The Data Recipe for Teaching AI New Skills (OLMo-3, Nemotron Ultra, GLM4.5/5, MAI-Thinking-1 review)

~A fail: I recorded everything with my proper mic turned off; sound will be better next time!~ How much machinery is required to get an AI good at some particular thing? Since GPT-1, we have known that large-scale data drives AI progress. What is less clear is how much targeted data is still needed today. If we want a model to become good at task X, will it work out of the box? Can we just apply RL? Do we need SFT traces? Or do we still need humans on platforms like Mercor to create expert data? While closed labs keep their data a secret, some amazing open-source training papers such as OLMo-3, Nemotron, GLM, and MAI-Thinking-1 show more insight. My conclusions: - Training remains highly task-shaped. Models improve when we can turn desired capabilities into measurable tasks, then generate high-quality data and feedback signals around those tasks. - We do have a working recipe for turning a measurable capability into usable AI behavior. The TMax paper (https://arxiv.org/pdf/2606.23321) provides a simple recipe for terminal agents. Starting with seed tasks from a skill taxonomy, they generate 15k of diverse tasks; after 2-3 days of training on several nodes they get a 9B model to get 27% on the 90 tasks of TerminalBench-2.0. - RL is not magic self-improvement: it still relies on lots of SFT data to prepare the model, and requires crafting high-quality, solvable, and diverse tasks closely related to the task of interest. - For self-improvement to take off, the hope is that training on specific tasks teaches reusable tools: reasoning patterns, search strategies, reflection, decomposition, verification, and abstraction. Right now, from the open-source papers evidence for this seems limited. - There is still the need for a teacher: Most papers rely on an SFT stage for tasks on which we will later RL or seed tasks on specific behavior. This teacher can be another LLM, but often times a human is still in the loop (just look at Mercor). 00:00 Intro 1:25 Top labs don't share data 2:24 Unclear where improvements come from 03:35 Pre-training is evaluation-shaped 07:24 Mid-training 08:40 Long-context 09:28 Teaching specific skills in mid-training 14:05 Once a skill is measurable, it's trainable 16:39 Crafting tasks for RL stage 19:34 A concrete recipe 22:35 Generalising beyond task of choice 24:26 Final lessons