Andrew Ng, the founder of Google Brain, Baidu Chief Scientist and general AI legend, has just released a youtube video "What's Next for AI Agentic Workflows". He begins by saying something we know well at Agentico:
"AI agents are a trend everyone building AI should pay attention to."
Agentic Reasoning Design Patterns
He gives four design patterns for achieving reasoning in AI agents:
Reflection
Tool use
Planning
Multi agent collaboration
These are all patterns we've used here at Agentico for about 6months now, glad to see they're becoming popular.
Planning
The fun new focus is 'planning', it enables new product development ideas, new revenue streams by using an LLM to manage a sequence of machine learning tools to get a job done, recovering from unanticipated errors as they occur. This is what we would expect of a real agent making arrangements on our behalf.
Research Agents
Using agents to conduct marketing research or product development research is already the most popular request for us at Agentico. I was interested to see Andrew Ng also say he uses agents in his daily workflow to seek and summarise research. Although it is not perfect yet.
"The set of tasks that AI can do will expand dramatically because of agentic workflows"
Cheaper if Faster
Andrew evidences how arranging GPT3.5 into agents makes it more effective at coding than GPT4, meaning its much cheaper! He also finishes by saying that "Fast token generation is important, more tokens from lower quality LLM can give good results".
I think he's alluding to what is now possible with Groq. Fast inference with cheap models allows trial and error at scale. In a previous blog we discussed how powerful it would be to have AI data scientists write many solutions in parallel, then simply select the best. I know Google is working hard on such 'brute force' solutions, they allow AI to be genuinely novel and creative.