Testing AI Reality

What I Learned at the AI + Expo

I. Welcome to the Future (Sort Of)

As I mentioned in last week’s post, I spent this past week at the AI + Expo. The event, organized by the Special Competitive Studies Project (

SCSP

), sprawled across the Washington Convention Center like a digital fever dream. It was huge, with all kinds of people and lots of options for how to engage.

A funny detail: the convention center was simultaneously hosting a gigantic real estate conference. The vibes were very different—real estate being one of the most in-person, relationship-driven businesses that’s probably least likely to be replaced by AI. I initially got lost and ended up on the real estate side, and you could definitely tell who was going to which conference just by looking at the crowds.

I found myself there as part of the International Strategy Forum Fellowship, a program jointly run by SCSP and former Google CEO Eric Schmidt’s Schmidt Futures. The fellowship is designed to bring people into conversations about AI’s intersection with policy and governance. About a dozen of us fellows have been selected to conduct research that will be published as part of a series called ISF Voices (read the first report on Ukraine’s digital development by my colleague here). I’m co-authoring a piece with a colleague from Serbia who’s a researcher and PhD student in the UK, examining how the sourcing of critical minerals undermines democratic institutions—particularly as AI, EV, and other new technologies drive demand—and what countries can do to strengthen their institutions while recognizing these are necessary technologies people are seeking. We even got to present about our research at the AI + Expo, which was very exciting.

As ISF fellows, we got extra sessions at the beginning and end of each day, then basically got to choose our own adventure in the AI + Expo. The expo had three main tracks:

1. The expo hall—the exhibition floor where companies and organizations had booths, basically doing sales, recruiting, or procurement, and showcasing their work. That’s where you’d see the robots and demonstrations.

2. Speaker series in various spaces throughout the convention center: makeshift areas at the edges of the expo hall, actual rooms surrounding everything, and larger auditoriums for big-name speakers like Eric Schmidt and Condoleezza Rice. These included themed areas like the Africa village (highlighting African AI entrepreneurs), a bookstore section with authors, and career help sessions, which I particularly appreciated given how many people in DC have been displaced and are transitioning industries.

3. The Hackathon, which I chose to participate in because I learn better in hands-on situations—a preference I’ve written about before.

This portrait-drawing robot isn’t taking away anyone’s job anytime soon

II. Walking the Expo Hall

Because I was doing the hackathon, I didn’t get to people-watch as much as I would have liked, but I did manage to spend a couple of hours at the end of the hackathon to focus on the expo hall, though I was pretty exhausted by then. The variety of AI integration across companies was striking—some had been training their models for years, while others seemed to have simply added AI labels to existing products without a clear understanding of how AI enhanced their offerings.

There were robotic dogs, drones, and even a robot that drew portraits, though it was really bad, definitely not taking anyone’s job anytime soon. The robotic dogs were cute, but the moment I started looking at them, I’d get a flashback to the “Metalhead” episode of Black Mirror and immediately feel terror. That image is seared into my mind—every time I see a robot dog, it triggers that association. I honestly don’t want to know what they’re practically for because I’m so terrified after that episode. It just made me want to hug my actual dogs a little tighter.

The crowds were fascinatingly varied. You had super casual people from the startup world or scrappy hacker academia, much more formal government folks, international entrepreneurs who weren’t fully Silicon Valley startup-y and thus a bit more formal, and consultants who were definitely suited up. It was a whole mix that reflected how many different worlds intersect with AI.

There was also a hacker school booth showcasing a free sponsored program that teaches cybersecurity skills—I thought that was very cool and appreciated that there were people selling training and career advice given where many federal workers are right now.

Most people assumed I had much less tech literacy than I actually do, probably because we were in DC, where people generally have less technological literacy than Silicon Valley or New York. Once people realized I knew more than what they assumed I knew—because I’m not an expert, I’m still learning, but for DC standards, I am a little bit more literate—conversations became much more substantive. Except for one older gentleman selling one of the least technologically sophisticated applications, who insisted on explaining SaaS to me after I told him I already knew what it was.

Despite all the controversy right now with DOGE and getting rid of government workers and implementing AI in government to replace them, based on the products that were available that I got to chat with, not a single one would be able to replace a government worker. But many could make their work more effective, especially if you have a seasoned government worker to provide critical oversight.

Most AI companies there had a military or government edge to them, which makes sense because it’s DC and they’re aiming to sell to the federal government and consultants who work with them—that’s more of the market in the city.

Robot dog vs. my actual dogs. The winners are clear (unless it comes to military warfare?)

III. Vibe Coding Through the Hackathon

The hackathon was sponsored by SCSP, OpenAI, Anthropic, Meta, AGI House, and the U.S. National Drone Association. I teamed up with an old college friend who actually knows how to code—he handled backend work while I focused on frontend and user experience design, drawing on my background in amateur web design.

It was refreshing that OpenAI’s prompt was “non-military uses of AI for national security.” I was grateful that OpenAI opened up a path to be more creative outside of the military focus, not because military applications aren’t useful or necessary (though I think there are valid criticisms of military use of technology and AI), but because it’s not the only way. There can be way more uses for AI in this space.

My friend and I, both recent federal government refugees, decided to build an app to facilitate declassifying documents. The concept was straightforward: AI analyzes documents and recommends declassification decisions, but humans retain full control over the process. This felt like exactly the kind of targeted bureaucratic AI solution that could address specific inefficiencies, not at the individual human level, but at structural levels that aren’t necessarily being addressed. I’ll develop this idea more in a future post.

My friend handled the backend document analysis while I focused on frontend design, drawing on behavioral science coursework to create incentives and break up tasks in manageable ways. We built in clear prioritization, plenty of space for humans to edit AI recommendations, and transparent reasoning from the AI system.

This is where I dove into “vibe coding.” My background in web design meant I had a strong understanding of website architecture and user interface design. Because I had taken a class in behavioral science and had web design experience, I was able to prompt the AI pretty well and had the confidence to handle many tasks myself.

When I needed to create a thousand dummy files in a specific format—something that might have taken me the entire first 24 hours in the past—I completed it within a few hours using AI assistance. I also did a lot of the scraping work because that’s stuff I know how to do from research. I have a sense of what we needed, and because I know what classified government documents look like, I knew exactly what to ask AI to help format the dummy documents.

For the presentation and UI/UX work, I was much more hands-on. I care about design and worked in Photoshop to create the final versions. I hate using emojis—it’s important to me to have sleek icons—so I did prototypes of the designs, but ended up doing the final version with Photoshop and editing extensively. AI helped, especially when I had 48 hours to come up with this, but it involved a lot of my sensibilities and experience as well.

If you want to see, here is the final presentation website with a dummy (not backend connected, pure prototype) interface. Would love feedback from people who actually know about these things so I can continue to learn!

Screenshot of our DECLASSIFY hackathon project website.

IV. What Vibe Coding Taught Me

The hackathon experience got me excited about learning to code in a way I hadn’t expected. After it ended, I decided to challenge myself by building iOS apps just to learn the process. That’s when the real lessons about vibe coding became clear.

The difference was dramatic. For web development during the hackathon, I could give the AI detailed, targeted guidance based on my existing understanding of web architecture. I knew exactly where I needed to intervene versus where AI could just do things for me, because sometimes giving things to AI is worse than doing them myself.

But with iOS development, I had no background knowledge. I really just had this very big picture idea. I couldn’t tell the AI exactly where to complement my skills because I didn’t have those skills yet. The AI would make mistakes that seemed uniquely AI-like—forgetting basic functionality like adding a button for a camera to shoot, or organizing files in completely wrong folder structures. These were weird structural oversights that anyone thinking about camera functionality would catch.

One thing I’ve learned is what things are pretty easy to do with Claude versus less easy, and it’s not necessarily what you’d think. Some things seem complex that are extremely easy to code, and things that seem simple that aren’t easy to code. I don’t have that intuition yet, but I’m building it through trial and error of vibe coding while being pretty active about it—going through the code, trying to understand what was given to me.

If I were completely passive about these things, I would just get something unusable. When people say “vibe coding is bad, these people aren’t thinking, they don’t know how to code,” I think there are real concerns—in terms of cybersecurity, maintenance, and development. There’s a risk of lower quality and less secure apps that may not survive software updates. The coding might just be weird. I understand those concerns.

But it’s impossible to vibe code a whole thing being completely passive, unless it’s something extremely simple, because there will be bugs. Even Opus, which I mainly use for debugging, is not perfect. I think the people who would be the best vibe coders would be people who are already pretty good at coding, because they’ll know exactly how to prompt, what to prompt, how to organize things, and where to start.

I’ve tried Lovable and Cursor too, and it’s the same for all of them. The first drafts are always really shitty. I compare that to what I asked AI to do for the hackathon presentation and UI/UX, which was much better, because I had a very detailed understanding of what needed to be there.

One practical challenge I think about in terms of the long-term consequences of vibe coding is dependency. My Claude Opus credits are limited on the pro plan—only four queries every five hours—so I’d wake up in the middle of the night when credits refreshed to debug issues. This dependency felt different from my other AI uses, where if the AI doesn’t work one day, I can do research just fine on my own.

I’m learning more through this process, even if I don’t know if I’ll ever have the incentives, structures, or need to not be dependent on AI for coding. That’s partly because this isn’t my career—I’m not starting over to become a software engineer. I’m doing this basically to learn and because I want to, and I’ve found it useful for creating little scripts that help me organize files and analyze things.

This distinction between what works for an established professional versus someone starting their career is important. In addition to being established career-wise and CV-wise, having the experience to know what is quality work and what isn’t is necessary to produce quality work with AI and to have AI enhance your work. These lessons I’m learning for myself don’t necessarily apply to other people or most of society.

One of the perks of being part of a fellowship with people from all over the world was that they were excited to see tourism in the city I live in. I ended up going to the mall at dark for the first time, and it was stunning.

V. The Missing Conversation: Silos and Solutions

What concerned me most at the expo was a siloing problem exemplified by one notable example: a booth by the International Committee of the Red Cross, positioned right in the middle of the expo floor, that specifically addressed the use of AI in warfare and asked people to think critically about it. In particular, they had some reports available that I'm linking in the footnote at the end of this paragraph to encourage more people to enter these conversations.¹

According to my colleagues who visited that booth, the only people stopping by were those already working in social sciences, humanities, or public service—people who were already thinking critically about these issues. The people making the tools and the policies, who have the most direct impact on these consequences, weren’t engaging with these critical perspectives.

This speaks to a broader problem: we need conversations between different communities, but they’re not happening enough. Coming from a public service background, I've learned a lot from understanding where tech people come from and how the technology works. But I don’t see that happening enough on the other side. The ICRC booth was an example of this gap—their criticisms are valid and necessary, but the engagement isn’t reaching the people who are calling the shots in tech development.

I think both communities would benefit from more engagement with each other. But the difference is that tech people are the ones building and deploying these systems, which means the responsibility for bridging these conversations falls more heavily on them.

What makes me so grateful to be part of the International Strategy Forum Fellowship is that it represents exactly this kind of big-tent approach—80+ people from all over the world with very different views and relationships to technology. Engineers, human rights reporters, philosophers, entrepreneurs, government officials, national security and foreign policy experts, business and development professionals, venture capitalists—it’s an incredibly diverse group with diverse contributions. I’ve learned something from every single fellow, and I think we need more of these forums to break down silos.

While I think there’s room to grow in making these conversations more attractive and seamless than just having the Red Cross booth standing alone, there are other ways to make cross-disciplinary engagement more effective. We need both individual AI literacy and institutional frameworks that support thoughtful integration rather than wholesale replacement.

The future being built in convention centers and hackathons will be shaped not just by technological capabilities but by how wisely we choose to integrate them into human systems. That requires bringing together technical developers, policy experts, ethicists, and affected communities in meaningful dialogue.

As I left the expo, I felt neither the euphoria of AI evangelists nor the despair of AI doomers, but something more nuanced: recognition that this technology has genuine potential if we can develop the literacy, guardrails, and cross-disciplinary collaboration necessary to use it thoughtfully.

In short, I came out of this expo learning a lot and with a lot of ideas that I look forward to developing and sharing with everyone in the next few weeks. Stay tuned!

1. Reports available included “War, law and outer space: pathways to reduce the human cost of military space operations” and “Eight Rules for Civilian Hackers During War and Four Obligations for States to Restrain Them.”