The videos embedded in this post are best viewed on steveblank.com
International Policy students will be spending their careers in an AI-enabled world. We wanted our students to be prepared for it. This is why we’ve adopted and integrated AI in our Stanford national security policy class – Technology, Innovation and Great Power Competition.
Here’s what we did, how the students used it, and what they (and we) learned.
Technology, Innovation and Great Power Competition is an international policy class at Stanford (taught by me, Eric Volmar and Joe Felter.) The course provides future policy and engineering leaders with an appreciation of the geopolitics of the U.S. strategic competition with great power rivals and the role critical technologies are playing in determining the outcome.
This course includes all that you would expect from a Stanford graduate-level class in the Masters in International Policy – comprehensive readings, guest lectures from current and former senior policy officials/experts, and deliverables in the form of written policy papers. What makes the class unique is that this is an experiential policy class. Students form small teams and embark on a quarter-long project that got them out of the classroom to:
select a priority national security challenge, and then …
validate the problem and propose a detailed solution tested against actual stakeholders in the technology and national security ecosystem
The class combines multiple teaching tools.
Real world – Students worked in teams on real problems from government sponsors
Experiential – They get out of the building to interview 50+ stakeholders
Perspectives – They get policy context and insights from lectures by experts
And this year… Using AI to Accelerate Learning
Rationale for AI
Using this quarter to introduce AI we had three things going for us: 1) By fall 2024 AI tools were good and getting exponentially better, 2) Stanford had set up an AI Playground enabling students to use a variety of AI Tools (ChatGPT, Claude, Perplexity, NotebookLM, Otter.ai, Mermaid, Beautiful.ai, etc.) and 3) many students were using AI in classes but it was usually ambiguous about what they were allowed to do.
Policy students have to read reams of documents weekly. Our hypotheses was that our student teams could use AI to ingest and summarize content, identify key themes and concepts across the content, provide an in-depth analysis of critical content sections, and then synthesize and structure their key insights and apply their key insights to solve their specific policy problem. They did all that, and much, much, more.
While Joe Felter and I had arm-waved “we need to add AI to the class” Eric Volmar was the real AI hero on the teaching team. As an AI power user Eric was most often ahead of our students on AI skills. He threw down a challenge to the students to continually use AI creatively and told them that they would be graded on it. He pushed them hard on AI use in office hours throughout the quarter. The results below speak for themselves.
If you’re not familiar with these AI tools in practice it’s worth watching these one minute videos.
Team OSC
Team OSC was trying to understand what is the appropriate level of financial risk for the U.S. Department of Defense to provide loans or loan guarantees in technology industries?
The team started using AI to do what we had expected, summarizing the stack of weekly policy documentsusing Claude 3.5. And like all teams, the unexpected use of AI was to create new leads for their stakeholder interviews. They found that they could ask AI to give them a list of leaders that were involved in similar programs, or that were involved in their program’s initial stages of development.
See how Team OSC summarized policy papers here:
If you can’t see the video click here
Claude was also able to create a list of leaders with the Department of Energy Title17 credit programs, Exim DFC, and other federal credit programs that the team should interview. In addition, it created a list of leaders within Congressional Budget Office and the Office of Management and Budget that would be able to provide insights. See the demo here:
If you can’t see the video click here
The team also used AI to transcribe podcasts. They noticed that key leaders of the organizations their problem came from had produced podcasts and YouTube videos. They used Otter.ai to transcribe these. That provided additional context for when they did interview them and allowed the team to ask insightful new questions.
If you can’t see the video click here
Note the power of fusing AI with interviews. The interviews ground the knowledge in the teams lived experience.
The team came up with a use case the teaching team hadn’t thought of – using AI to critique the team’s own hypotheses. The AI not only gave them criticism but supported it with links from published scholars. See the demo here:
If you can’t see the video click here
Another use the teaching team hadn’t thought was using Mermaid AI to create graphics for their weekly presentations. See the demo here:
If you can’t see the video click here
The surprises from this team kept coming. Their last was that the team used Beautiful.ai in order to generate PowerPoint presentations. See the demo here:
If you can’t see the video click here
For all teams, using AI tools was a learning/discovery process all its own. By and large, students were largely unfamiliar with most tools on day 1.
Team OSC suggested that students should start using AI tools early in the quarter and experiment with tools like ChatGPT, Otter.ai. Tools that that have steep learning curves, like Mermaid should be started at the very start of the project to train their models.
Team OSC AI tools summary: AI tools are not perfect, so make sure to cross check summaries, insights and transcriptions for accuracy and relevancy. Be really critical of their outputs. The biggest takeaway is that AI works best when prepared with human efforts.
Team FAAST
The FAAST team was trying to understand how can the U.S. improve and scale the DoE FASST program in the urgent context of great power competition?
Team FAAST started using AI to do what we had expected, summarizing the stack of weekly policy documents they were assigned to read and synthesizing interviews with stakeholders.
One of the features of ChatGPT this team appreciated, and important for a national security class, was the temporary chat feature – data they entered would not be used to train the open AI models. See the demo below.
If you can’t see the video click here
The team used AI do a few new things we didn’t expect – to generate emails to stakeholders and to create interview questions. During the quarter the team used ChatGPT, Claude, Perplexity, and NotebookLM. By the end of the 10-week class they were using AI to do a few more things we hadn’t expected. Their use of AI expanded to include simulating interviews. They gave ChatGPT specific instructions on who they wanted it to act like, and it provided personalized and custom answers. See the example here.
If you can’t see the video click here
Learning-by-doing was a key part of this experiential course. The big idea is that students learn both the method and the subject matter together. By learning it together, you learn both better.
Finally, they used AI to map stakeholders, get advice on their next policy move, and asked ChatGPT to review their weekly slides (by screenshotting the slides and putting them into ChatGPT and asking for feedback and advice.)
The FAAST team AI tool summary: ChatGPT was specifically good when using images or screenshots, so in these multi-level tasks, and when you wanted to use kind of more custom instructions, as we used for the stakeholder interviews. Claude was better at more conversational and human in writing, so used it when sending emails. Perplexity was better for researchers because it provides citations, so you’re able to access the web and actually get directed to the source that it’s citing. NotebookLM was something we tried out, but it was not as successful. It was a cool tool that allowed us to summarize specific policy documents into a podcast, but the summaries were often pretty vague.
Team NSC Energy
Team NSC Energy was working on a National Security Council problem, “How can the United States generate sufficient energy to support compute/AI in the next 5 years?”
At the start of the class, the team began by using ChatGPT to summarize their policy papers and generate tailored interview questions, while Claude was used to synthesize research for background understanding. As ChatGPT occasionally hallucinated information, by the end of the class they were cross validating the summaries via Perplexity pro.
The team also used ChatGPT and Mermaid to organize their thoughts and determine who they wanted to talk to. ChatGPT was used to generate code to put into the Mermaid flowchart organizer. Mermaid has its own language, so ChatGPT was helpful, so we didn’t have to learn all the syntax for this language.
See the video of how Team NSC Energy used ChaptGPT and Mermaid here:
If you can’t see the video click here
Team Alpha Strategy
The Alpha Strategy team was trying to discover whether the U.S. could use AI to create a whole-of-government decision-making factory.
At the start of class, Team Alpha Strategy used ChatGPT.40 for policy document analysis and summary, as well for stakeholder mapping. However, they discovered going one by one through the countless numbers of articles was time consuming. So the team pivoted to using Notebook LM, for document search and cross analysis. See the video of how Team Alpha Strategy used Notebook LM here:
If you can’t see the video click here
The other tools the team used were custom Gpts to build stakeholder maps and diagrams and organize interview notes. There’s going to be a wide variety of specialized Gpts. One that was really helpful, they said, was a scholar GPT.
See the video of how Team Alpha Strategy used custom GPTs:
If you can’t see the video click here
Like other teams, Alpha Strategy used ChatGPT to summarize their interview notes and to create flow charts to paste into their weekly presentations.
Team Congress
The Congress team was exploring the question, “if the Department of Defense were given economic instruments of power, which tools would be most effective in the current techno-economic competition with the People’s Republic of China?”
As other teams found, Team Congress first used ChatGPT to extract key themes from hundreds of pages of readings each week and from press releases, articles, and legislation. They also used for mapping and diagramming to identify potential relationships between stakeholders, or to creatively suggest alternate visualizations.
When Team Congress weren’t able to reach their sponsor in the initial two weeks of the class, much like Team OSC, they used AI tools to pretend to be their sponsor, a member of the defense modernization caucus. Once they realized its utility, they continued to do mock interviews using AI role play.
The team also used customized models of ChatGPT but in their case found that this was limited in the number of documents they could upload, because they had a lot of content. So they used retrieval augmented generation, which takes in a user’s query, and matches it with relevant sources in their knowledge base, and fed that back out as the output. See the video of how Team Congress used retrieval augmented generation here:
If you can’t see the video click here
Team NavalX
The NavalX team was learning how the U.S. Navy could expand its capabilities in Intelligence, Surveillance, and Reconnaissance (ISR) operations on general maritime traffic.
Like all teams they used ChatGPT to summarize and extract from long documents, organizing their interview notes, and defining technical terms associated with their project. In this video, note their use of prompting to guide ChatGPT to format their notes.
See the video of how Team NavalX used tailored prompts for formatting interview notes here:
If you can’t see the video click here
They also asked ChatGPT to role play a critic of our argument and solution so that we could find the weaknesses. They also began uploading many interviews at once, and asked Claude to find themes or ideas in common that they might have missed on their own.
Here’s how the NavalX team used Perplexity for research.
If you can’t see the video click here
Like other teams, the NavalX team discovered you can customize ChatGPT by telling it how you want it to act.
If you can’t see the video click here
Another surprising insight from the team is that you can use ChatGPT to tell you how to write better prompts for itself.
If you can’t see the video click here
In summary, Team NavalX used Claude to translate texts from Mandarin, and found that ChatGPT was the best for writing tasks, Perplexity the best for research tasks, Claude the best for reading tasks, and notebook LM was the best for summarization.
Lessons Learned
Integrating AI into this class took a dedicated instructor with a mission to create a new way to teach using AI tools
The result was AI vastly enhanced and accelerated learning of all teamsIt acted as a helpful collaborator
Fusing AI with stakeholders interviews was especially powerfulAt the start of the class students were familiar with a few of these AI tools
By the end of the class they were fluent in many more of them
Most teams invented creative use casesAll Stanford classes we now teach – Hacking for Defense, Lean Launchpad, Entrepreneurship Inside Government – have AI integrated as part of the course
Next year’s AI tools will be substantively better
Filed under: Customer Development, Gordian Knot Center for National Security Innovation, National Security, Technology Innovation and Great Power Competition |