tldr; There is a large language model trained on me. It is increasingly capable of responding to questions on cybersecurity, international security, counterterrorism, technology, and entrepreneurship. You can access it here: ALTzero Project – MattGPT
Setting the Stage
Two decades ago I realized that in the future technology might enable me to build an interactive knowledge base to query documents I’ve collected on a variety of issues. In preparation for that future capability, I created a directory called knowledge and became a bit of a data hoarder collecting thousands of documents on numerous issues within the security and technology domains.
When ChatGPT rolled out last fall, the capabilities of AI technology made a surprising great leap forward. Anyone that connected with ChatGPT found it remarkably capable and fun to interact with, yet it had one key flaw in that it was trained off data collected on the internet. As my longtime friend Bob Stratton has remarked for 30 years, the internet is a bad neighborhood. So while ChatGPT was a lot of fun and even useful, it was hard to trust.
Domain Context is Key
At OODA we quickly shifted our future focus on the value of LLMs like ChatGPT to use cases where the training data is drawn from domain specific content. For example, if you are a lawyer, you want your LLM trained on the law or a doctor wants it trained on medical journals, etc.
As it relates to the core issues we focus on with the OODA Network, the OODAloop.com site becomes a great trusted domain source. Additionally, I’ve personally written on these issues for over three decades and have been collecting domain specific trusted content that influenced my analysis for over twenty years. The ability to load this domain-specific data into an LLM could produce interesting results.
Over the past two weeks, I’ve been working with Delphi.ai to build an LLM based on me. The primary personality and expertise of the AI is based on what I’ve written, interviews I’ve given, presentations I’ve delivered and direct source material drawn from blog and social media posts.
What makes this unique is that the model is also trained on thousands of documents I’ve personally provided it from my knowledge database that include historical reports like the Ware report, the highly impactful Chinese look at Unrestricted Warfare, hundreds of threat intelligence reports, intelligence estimates, international government reports, technology briefs, and hundreds of other documents that have influenced me as a professional.
The result is a surprisingly capable AI that frames things as I might drawing on knowledge that I’ve consumed and providing footnotes on the sourcing that influenced its response. For example, you can interact with it in a variety of ways:
- As a CISO looking for advice on managing cyber risk. You can get specific and ask in the context of a NIST framework or what controls you should focus on based on Mitre ATT&CK.
- As an analyst looking to study a particular threat actor dynamic, terrorism risk, or geopolitical issues.
- As an academic researcher seeking historical context on a variety of risk issues.
- As an entrepreneur wanting to talk through growing a business.
- As a job-seeker to help prepare for an interview.
- As a researcher looking for a way to augment your efforts.
- A technologist looking to explore concepts like AI, blockchain, or quantum computing.
In a week of testing with colleagues, the AI has produced pretty remarkable results navigating a variety of complex issues. Over the coming weeks it will grow even more capable as I continue to feed the model with training data including the OODAloop.com website, the old Terrorism Research Center analysis and some of the remaining 2000 documents in my knowledge folder. With enough time, I can also start to weight some documents as being more valuable than others and thus further fine tuning the model’s responses.
In the past week OpenAI released a custom GPT feature that will let folks develop and train personal GPTs. While this approach is less capable that what Delphi has developed in that you can only lightly influence the model by uploading some documents, it marks a rapid shift forward into domain specific AI technology development. Over the next year, we will see the rapid deployment of persona, role, and organization specific AI technologies.
As domain specific AIs emerge, we will also see the emergence of oracle aggregator AIs that soley exist to query other trusted models. A council of AI elders if you will.
The persona-based AI will have increasing utility and soon we will be able to launch private AIs that are trained on our more sensitive personal information that serves as a personal repository. I articulated the value of these private AIs back in 2013 with my essay “I am big data and so are you” I look forward to a private AI that has access to my email, internal documents, etc.
Also, we’ll likely see the rise of unauthorized or adversary-focused AIs. For example, a public persona might have an AI built for them without their involvement or imagine the interest in asking a Putin.AI questions and yes, it inevitably means we’ll be confronted with Hitler.AI as well.
Regardless, the brave new world of persona and domain specific AI is here to stay.
You can access the beta of MattGPT here: ALTzero Project – MattGPT