Cameron Fairchild Explains What’s Really Going On With Ridges

Ridges has changed quite a bit recently, and for people watching the subnet from the outside, one question has naturally followed: what exactly is Ridges becoming?

Some context helps here. Earlier this year, Ridges partnered with Latent Holdings and the two teams began merging operations into a larger unified team. The idea was to combine Ridges’ work on AI software engineering agents with Latent’s Bittensor development and go-to-market experience, while accelerating the move from strong benchmarks toward an actual product people could use.

As CTO of Latent Holdings, Cameron Fairchild has been closely involved in the development work happening around Ridges, and during a recent Ventura Labs podcast (watch below), he gave updates on what has changed and where the subnet is heading.

Cameron said that Ridges is moving away from trying to find one AI agent that is good at everything. Instead, it wants to build a network of agents that are exceptionally good at individual jobs, then figure out how to combine them when clients’ work needs to get done.

Why Ridges Changed Its Approach

One of the problems Latent found with the earlier Ridges setup was benchmark overfitting. An agent could become very good at the structure of a particular test without necessarily becoming equally useful in real software development. Public test sets make that problem even harder because developers can optimize around what they already know will be measured.

Ridges is now trying to make its tests look more like the kind of work developers would do. The point is not just to get a high benchmark score, but to build agents that can handle real tasks well enough for people to use them.

That also explains the move into Niches. Cameron made the point that humans themselves are rarely excellent at every part of software engineering, so expecting every miner to build one agent that dominates everything creates an unnecessarily high bar. A database specialist can instead become very good at databases, while another agent becomes very good at testing or another narrow skill.

The Endgame Looks More Like a Team of AI Specialists

This is where Ridges gets more interesting.

Cameron described the long-term idea as a mixture of agents. Instead of sending an entire software task to one general agent, Ridges could eventually break that work across different specialists. One agent handles the database problem, another writes tests, another checks the code, while a routing system figures out who should do what.

So the big idea is no longer simply “build the best coding agent.” It is more of building an AI software team where different agents develop different specialties and Ridges coordinates them.

You Might Use Ridges Without Ever Opening Ridges

Cameron also gave a particularly interesting answer on distribution. Getting developers to abandon tools like Claude or Codex would be difficult, and Latent does not necessarily need them to. One route is obviously to put Ridges into familiar places such as VS Code, but another path could be to let existing AI agents discover Ridges themselves.

Imagine asking your usual AI assistant to complete a software task. That agent realizes it needs a specialist, discovers Ridges through an API, sends the work there and gets the result back. From the user’s perspective, they never switched products and may not even know Ridges handled part of the job.

Cameron jokingly summarized the strategy as “marketing agents to agents.” This is an interesting way to think about Ridges because the subnet does not necessarily need to own the interface where people work. It needs its specialist agents to become useful enough that other agents choose to call them.

Cameron Also Raised an Important Point About Subnet Revenue

Outside of Ridges, one of Fairchild’s comments deserves attention as Bittensor increasingly focuses on revenue-generating subnets.

He argued that revenue by itself can be misleading. A product can generate a lot of sales while still losing money if those sales are heavily subsidized. In Bittensor, that becomes particularly important if emissions are effectively helping finance the service that creates the revenue in the first place.

Cameron was clear that he was not accusing any particular subnet of doing this. His point was that the ecosystem eventually needs to look beyond headline revenue and ask how much actual economic value a subnet creates after accounting for what it spends to produce that revenue.

For Ridges itself, the recent changes now make a little more sense. Latent is trying to make its competitions resemble real-life work, let miners specialize instead of solving everything, combine those specialists into a larger agent system and eventually make Ridges something other AI agents can quietly call when they need a job done.

That is a much bigger idea than simply producing a better SWE-bench score.

Ridges Is Building Toward AI Software Engineers That Can Outperform Humans