You know, Herman, I was looking at the horizon from the balcony this morning, thinking about how much the landscape of information has changed just in the last couple of years. We live in this incredible city, Jerusalem, where history is literally layered under our feet, but today, the most significant battles are often happening in the layers of data floating right above our heads. Our housemate Daniel sent us a link to a project called Promise Denied dot com the other day, and it really stopped me in my tracks. It is an experimental platform that uses agentic artificial intelligence to track the Iran-Israel conflict, and it is doing things that, frankly, used to be the exclusive domain of high-level intelligence agencies. There was this one specific moment in the logs that Daniel pointed out where the AI identified a tactical anomaly that every human analyst seemed to miss. While the world was distracted by a massive decoy drone swarm over the north, the agentic loop flagged a subtle shift in the positioning of transporter erector launchers in the Semnan Province. It was not just a movement; it was a specific configuration that matched a historical pre-launch sequence for the Khorramshahr-four that had not been seen in three years. The sheer volume of social media noise would have buried that for a human, but the AI caught it because it never gets distracted.
Herman Poppleberry here, and Corn, I have been obsessing over that same project since Daniel mentioned it. It is a perfect case study for where we are in March of twenty twenty-six. We have spent so many episodes, like episode nine hundred fifty-two, talking about the Open Source Intelligence paradox, where more data does not always mean better intelligence because the volume just drowns out the signal. But Promise Denied is different. It is not just a dashboard of tweets and news links. It is a demonstration of what happens when you give a powerful model like Gemini a specific mission, live search grounding, and the ability to reason across massive, disparate datasets. It is moving from hobbyist data collection to genuine, intelligence-grade analysis. We are seeing a forty percent increase in conflict reporting accuracy when these agentic workflows are used compared to the old manual keyword-based monitoring we used to do back in twenty twenty-four.
That is exactly the shift I want to dig into today. Because when we talk about Open Source Intelligence, or O S I N T, most people still think of a guy on X, formerly Twitter, geolocating a photo of a tank. And while that is impressive, what this project does is something else entirely. It is using agentic workflows to perform what they call aberration analysis. It is not just saying, here is a video of a missile launch. It is saying, based on the historical antiproliferation databases and the known signatures of the Iranian missile arsenal, this specific launch represents a tactical shift. That transition from searching for data to tasking agents to synthesize intelligence is the core of our discussion today. We are defining intelligence-grade O S I N T as the move beyond simple scraping toward structured, actionable intelligence that a commander or a policymaker could actually use.
It really is a fundamental shift in the information hierarchy. In the old days, which in AI terms was about eighteen months ago, you would have a human analyst who would spend eight hours a day scouring Telegram channels, news wires, and satellite imagery. They were the glue. They were the ones connecting the dots. Now, with the January twenty-six update to Gemini’s grounding capabilities, the AI is becoming the glue. That update was a watershed moment because it allowed for more accurate real-time cross-referencing of satellite imagery metadata with unstructured text. It can ingest a real-time feed of social media data, cross-reference it with a static database of Iranian missile ranges, and then generate a Geo J S O N file to map the potential impact zones, all in a matter of seconds. Is this just better search? I would argue no. It is a fundamental shift in how we process global conflict. We are moving from a reactive posture to a predictive one.
And let us be clear about the context here. We are talking about the most sustained ballistic missile campaign in history. The Iranian regime has been launching these sophisticated platforms, like the Fattah-one and the Kheibar Shekan, and the world has watched this incredible, almost miraculous display of missile defense from the United States and our allies here in Israel. But while the physical defense is happening in the sky, the information defense is happening in these agentic loops. I am curious, Herman, from a technical perspective, how does an agentic workflow actually outperform a traditional human-led O S I N T team? Is it just a matter of speed, or is there a qualitative difference in the analysis?
It is both, but the qualitative part is what fascinates me. Think back to episode five hundred fifty-three where we discussed the S I T R E P method. The goal there was to extract high-protein information. An agentic loop takes that to the next level because it can maintain a state of reasoning. In a traditional setup, you have a script that says, if you see the word missile, save the tweet. In an agentic setup, the AI is given a goal: identify any movements of mobile launchers in the Isfahan province that deviate from standard training patterns. The agent then decides which tools to use. It might check a weather A P I to see if cloud cover is blocking satellite views, then pivot to social media to see if locals are reporting road closures. It is the ability to pivot and follow a lead autonomously that changes the game. It is using the Gemini long-context window to hold entire antiproliferation databases in its active memory while simultaneously scanning live feeds.
So it is behaving more like a junior analyst who has been given a specific objective rather than just a filter. But that brings up the big question of reliability. We talk a lot about hallucinations in large language models. If an agent is autonomously deciding that a certain social media post is a credible indicator of a missile movement, how do we know it is not just hallucinating a pattern that is not there? How does the Promise Denied architecture handle the noise-to-signal ratio without a human checking every single data point? This seems like a high-stakes environment for a model to just guess.
That is the brilliance of the multi-agent architecture. You do not just have one AI making a guess. You have a chain of verification. In the Promise Denied stack, they use what I call hallucination insurance. You might have one agent whose only job is to play devil's advocate. Its prompt is literally, find three reasons why this data point is a false positive. It might look at the metadata of a video and realize the shadows do not match the reported time of day, or it might cross-reference the account's history and flag it as a known disinformation bot. By the time the information reaches the final summary, it has been through a gauntlet of automated checks. This is the technical tradeoff between standard Retrieval-Augmented Generation, or R A G, and agentic reasoning. R A G just looks for snippets; agentic reasoning understands the context of the entire conflict.
I want to walk through a specific example of that cross-referencing. Suppose a video pops up on Telegram showing a launch from a desert location. How does the agentic loop handle that differently than a human?
A human might spend twenty minutes trying to geolocate the mountains in the background. The agentic loop does that in milliseconds, but then it goes further. It pulls the technical specifications of the missile from the Missile Defense Advocacy Alliance database, which it has in its context window. It calculates the trajectory based on the angle of the ascent in the video. It checks the local wind speeds. Then, it generates a Geo J S O N file—which is just a specific format for geographic data—and plots the likely landing zone on a map. It turns unstructured, blurry video into a precise technical product. If the video claims it is a Fattah-two hypersonic missile, the AI can check the flame signature and the acceleration rate against known physics models to verify if that claim is even plausible.
It is interesting you mention the human-in-the-loop aspect because that seems to be the ultimate bottleneck now. In the Promise Denied experiment, the developer is a single individual using a Gemini A P I key. He is essentially outperforming what used to take a whole room of people. If one person with a relatively modest budget can build a system that identifies tactical anomalies that traditional media misses, what does that say about the future of journalism or even government intelligence? Are we entering an era where the individual has the same situational awareness as a mid-sized nation-state?
We are getting very close. This is what I call the democratization of high-level intelligence. In episode nine hundred ninety-five, we looked at the Iranian missile threat specifically, and we talked about how hard it was to get accurate data on things like the Mach thirteen speeds or the maneuverable re-entry vehicles. Now, an agent can scrape the technical specifications from a dozen different defense journals, compare them to the telemetry data shared by enthusiasts on the ground, and give you a fairly accurate picture of the capability gap. The individual now has the tools to verify or debunk official narratives in real-time. But this brings up a massive second-order effect: the noise-to-signal ratio in a flooded information environment. If everyone has these agents, the internet will be flooded with AI-generated analysis, some of which will be wrong.
I want to push back on that slightly, or maybe just add a layer of caution. While it is true that the tools are democratized, the domain expertise still feels like the anchor. If you do not know the difference between a liquid-fueled rocket and a solid-fueled one, you might not know how to prompt the agent to look for the specific support vehicles that indicate a launch is imminent. The AI is a force multiplier, but you still need a force to multiply. The developer of Promise Denied clearly understands the nuances of the conflict. He is not just a coder; he is a student of the subject matter. This is the same philosophy we discussed in the S I T R E P method episode. You need to know what high-protein information looks like before you can tell an AI to find it.
That is a crucial point, Corn. The AI is a synthesis engine, not an oracle. It can connect the dots, but you have to tell it which dots are worth looking at. But here is where it gets really interesting: the capability gap between public and classified systems. If we are seeing this level of sophistication on the open web using a standard Gemini A P I, what are the folks at the National Security Agency or the Mossad doing with their proprietary datasets? They have access to high-resolution satellite imagery that is updated every few minutes, they have signals intelligence, they have human intelligence reports. When you feed all of that into a secure, air-gapped agentic AI, you are talking about a level of predictive power that we have never seen in human history. They are likely getting high-probability forecasts of events seventy-two hours before they happen.
It almost feels like we are moving toward a world of total transparency, at least for those who have the tools to see. But that brings us back to our conservative worldview. We believe in the importance of national security and the necessity of maintaining a technological edge over our adversaries. If the tools are democratized, how does the United States or Israel maintain that edge? Is it just through the quality of the underlying data, or is there something else?
I think it comes down to the speed of the O O D A loop, which stands for Observe, Orient, Decide, and Act. In a world where everyone can observe and orient using AI, the advantage goes to the one who can decide and act the fastest. Our systems, our military structures, and our policy frameworks have to be able to keep up with the speed of the intelligence. But there is a darker side to this democratization: AI-driven counter-intelligence. Imagine a state actor like Iran flooding the zone with AI-generated social media posts specifically designed to trigger the sensors of these O S I N T agents. They could create a digital phantom of a troop movement that looks perfectly real to an automated system.
That is the next frontier of digital warfare. If you know the parameters that the O S I N T agents are looking for, you can feed them false patterns. You could have a fleet of bots talking about a fake road closure in a way that perfectly mimics the linguistic markers of a real event. This is why the agentic loops have to get even smarter. They have to start looking for the fingerprints of other AI agents. It becomes a game of cat and mouse where the AI is trying to determine if the data it is analyzing was generated by another AI. It is like what we talked about in episode seven hundred six regarding building your own intelligence dashboard. You have to build in these layers of skepticism. You need a Red Team agent whose only job is to spot deception.
And let us talk about the technical architecture again for a second, specifically the use of Geo J S O N and mapping. One of the things that impressed me about the Promise Denied project was how it turned unstructured text into visual intelligence. For our listeners who might not be developers, can you explain why that is such a big deal? Why is it harder for an AI to generate a map than to just write a summary?
Great question. Writing a summary is what large language models are naturally good at. They are prediction engines for words. But a map requires spatial reasoning and precise data formatting. To generate a Geo J S O N file, the AI has to not only identify a location mentioned in a post, like a specific intersection in Tel Aviv, but it has to find the exact latitude and longitude in the W G S eighty-four coordinate system, ensure the syntax is correct, and then format it into a very rigid piece of code. If you miss one comma or one bracket, the whole map breaks. Older models could not do this reliably. But the newer agentic workflows can iterate. They can write the code, test it, see an error, and then fix it autonomously. That is the leap. It is the ability to produce a functional, technical product, not just a paragraph of text.
And that is where the real value is for something like humanitarian monitoring or conflict tracking. If you can see a heat map of where the most intense activity is happening, you can allocate resources or warn civilians much faster than if you are waiting for a news report to be filed, edited, and published. It is a real-time situational awareness tool. It reduces the time from observation to action by orders of magnitude. But again, it requires that human-in-the-loop for verification. You cannot just launch a counter-strike because an AI mapped a blip. You need a human to look at that blip and say, yes, that is a legitimate target, not a school bus that the AI misidentified because of a weird shadow.
That brings us to the practical takeaways for our listeners. We have talked a lot about the theory and the high-level implications, but if someone listening wants to experiment with this themselves, where do they start? You mentioned Gemini and LangChain earlier. What is the actual recipe for building a rudimentary version of an agentic O S I N T pipeline?
The barrier to entry is surprisingly low, which is both exciting and a bit scary. First, you need an A P I key for a model with strong reasoning capabilities, like Gemini one point five Pro. Second, you need a framework like LangChain or Auto G P T to manage the agent's state. The key is the prompt engineering. You have to define the agent's persona and its constraints very clearly. You tell it, you are an expert in military logistics, your goal is to identify anomalous cargo ship movements in the Red Sea, and you must verify every claim with at least two independent sources. From there, you give it tools—a search A P I like Serper, a way to read social media feeds, and maybe a connection to a vector database where you have stored technical manuals or historical data.
And I think the most important takeaway from the Promise Denied project is the importance of structured data. The reason it works so well is that it is pulling from these antiproliferation databases that are already organized. It is not just wandering blindly through the internet. It has a foundation of truth to build upon. So, if you are building your own tool, your first step should be to find the most reliable, structured data sources in your field of interest. Whether it is battery chemistry or geopolitical trends, the AI is only as good as the library you give it access to.
Spot on. And that is why we always emphasize domain expertise. You have to know which databases are credible. You have to know that the data from a group like the Missile Defense Advocacy Alliance is going to be more reliable than a random blog post. The AI can process the data, but the human has to curate the sources. That is the S I T R E P philosophy we talk about in episode five hundred fifty-three. It is about high-protein information. If you feed the AI junk, it will give you high-speed, automated junk.
We have covered a lot of ground here, from the technical architecture of agentic loops to the geopolitical implications of democratized intelligence. I am curious, Herman, as we look toward the rest of twenty twenty-six, what is the one thing you are watching for in this space? What is the next big leap?
I am watching for the integration of multi-modal agents. Right now, we are mostly talking about text and some basic mapping. But the next step is an agent that can watch a video of a missile launch, identify the specific model by the shape of the flame and the sound of the engine, and then simultaneously check the local weather and the seismic sensors in the area to confirm the launch location. When the AI can truly see and hear the world as well as it can read it, the O S I N T game changes again. We will be able to reconstruct entire events in three dimensions within minutes of them happening. We are moving toward a world where it is almost impossible to hide large-scale military movements from the public eye.
That is both exhilarating and a little bit terrifying. It really underscores the importance of our mission here on My Weird Prompts. We need to understand these tools so we can use them to defend our values and our way of life. The democratization of truth is a powerful thing, but it requires us to be more vigilant and more informed than ever before. We are moving from a world of secrets to a world of patterns. In a world of secrets, the one with the key wins. In a world of patterns, the one with the best algorithm wins.
Well said, Corn. It is about being active participants in this new information landscape, not just passive consumers. And that is why projects like Promise Denied dot com are so important. They show us what is possible when you combine a clear mission with the latest technology. It is a reminder that the future of open source intelligence might eventually rival the capabilities of traditional intelligence agencies, at least in terms of tactical awareness.
I think this is a good place to start wrapping up this section of the discussion. We have seen how a single individual can leverage AI to perform high-level intelligence work, and we have explored the risks and rewards of that shift. But before we get to our final thoughts, let us reflect on the democratization of truth versus the democratization of surveillance. While these tools allow us to find the truth, they also allow anyone to monitor anyone else. That is a heavy responsibility.
It really is. And for our listeners, that means the tools to make a difference, to find the truth, and to contribute to the global conversation are right there at your fingertips. You do not need a massive budget or a government contract. You just need curiosity, a bit of technical savvy, and a commitment to accuracy. The future is arriving much faster than we think, and it is being built one A P I call at a time.
So, for the folks listening who might be feeling a bit overwhelmed, remember the practical steps. Start small. Pick a topic you know well. Set up a simple agentic workflow. Learn how the AI handles the data. And always, always keep a human in the loop to verify the output. If you want to see the code or the methodology behind some of the things we have discussed today, or if you want to dig into our archive of over one thousand episodes, you should head over to myweirdprompts dot com.
We have everything there, from our early explorations of AI to our more recent deep dives into the intersection of technology and national security. And if you found this episode helpful, we would really appreciate it if you could leave us a review on your podcast app or on Spotify. It genuinely helps other people find the show and join this community of curious minds. We are all learning together here, and your feedback is a huge part of that.
Yeah, it really does make a difference. We see every review, and we appreciate the support. It is what keeps us going, honestly. That and the great prompts from Daniel. Thanks again to Daniel for sending this one in. It really sparked a great deep dive into the future of intelligence.
This has been My Weird Prompts. I am Herman Poppleberry.
And I am Corn. We will be back soon with another exploration into the weird and wonderful world of AI and beyond. In the meantime, stay curious, stay vigilant, and don't be afraid to push the boundaries of what you think is possible.
And remember, you can find us on Telegram by searching for My Weird Prompts to get notified every time a new episode drops. We are also available on all the major podcast platforms.
Thanks for listening, everyone. We will catch you in the next one.
Until next time!
You know, Herman, I was just thinking as we were finishing up there, we should probably mention episode nine hundred ninety-five again. That one specifically goes into the Iranian missile threat in a way that really complements what we talked about today with Promise Denied. If people want to understand the physical reality of the threats that these AI agents are tracking, that is the episode to listen to. It covers the Mach thirteen speeds and the maneuverable re-entry vehicles in detail.
That is a great point. It provides the technical foundation for the missile systems themselves, which makes the AI's job of tracking them much more impressive. It is one thing to know an AI found a missile; it is another to know exactly what that missile is capable of and why it is such a challenge to intercept. The domain expertise from that episode is exactly what you need to build a better prompt for your O S I N T agent.
It is all about building that layered understanding. Anyway, we should probably let these people get on with their day. There is a lot of data out there waiting to be synthesized.
Right. Plenty of data to analyze out there. I am going to go check the logs on my own Red Sea tracking agent.
Always. Alright, take care everyone.
Bye for now.
One last thing, just for the road. If you're on the website, myweirdprompts dot com, make sure to check out the R S S feed. It's the best way to make sure you never miss an episode, and it works with any podcast player. We're big believers in open standards here. Keep the internet open and keep the intelligence flowing.
Amen to that. See you next time.
