WHY DO PROPRIETARY RL ENVIRONMENTS MATTER? My boy Prakash seems unconvinced. TIME TO MAKE MY CASE First, what's one of these fancy 'RL ENVIRONMENTS.' What does that mean? Lots of slop around these atm so lemme help. An RL environment is like a sandbox where an agent can practice f***ing up a task AND score itself on how it did. Think of like a flight simulator for an AI agent. The env hands the model a job, like 'complete this tax return'. The agent takes action via tools, and at the end up grader checks the result and hands back a reward. THIS DERIVES YOUR EVAL METRICS (word evals being thrown around alot these days too). Run that bad boy like a million times and you theoretically should be able to improve your agent. PROPRIETARY ONES ARE ALPHA. Only YOU know how well this agent is doing at this task, and if it is defined / implemented by subject matter experts man oh man you have something special in the AI AGE! It is hard to find SMEs + RL Env experts who combined can build the sandbox necessary to actually test the frontier. Prop rl envs matter to a variety of market participants. Let's start with the one I was originally referencing: @llmh. Long Lake is an AI PE fund/holdco that buys companies (congrats on amex @alextaubman) like property management cos and improves them via ai agents. AI first orgs (AI ROLL UPS) NEED prop RL environments for a variety of reasons: 1) DON'T DESTROY YOUR COMPANY WHILE IMPLEMENTING AI. How do you test your AI agents without destroying your business OTHER than an RL environment? You need a playground / sandbox for your agents to make mistakes in so you can MAKE THEM GOOD AT STUFF. 2) How do you know if your AI agents are ACTUALLY WORKING before you deploy them? Once you have AI agents, you need benchmarks / environments in which to say 'we are ready to go. we think we will see these gains'. 3) How do you convince the frontier labs to improve the models in ways that actually derive utility for you WITHOUT them. You can take your RL env (which always include prop bench marks) to openai and say 'hey sam, wtf, my ai accountant is doing this dumb shit and its on your model to fix it' NOW, that was an AI first org, how about the data companies themselves? @joinHandshake and @GarrettLord take a fascinating approach. 1) PROP RL environments can help you understand the GAPS in the frontier so that you can BUY that data and FILL those gaps. 2) They help you advertise: 'Hey OpenAI, we invented this new RL environment with this benchmark and you guys AREN'T GREAT at that' 3) They help you compete: If you have rl environments your comp doesn't have (@mercor, @micro1_ai, etc) you can find data to win in those environments that your competitors can't. ANYWAY that's why RL environments matter a LOT and will continue to matter a lot. You will see a race in these AI roll ups / AI services companies to hire the best RL env / post training people soon. This is also why you see so many YC companies building these RL environment companies.