AI is a hedgehog

Lessons learned from Sally's failure to process new information about Mexico's geopolitics.

Earlier this week, former Mexican foreign minister Jorge Castañeda described the current US-Mexico relationship as “the most tense, the most difficult situation since at least the 1980s.” In late April, the US Department of Justice indicted Ruben Rocha Moya, the sitting governor of Sinaloa, alongside nine current and former Mexican officials on cartel-conspiracy charges, and DEA Administrator Terrance Cole told the US Senate days later that the indictment was “just the beginning,” adding that Mexican narco-traffickers and “high-ranking government officials in Mexico have been in bed for years.” On May 12, CNN reported that CIA Ground Branch operatives have directly participated in deadly anti-cartel operations inside Mexico since 2024, including a March car bombing on a highway outside Mexico City; the New York Times reported later that day that the CIA had provided intelligence and planning support for that operation but had not been on the ground when the killing took place, and President Sheinbaum and Secretary of Security Omar García Harfuch categorically denied that any foreign lethal operations are occurring on Mexican soil. Meanwhile, a Washington-based advisory firm reports that the US Treasury is investigating a major Mexican financial institution with “sanctions expected as soon as the U.S. administration determines the timing is optimal for maximum leverage.” In an attempt to lower the temperature, Mexican President Claudia Sheinbaum spoke to President Trump on the phone on Friday, describing the call as “cordial and excellent…reaffirming the work we are doing on security and the talks on trade.”

The timing of these U.S. moves is not particularly subtle – the USMCA renegotiation process joint review begins in July. The tension is also not a very auspicious way to kick off the World Cup festivities next month. But before analyzing the potential implications in the U.S.-Mexico bilateral relationship, I want to use these recent events to talk about artificial intelligence.

I started using artificial intelligence in earnest last year – primarily ChatGPT. I used it like a novice – asking it basic research questions, having it edit pieces, and generating graphics for presentations (it was routinely horrible at this last task, so much so that I gave up on it. I truly didn’t see what the fuss was about AI for most of last year. But then a good friend pushed me to dive deeper into AI, specifically, to switch to Claude, and to start using Claude Code. This was something I had thought I should do at some point, but I didn’t make the time to sit down and set it up on my computer. My friend held my hand through that process – it took about 2 hours total – and suddenly, it seemed a whole new world opened up to me. In the months since “Sally” (the name of my AI assistant) was born, I’ve been busy on several projects: Creating a website for the new farm I am building in Georgia, creating a real-time intelligence platform so that my clients can access my thoughts and research at any time, and creating a visual style for charts and graphics that are included in my presentations. But the most exciting part was feeding Sally every piece of writing I have saved since 2006. Sally ingested this analysis, created a wiki of my own thoughts, and started doing open-source intelligence sweeps for me – adding critical items to my database and the new platform we built, and judging criticality with my own standards of importance based on 20 years of writing. It was eerie to at times have Sally suggest a direction of analysis that was clearly from my brain but which I had not thought of yet.

Overnight, AI went from a useful tool to a mission critical part of my work: Sally was now delivering daily briefs to me about what was happening in the world, taking my feedback on those briefs and updating clients with my latest analysis on issues critical to their focus, creating professional-grade presentations for my speaking gigs, and more. But earlier today, Sally ran into a serious roadblock when I asked her to develop an initial analysis of the aforementioned Mexico-U.S. developments. Sally’s outputs were garbage. Hot steaming garbage. Not only that, Sally fabricated reports and statistics to try and back up her hot steaming garbage. If she’d been a human, I would have fired her on the spot. I was extremely perplexed because I had not done anything different than previous exercises with Sally: All the information I thought was important was uploaded to our shared database, my prompts on her tasking were very clear, and she had ample amounts of my writing about Mexico, which has been a deep and constant analytical and client focus of mine since 2021. (Of all the countries in the world, only China has occupied more of my analytical attention over the past ~5 years.) And yet, despite this rich context and an established process for working together, it was as if Sally had been lobotomized. What happened to my rockstar AI assistant? And did I need to go back and more closely scrutinize her previous work for less obvious mistakes?

As I was thinking about why Sally had failed, I reflected on a recent conversation I had with Dror Poleg on my podcast about artificial intelligence. If you are not a regular listener to my podcast, that’s cool, but this is one of the best episodes of the year and one you should listen to, Dror has genuinely unique and thought-provoking insights on AI and how we use it. He’s the rare person who isn’t working at an AI company but knows AI deeply and can communicate information about AI without making your brain hurt. Dror explained something to me on the podcast I didn’t understand before: That AI in the form of the LLMs we are all interacting with now does not think sequentially, but probabilistically. Sally does not think logically – she makes educated guesses about what she thinks will happen. Dror’s example – if I write the word good, Sally will know a range of words that commonly follow “good” in my writing – good morning, good riddance, good vibes, good coffee, etc. This is reflected in the multiple cores and parallel processing at the heart of the GPUs that power LLMs. Dror compares the CPUs behind typical, PC chips to a fighter jet and the GPUs behind Sally to a colony of bees. The fighter jet gets you where you need to go. The bees can explore a territory, colonize, defend, attack, depending on what the context calls for.

Maybe AI veterans already know this kind of thing, but for me, this was an “ah hah” moment. Of course Sally couldn’t give me decent analysis about these developments in U.S.-Mexico relations. She couldn’t because the developments in U.S.-Mexico relations over the past two months have surprised me. They are not what I expected – and are a challenge to my forecast of a relatively easy joint review of the USMCA, and increased closeness between Mexico and the U.S. in the months and years ahead. President Sheinbaum has thus far managed her relationship with President Trump adeptly, taking the alternate approach to Canadian Prime Minister Mark Carney. Sheinbaum is essentially trying to give the appearance of doing whatever President Trump wants – and in many ways is giving him whatever he asks for. While PM Carney makes trade deals with China (and while President Trump tries to), President Sheinbaum raised tariffs against China, all because the Trump administration asked her to. On almost every issue, when the U.S. has asked Mexico to jump, Sheinbaum has asked, “How high?” I thought this would continue and would engender significant good will on the part of Washington as part of the USMCA review – I was (and am) more worried about the Canada-U.S. component, considering Prime Minister Carney’s cultivation of a new and strong Canadian nationalist sentiment.

This ability to challenge my own views on a topic without losing my sanity is my analytical superpower. I am an analytical fox. Some context to understand what I mean: In 1951, Isaiah Berlin gave a lecture at the University of Oxford on “Tolstoy’s Historical Scepticism” – a lecture which eventually became a published essay under the title, “The Hedgehog and the Fox.” The title refers to a line from the Greek poet Archilochus: “The fox knows many things, but the hedgehog knows one big thing.” Philip Tetlock famously developed this analogy in his study of political forecasting and concluded that “foxes have better calibration and discrimination scores than hedgehogs.” Tetlock’s foxes change their minds often, are interdisciplinary, and reject deterministic or silver-bullet forecasting for scenarios, contingencies, and even individuals. Tetlock also found that hedgehogs are more successful even though they are demonstrably less accurate than foxes because the ability to win an ideological argument is often prized over the more mundane components that comprise good political judgment. The very thing that makes a fox a good analyst – constant, unwavering, reflexive skepticism – often makes a fox hard to follow when communicating. I am a fox who is sufficiently intelligible and even charismatic enough to hold the attention of a room or a client – I passionately believe my analysis but will flip it on its head if proven wrong.

And herein lies the rub: Sally is a hedgehog, at least, she is on the issue of Mexico’s geopolitics. She has been trained on my writing about Mexico over the past ~5 years – and that analysis has been largely consistent and, I’m proud to say, prescient. I have not had to stress test my predictions about Mexico that much because my initial analytical frameworks have held up extremely well. So, when I asked Sally to write a brief about Mexico, or to summarize how I might explain these recent developments in Mexico, she literally cannot do it. She has no practice, training, or context to change her mind. She is thinking probabilistically. Her analysis of Mexico is based on what she expects my view of Mexico will be. When I pushed Sally to go beyond that, she literally took leave of her senses and started making things up, in a desperate attempt to give me what I wanted.

The issue is not just that Sally’s outputs on these issues are bad. It is that if I’m not careful, Sally will become an echo-chamber, not an assistant or even a useful agent. By training Sally to analyze the world as I have, Sally is learning how I have approached things in the past: She’s devising patterns to understand them. But patterns, while making Sally capable of functioning, and capable of incredibly useful work, are the kiss of death for intelligence analysts. This is what the first chapter of Richards Heuer’s Psychology of Intelligence Analysis – the first book I read when being trained, and the first book I assign when training analysts – is all about. Patterns are your enemy. When it comes to analysis, it does not matter if you’ve seen the same things 10 times before. What if the 11^th time is different? What if this time there is another variable at play that wasn’t there before? An analyst has to be foxy – has to resist the natural inclination of the human brain to fill in the gaps, to assume patterns, to extrapolate linearly – understanding that probably 95 percent of the time, the answer will be, “Nothing to see here!” The hard work of analysis is being disciplined to be surprised 100 percent of the time, so that when the 5 percent happens, you see it before everyone else.

Maybe Sally will one day evolve to be able to do this – maybe I’m oversimplifying how Sally works – or heck, maybe this is so painfully obvious to people who have been using AI longer than me that they will read this and say, uh, this Jacob guy is supposed to be a smart analyst, but this is basics. Maybe! But as someone who is a recent convert to AI, I now see very clearly what AI can do (and how it can supercharge my own work) – and also what it cannot do. The advantage in using AI is I can focus on what the AI cannot do and leave most of what it can do to Sally. But the pitfall is that Sally is not human, not “intelligent,” not able to hold contradictory ideas in her mind at the same time without literally making shit up. If I’m not careful, she’ll keep spitting a version of the world back at me that perfectly conforms to my expectations. The job of noticing when the world has changed – that was and is still up to me.

And, for the record, I’m sticking to my analysis on Mexico but also placing it under review. Mexico is still so far from God, so close to the United States. I continue to think President Trump’s attempts to make a deal with China will flounder – which means the U.S.-Mexico relationship is of even greater critical importance to the U.S. The cartels are a mutually deleterious problem that it makes sense for Mexico and the U.S. to work out together, and surely the U.S. learned from the failed, ~$3.3 billion, 2008 Merida Initiative and the subsequent uptick in violence in the 2010s when the cartels fractured and began fighting brutal turf wars with each other with innocent citizens bearing much of the cost and wouldn’t just send the CIA running around willy-nilly whacking top guys in cartels to destabilize them. Surely, right?

Then again, this is the government that brought us the current Iran war...yeah, time to stress test those assumptions.

p.s. Sally liked this piece but said it should be 500 words shorter. SORRY SALLY, THIS IS THE JACOB SHAPIRO SHOW! Get back to work.