Next Biz Thing #387 rl-list.com

RL List https://www.rl-list.com/

This episode of The Next Biz Thing looks at RL List, a directory and ranking platform mapping the fast moving market for reinforcement learning environments. Host Markus J. Diplama walks through how the site covers thirty eight vendors, why its filters for funding, team size, SOC 2 status and focus area match what buyers actually ask, and what makes its published methodology unusual. Confidence tags on every data point, honest blanks instead of guesses, and per vendor update dates turn a plain directory into a usable procurement tool. Worth a look for anyone working near AI training infrastructure.

Here is a question I did not expect to find myself asking this year. If artificial intelligence agents learn by practising, then who is building the practice rooms? Not the models. Not the chips. The rooms. The simulated worlds where an agent tries something, gets it wrong, tries again, and slowly becomes competent. Somebody has to build those, and it turns out that quite a lot of somebodies now do.

Welcome back to The Next Biz Thing. I am Markus J. Diplama, and this show exists to shine a light on the businesses building the useful, unglamorous, load bearing pieces of whatever is coming next. Today's subject is a good example, because it does not build the technology itself. It builds the map. The site is called RL List, and it is a directory and ranking platform for the reinforcement learning environment market.

Let me back up and explain the world this lives in, because the value of what RL List does only makes sense once you understand the problem.

Reinforcement learning is a way of training a system through trial, feedback, and repetition rather than through examples alone. If you want a model to become good at writing code, or at operating a browser, or at completing a long multi step workflow inside a business, you need somewhere for it to try. You need a task, an environment, a way to check whether the attempt succeeded, and a signal that tells the model whether it did well. That whole package is what people in the field call an RL environment.

Over the past couple of years, building those environments has quietly turned into an industry of its own. And it is an industry with a genuinely awkward shape, because the environments are hard to build, expensive to build well, and almost impossible to evaluate from the outside. If you are running a frontier AI lab, or a large enterprise trying to train agents on your own workflows, you now face a procurement question that did not exist three years ago. Who do you buy environments from?

That is the question RL List answers.

The site is a curated directory covering thirty eight companies that build RL environments. You can browse them, you can filter them, and you can compare them. The filters tell you a lot about what the buyers actually care about. You can filter by funding raised. By team size. By customer base. By whether the company holds SOC 2 certification. By focus area, meaning what kind of environments they specialise in.

Notice how practical that list is. That is not a set of filters designed to look impressive on a homepage. Funding tells you whether a vendor will still exist in eighteen months. Team size tells you whether they can support you. SOC 2 tells you whether your security team will approve the contract. Focus area tells you whether they have actually built the kind of environment you need. Those are the four questions a real buyer asks, and the site is organised around them.

On top of the directory sit rankings, use case guides, and a documented methodology. The use case guides are split into the three areas where demand is most concentrated right now. Coding agents. Computer use and browser agents. Enterprise workflows. Anyone paying attention to where agents are actually being deployed will recognise those as exactly the right three buckets.

RL List describes what it offers as a cited, first pass procurement shortlist for 2026, built from public and vendor shared data. I want to pull that sentence apart because every word in it is doing work.

Cited. Every claim points somewhere.

First pass. This is not pretending to make the decision for you. It gets you from thirty eight options down to a handful worth a conversation.

Procurement shortlist. It knows who its reader is. The reader is someone who has to go and actually buy this.

Built from public and vendor shared data. It is telling you the provenance up front rather than presenting a black box.

That is an unusually honest positioning statement, and it leads directly to the thing I find most impressive about this business, which is the methodology.

RL List publishes how it does the work. The data comes exclusively from public sources. No circumventing paywalls, no scraping behind logins. Evidence comes from company websites, financial filings, audit registries, Crunchbase, and reputable press coverage. That is a real editorial standard, written down where anyone can hold them to it.

The score itself combines three things. Scale and traction, which looks at funding raised on a logarithmic scale and at customer acquisition. Signals, which covers research team credentials, verified customers, connections to frontier labs, and security certifications like SOC 2 and ISO. And data confidence, which is an assessment of how well verified the underlying picture is, with a bonus where claims are backed by primary sources.

That third component is the one I want to dwell on, because it is rare and it is smart. Most rankings treat all data as equally solid and quietly hide the guesswork. RL List instead makes verification strength part of the score itself. A company with strong but poorly evidenced claims does not get to ride on assertion alone.

They extend the same idea down to individual data points. Every single one carries a confidence tag. Confirmed, meaning it comes from a primary or official source. Reported, meaning a credible third party said it. Estimated, meaning it is a labelled inference, used sparingly. And Unknown, which is left honestly blank rather than filled in with a guess.

I want to give that last one its due, because in my experience it is the hardest discipline in this kind of work. A blank cell looks like a gap in your product. The temptation to fill it with something plausible is enormous. Choosing to leave it empty, and to label it as empty, is a decision that costs you in the short term and buys you credibility in the long term. That is a business making a bet on being trusted rather than on looking complete.

There is a scoping decision worth mentioning too. The rankings focus on dedicated, pure play RL environment vendors, and deliberately exclude data labelling incumbents, infrastructure providers, and open source projects, on the grounds that they are not comparable groups. That is the right call analytically. Comparing a specialist environment builder to a general purpose data labelling giant produces a number that means nothing. But the site does not simply discard those categories. It still covers commercial, open source, incumbent, and infrastructure focused vendors in the directory, it just does not pretend they belong in the same ranked column.

Then there is the update cadence, which in this market may be the single most valuable feature of all. The site notes something that anyone watching this space will recognise immediately. The RL environment market moves fast. Companies raise, pivot, and get acquired within weeks. RL List handles that with rolling verification, with each vendor page carrying its own last updated date, and with the complete snapshot refreshed as a whole. The most recent full refresh w...