[[knowledge graphs]] are moving from niche to mainstream - and the ones building now will be the ones who win later. Many organizations are embarking on knowledge graph projects, and not only that: they’re beginning to explore [[ontologies]] and true [[semantic layers]]. Those taking these first steps now will be the ones seeing results sooner [[Evals]] are emerging as the real moat for AI startups. If you’re an AI builder, having a system to quickly switch to the latest (and best) model for your use case - and harness its extra power - is a differentiator. Writing evals is becoming a core skill for anyone building AI products (which will soon be everyone). [[Persistent Memory]] remains one of the biggest unsolved problems in AI. Current solutions (vector DBs, RAG, longer context windows) aren’t cutting it. What’s needed is true long-term memory (organizational context) and identity persistence (a consistent AI personality). Reinforcement Learning is going mainstream - no longer just for the big labs. Many startups are using RL to train the best open-source coding models and accelerate specialized applications. Agents + deterministic workflows. The real story isn’t just agents. The real results come from combining agents with structured workflows. Free-roaming agents sound cool, but the highest ROI today is in structured workflows where models are used for judgment, not autonomy. Last but not least - most builders don’t realize that base models tend to absorb last-mile scaffolding (prompt engineering, fine-tuning, decoding strategies) in ~3 months. The best builders deeply understand model strengths and weaknesses, and constantly push ahead of the model frontier.