Following up with (likely a 3 part series) on Digital Twin.
Starts with a high-level system prompt where you feed it details about who you are, where you are, what you want, and what your current deployed tactics are, etc...
Going to show basic examples in a Gemini Gem interface, but OAI and the other houses have their own versions of these (OAI calls them GPTs).
Then we're scaling the complexity curve to a Notebook environment. That gives the ability store our "strategic context" in Drive folders, load them up programmatically, etc...
The final frontier is a custom-built app that visualizes your digital twin, gives you surface area to engage with them + see the work they are driving / orchestrating / looking at themselves.
The 'CEO + departments' framing for a digital twin is way smarter than most personal AI concepts I've seen. The specialization makes sense when you realize general-purpose models are terrible at sustained context. But the real unlock is the inter-twin protocol - that's where network effects kick in and the whole thing becomes exponentially more usefull. Most people are gonna resist this till they see someone else crushing it with 4 extra hours of deep work daily.
This changed the way I look at Agents. Agents were a non-starter for me, they seemed very clunky and have the potential for recurring and expounding hallucinations.
I feel like this method of inter-agent communications within your personal stack is the way to reduce the hallucinations and reduce the compounding effect as well. IE the CEO catches a logical error and orders a corrected report.
I got Grok to help me with construction takeoffs. Usually Grok got within 1% of my takeoff in seconds as opposed to my hour but then again in some cases Grok had us installing tile on... air (we will install tile where there is no drywall return).
I'm hoping these tools can alleviate these weird logical errors.
Great article, thanks for sharing.
Will you provide more details on how you built it in practice ?
Appreciate you reading.
Following up with (likely a 3 part series) on Digital Twin.
Starts with a high-level system prompt where you feed it details about who you are, where you are, what you want, and what your current deployed tactics are, etc...
Going to show basic examples in a Gemini Gem interface, but OAI and the other houses have their own versions of these (OAI calls them GPTs).
Then we're scaling the complexity curve to a Notebook environment. That gives the ability store our "strategic context" in Drive folders, load them up programmatically, etc...
The final frontier is a custom-built app that visualizes your digital twin, gives you surface area to engage with them + see the work they are driving / orchestrating / looking at themselves.
The 'CEO + departments' framing for a digital twin is way smarter than most personal AI concepts I've seen. The specialization makes sense when you realize general-purpose models are terrible at sustained context. But the real unlock is the inter-twin protocol - that's where network effects kick in and the whole thing becomes exponentially more usefull. Most people are gonna resist this till they see someone else crushing it with 4 extra hours of deep work daily.
This changed the way I look at Agents. Agents were a non-starter for me, they seemed very clunky and have the potential for recurring and expounding hallucinations.
I feel like this method of inter-agent communications within your personal stack is the way to reduce the hallucinations and reduce the compounding effect as well. IE the CEO catches a logical error and orders a corrected report.
I got Grok to help me with construction takeoffs. Usually Grok got within 1% of my takeoff in seconds as opposed to my hour but then again in some cases Grok had us installing tile on... air (we will install tile where there is no drywall return).
I'm hoping these tools can alleviate these weird logical errors.
Thanks for the brainfood Matt!