eg
write an essay outline on topic X
do you need any web research?
write a first draft
consider what parts need revision or more research
revise your draft
agentic AI thinks and reflections output and uses it as input again for next step
agentic workflows deliver better results and accuracy in tasks
Planning - decision tree and decide on steps for task
reflection
tool use (api calls)
multi-agent collaboration
1. reflection with llms
take baseline performance and lift it to better performance
e.g. please write code for (task)
ask the LLM to improve the for prompt and give feedback to LLM
use critique and use multi-agents to prompt an LLM to be a coder and another a code reviewer
different personas/agents as a reflection design pattern
2. tool
3. planning/reasoning
long instruction
complex request, pick sequence of action to delivery on a complex task
4. multi-agents
has been demonstrated there's an accuracy improvement
one LLM to specialize
eg. multiple CPUs, pick a task and break it down
hire different agents to pick different pieces
4 agentic design patterns.
large multi-modal based agents
vision agent
AI stack:
semiconductors:
nvidia, amd, intel
cloud infrastructure:
aws, google cloud, azure, snowflake
foundation models:
openai anthropic meta
agentic orchestration layer:
langchain crew.ai AG
eaiser for developers to build applications on top
applications:
credo.ai, speechlab, woebot health
four AI trends:
1. agentic workflows consumer a lot of tokens, and will benefit from faster, cheaper token generation (e.g. Sambanova, Cerebras, Groq)
2. today's agents are built by taking LLMs trained to answer questions and retrofitting them into an iterative wofklow. More LLMs will be finetuned for use in agentic workflows, such as to use tools, to plan/reason (e.g. openai o1) or to use
3. data engineering's important is rising, particularly on management of unstructured data (text, images).
4. The text processing revolution has arrived. The image processing evolution is coming, and will enable many now visual AI applications in entertainment and self-driving, etc.
Comments