Grounded Rewards: There is something brewing here, it's starting to move really fast, and few understand it rn. Candidly, myself included: I barely grok it, but see enough crumbs to have an appoximate sense of its (very large) upcoming impact. And to be clear, grounded post training is not an industry term I've come across, it just captures the essence of my recent observations. At the heart of it: Grounded rewards from real world characterization data enables LLMs to reason out of distribution. Anthropic, OpenAI et al are unlikely to resource this work to a necessary degree, because the billions they generate in 'classical' inference represents their genuine innovator's dilemma. Else, they may simply be unable to make OOD reasoning converge because they are married, confined, and increasingly incarcerated by reward methodologies that must be instantiated, executed, evaluated, and backpropped in a digital domain - in software. Of course, if it proves true that RL post-training on math and code alone is sufficient to generalize up to AGI, then real world characterization data doesn't matter and we're all cooked anyway. But the more I learn about these concepts, the more I am starting to build a core conviction that AGI, by definition, must reason out of domain, and that we therefore need novel experimental design and simulation infrastructure to support the extension of already post-trained LLMs into real-world characterization land. We need to give them the tools to experiment, verify, and iterate on out of domain tasks and characterizations. I suspect the first wave here will be to literally post-train an LLM to use real world experimental tools (exposed via API function call) to produce real world / wet lab characterization data, in pursuit of (dis)proving an LLM generated hypothesis. Then, reward the model on the process it used to generate a good (i.e. expected) characterization, and rinse & repeat. In doing so, you've established action, state, trajectory, agency, reward, and value - all the ingredients necessary to get a nastythicc RL loop converging. Grounded rewards suck today because they take hours, days, weeks, months, or even years to conclude, and generate signal from which to reward. They are extremely long duration, yet often information dense. Is there a sweet spot? I think so.