Bob Gourley
The age of continuous crisis has placed incredible demands on intelligence analysts. Automation has not kept up in helping analysts meet the ever expanding demands for quality intelligence. Search tools for the enterprise are supposed to help intelligence analysts, but almost all broken and underperform, especially over unstructured text.
We have been reporting on hopeful new advances that can address some of these challenges. For example, new approaches to Natural Language Processing hold great potential to help (See: What Leaders Need to Know About the State of Natural Language Processing). Watching these developments closely has led us to believe very soon every enterprise will have high end AI based tools to empower their intelligence analysts with the right information for them to do their jobs.
Like the rest of the Internet we have also been evaluating the new OpenAI capabilities like GPT-3 and ChatGPT, which go beyond search to produce results to requests that can seem conversational. We have been participating in dialog with many in the OODA Network on these topics including exchanging prompts and replies from these systems that help show some of the incredible potential of these large language models.
To continue to explore the relevance of these tools to intelligence missions we built a web application designed to leverage the power of GPT-3 to the domain of intelligence analysis. The application is available for use at UnrestrictedIntelligence.com
This application is set up to allow any user to ask any question or make a request for intelligence. In response the system delivers a succinct intelligence product based on the information GPT has been trained on.
Here is a sample result when the system was tasked to “Explain in simple terms the threat from China to the US”:
This report is very basic but that is what we asked for! It certainly does not contain the wisdom of an analyst who has a deep expertise in China. But overall not bad for a robot. Consider the fact that reports like that can be produced on any country in the world in a matter of seconds. And for those that want more it is easy to interact with this system and ask iterative questions to extract insights.
Here is a sample report when asked about the threats to new manufacturing businesses in Brazil:
You can ask the Unrestricted Intelligence app any question about threat actors, future threats, adversary capabilities, adversary history. This is not limited to tasks of interest to the national intelligence community. Open source analysts and cyber threat analysts can also leverage this tool to produce fast analysis.There are huge weaknesses in this system. The data this is trained over is all older than September 2021. And the models of GPT-3 can fail in strange ways producing results that are just wrong. Additionally, just like leaders can ask bad questions of intelligence analysts, users of this system can ask bad questions that deliver ridiculous results.
All this means the reports from the Unrestricted Intelligence site really need to be assessed by humans before any decisions are made over them.
Keep in mind that GPT-4 is expected to come out in the spring and is believed to be at least 100 times more powerful. This will make this site even more helpful to all source analysts and decision-makers in need of intelligence support.
How does it work? The questions a user asks are wrapped in some guidance and then passed to OpenAI via their API for GPT-3. The guidance that goes with the query requests GPT respond in the language of an intelligence professional, using, to the greatest extent possible, the standards for intelligence products promulgated by the Office of the Director of National Intelligence (ICD203). GPT is also tasked to use models of intelligence analysis like analysis of alternatives and to discuss potential ways that bias or deception may impact the analysis. The system also suggests follow on questions to ask GPT.
What comes next? GPT-4 is expected to come out in the spring and is believed to be at least 100 times more powerful. This will make this site even more helpful to all source analysts and decision-makers in need of intelligence support.
For now we could use some help testing the system out. Please think of hard questions to ask the system in the domains of geopolitical intelligence, cyber threat intelligence, competitive intelligence and any other subject dealing with advanced analysis. Would really appreciate your feedback.
No comments:
Post a Comment