How to get your AI outputs to be 42% more accurate:
Example: People complain ChatGPT can't write essays as well as humans because the results are too generic and inaccurate.
I know what you're thinking - it depends on how well you craft the prompt, right?
Yes, but that's only one part of it.
Here's the bigger point (full credit to Matt Berman):
👉 If you ask GPT to do the whole job, then you are more likely to have quality diminished by hallucinations.
The better way to work is agentically, meaning you break up the assignment into individual tasks and have different LLMs act as agents assigned to each.
So, for this example:
👉 One writes the essay outline
👉 Another handles the web research
👉 Another writes the first draft
👉 Another "reflects" on that draft
👉 Another calls up the AI tools needed
- e.g., web scraping tool, stock ticker tool, etc.
👉 Another revises that draft
👉 Another "reflects" on that draft
and so on...
Now you've got (in theory) 7 independent AI "minds" working iteratively on a task, each taking a fresh look and honing as needed.
The chart below shows how drastically that improves performance.
As Andrew Ng shared in this example, a classic math assignment, each added agent improved outcomes, so that even ChatGPT 3.5's answers improved from 48% to 95% (!!) accurate:
(The left side "Zero Shot" is 3.5 by itself, and as you move further to the right, more agents are added)
To be fair, this isn't completely easy to do yet.
And agents today are still finicky with their own unique issues.
But increasingly, there are tools fine-tuned for specific tasks becoming available.
💡💡 The key takeaway💡💡:
The future of AI is agentic, and the faster we start thinking and working with LLMs that way, the greater success we'll have.
...
Now what did I miss or get wrong?
hashtag#ai hashtag#chatgpt hashtag#openai hashtag#marketing
...show more