Not metaphorically. I mean literally watching a neural network try to figure out if a three-bedroom in Austin was overpriced. It was bad. Embarrassing. The model kept recommending properties based on how nice the photos looked. That was 2023. Then Anthropic did something I didn't expect. They started treating real estate like a training problem, not a prediction problem. Here's how their strategy works. And why it might actually break the market. Step 1: Stop predicting prices. Start modeling irrationality. Every real estate algorithm I've seen tries to forecast "fair market value." This is stupid. Markets aren't rational. They're emotional, seasonal, and full of humans making bad decisions because they fell in love with a kitchen island. Anthropic's approach flips this. They don't train on historical sales data alone. They train on why people overpaid. Their models ingest listing descriptions, agent commentary, neighborhood sentiment from social media, even the tone of open house conversations. Practical tip: If you're building your own system, collect the noise. The screaming match between buyer and agent? That's signal. The Zestimate? That's noise. Common pitfall: Assuming more data helps. It doesn't. Garbage in, garbage out. The internet is terrible. You need to filter aggressively. Step 2: Build the "Iron Man suit" for deal discovery. Here's where Anthropic's strategy gets interesting. They don't automate buying. They augment it. Their system works like this: The AI scans thousands of off-market listings, foreclosure notices, probate filings, and private sales. It flags properties where the seller's situation creates leverage — divorce, estate liquidation, relocation, motivated downsizing. Then a human agent steps in. Not to negotiate. To witness. To verify the AI's recommendation with real-world inspection. This is the key insight Anthropic stole from Tesla's Autopilot playbook: The system handles 95% of the work. The human handles the 5% that would break everything. Practical tip: Set your AI to optimize for "seller pain" not "property quality." A perfect house with a desperate seller beats a great house with a patient one every time. Common pitfall: Letting the AI close the deal. Don't. The moment you automate the final handshake, you lose the human judgment that catches the rotting foundation or the neighbor's meth lab. Step 3: Exploit the jagged intelligence of LLMs. This is the part most people miss. LLMs are terrible at some things and superhuman at others. Anthropic figured out exactly where the jagged edge cuts in real estate. Their models can: - Read a thousand pages of zoning regulations in seconds - Cross-reference tax records with flood zone maps - Spot legal loopholes that would take a human lawyer days to find Their models cannot: - Tell if a crack in the foundation is structural or cosmetic - Know if that weird smell is fixable - Judge whether the neighborhood is actually improving or just gentrifying fast enough to flip So they designed the workflow around the AI's strengths and the human's strengths. The AI does the reading. The human does the looking. Practical tip: Map your AI's failure modes before you deploy. Run a thousand test cases. Sort them by loss descending. You will find the weird stuff. Common pitfall: Assuming the AI's weaknesses are temporary. They're not. The jagged edge is a feature, not a bug. Step 4: Build a data flywheel. This is the real moat. Every deal Anthropic's system touches generates new data. The AI recommends a property. The human inspects it. The deal either closes or doesn't. That outcome feeds back into the model. Over time, the system learns which signals actually predict successful acquisitions. Not just which properties look good on paper, but which ones survive the inspection, the appraisal, the negotiation. This is why traditional brokerages are screwed. They have transaction data. Anthropic has decision data. They know why you said no, not just what you bought. Practical tip: Track every rejection. Build a database of "almost bought but didn't." That's where the real learning lives. Common pitfall: Only collecting positive outcomes. Your model will learn to be optimistic and wrong. The bottom line. Anthropic's strategy isn't about replacing agents. It's about making agents 10x more effective by giving them superhuman pattern recognition and letting them focus on the 5% that matters. The market won't be disrupted by AI that buys houses. It'll be disrupted by AI that finds the right houses and lets humans make the final call. That's the Iron Man suit approach. And it works. I'm still not sure if this makes real estate better or worse. But I know it makes it different. And different is scary when you're the one holding the bag.