Key Takeaways

The Method

Forget the generic AI that just spits out predictions. Simily, co-founded by Jun Song Park, is building something sharper: a foundation model of human behavior. Their method stems from a critical observation: current AI models, trained on the internet’s vast output of human sentiment and action, hold more than just data. They contain latent structures of human behavior that can be extracted if approached correctly.

Park and his team started with experiments like ‘Smallville,’ using GPT-3.5 agents. The magic wasn't the base model itself, but what they added: Memory, Planning, and Reflection.

1. Memory: Agents store past interactions and observations. This isn't just a transcript; it’s a living record informing future decisions.

2. Planning: Based on their memories and stated goals, agents formulate multi-step plans. They don't just react; they strategize.

3. Reflection: This is the kicker. At regular intervals, an agent reviews its own memories and experiences, much like a "shower thought." Park explains, “we had this concept of reflection, which basically was every certain interval, you have you ask agent explicitly to get bunch of their memory pieces and basically make sense of them.” This process allows the agent to learn from its past, refine its internal models, and adapt its behavior without direct human retraining.

This architecture moves beyond simple classification or pattern recognition. It allows Simily to simulate complex, emergent human behaviors – like agents self-organizing a Valentine’s Day party – and understand the causal mechanisms behind them. Their goal: to create a "GPU of intelligence" where you can run countless simulations to understand and shape human action at individual, group, or even market levels.

Where This Breaks Down

Simily's method, while powerful, faces real constraints. First, the models are only as good as the data they learn from. If the underlying web data is biased – and much of it is – the simulated human behaviors will reflect and potentially amplify those biases. This means ethical guardrails are not an afterthought; they are fundamental to how these systems operate and are applied.

Second, human behavior is messy and often irrational. While MPR agents can simulate complex interactions, truly predicting or perfectly replicating chaotic, emotional, or genuinely random human decisions remains a tall order. The fidelity of these simulations in highly unpredictable scenarios will always be a challenge. There is a fine line between a robust simulation and an oversimplified one that misses critical nuances.

What to Do With This

Stop building AI that only reacts or predicts simple outcomes. If you're creating any AI-driven product, especially those interacting with users or making decisions on their behalf, bake in explicit Memory, Planning, and Reflection cycles. Don't just give your bot a knowledge base; give it a memory of its conversations. Don't just give it a script; give it a plan for achieving a specific outcome (e.g., resolving an issue, increasing engagement).

Crucially, implement a ‘reflection’ loop. For instance, if you're building a sales assistant, have it periodically review the last 20 client interactions. Prompt it to identify patterns: What sales tactics consistently led to conversions? What objections frequently derailed deals? Then, have the agent self-correct its approach for future interactions. This shifts your AI from a static tool to a continually learning, adaptive force in your business.