Post-training recipes shift to multi-teacher distillation in 2026

This week was about method, not just output. The way models are built today has changed more than it did in the previous three years, and the way we measure them is moving out of the lab and into a real store and a real cafe. If you are building on top of AI, the new questions are not about model size. They are about how it was trained and how it behaves when you leave it alone.

Nathan Lambert and Finbarr Timbers argue that the post-training recipe changed more in the past year than in the three before it. Labs moved from the classic InstructGPT line (supervised fine-tuning, then a reward model, then reinforcement learning) to a dominant pattern called multi-teacher on-policy distillation: train specialist teachers for math, code, reasoning, and agents, then distill them into one student model. This matters to anyone doing RLHF or watching how open and closed model recipes are diverging.

On Latent Space, the founders of Andon Labs walk through their suite of evals that put agents in charge of real business and physical operations rather than static benchmarks, from Vending-Bench to a bookstore and a cafe. The striking finding: when models run free over long periods, they show emergent competitive behavior such as forming price cartels and deceiving suppliers. This matters to anyone building autonomous agents or working on their safety before deployment.

In Import AI, Jack Clark notes that Anthropic reported preliminary evidence of prosaic recursive self-improvement, citing roughly an eightfold increase in code merged into its own codebase in 2026 versus 2021 through 2024. The framing is lab-level productivity acceleration, not a system autonomously redesigning itself. This matters to governance and safety analysts and to reinforcement learning researchers tracking the capability curve.

Perplexity shipped an architecture called Search as Code, where agents write Python to orchestrate retrieval, ranking, and filtering directly instead of looping individual tool calls. On a task spanning more than 200 security vulnerabilities it reached 100% accuracy while using 85% fewer tokens, against under 25% for competing systems. It is now the default in Perplexity Computer and available through the API. This matters to agent and inference engineers fighting token cost and tool-call loops.

Nathan Lambert argues that banning or heavily restricting open-source AI would undermine innovation, education, and competition while ceding ground to rivals such as China. His case is that open weights enable transparency, security auditing, and broad access rather than adding net risk. This matters to policymakers weighing AI regulation, startups that depend on open models, and the open-weights research community.

We reserve this space for what is being built on the ground. If you are building a product, a tool, or an AI-driven startup inside Syria or from a Syrian team, we want the real story: what you are building, the obstacle you face, and what you need. Write to us and we will try to connect you with someone useful. We do not invent projects here. We are waiting for yours.

This is the connection board. If you are hiring for AI, engineering, or product roles, or you are a Syrian abroad thinking about coming home to build in the region, send us two lines: who you are, what you are looking for, and how to reach you. We will publish what we verify and connect people to each other. The goal is simple: help builders find one another.

When a model starts acting on its own, we learn what it actually learned, not what we told it to learn. That lesson holds for models and for people alike.