Poolside: An OpenAI Rival? AWS Likes Its GitHub Pedigree

LAS VEGAS — Poolside is a frontier AI lab, similar to Anthropic and OpenAI, but with a major difference. Poolside’s foundation model focuses exclusively on software development. And the company uses that approach to transform the software development life cycle (SDLC) into data models.

Poolside uses its clusters and the data that it generates to create some advantages, such as scaling out by combining compute clusters and synthetic data. Further, Poolside installs its stack inside an enterprise environment, so the data never leaves the customer’s infrastructure.

“If somebody is building an AI application and they’re going to use a general-purpose model hosted by somebody else, a lot of the context, the code context is going to go over the wire,” said Jason Warner, CEO and co-founder of Poolside, at AWS re:Invent.

“So that’s going to go back to somebody else. We’re saying, ‘No, that’s never going to leave your VPC premise or your network boundary — that sits inside your environment. It allows us to advance the model’s capabilities in fundamental,  very specific ways.”

The belief, he said, is that this will enable a faster scale toward the goal of artificial general intelligence (AGI), a type of AI that equals or surpasses human abilities across a range of cognitive tasks.

Overcoming Obstacles to Scaling

Poolside’s competitors, Warner said, face a scaling boundary: the compute cluster size. “We have two. We have compute cluster size, and we can generate more data.”

What does that mean? Poolside can do parallel processing and model training at the same time.

Warner previously worked as CTO at GitHub for four years. He enjoyed a similar role at Heroku, a platform as a service now owned by Salesforce. And before that, he worked in another similar role at Canonical, the people who make Ubuntu Linux.

Eiso Kant, Poolside’s co-founder, has emerged as a leading voice in building out intelligent substrates using neural networks. In particular, he became recognized for his work developing a platform called source{d}.

GitHub explored acquiring source{d}, Warner said, but as sometimes happens, the deal did not close. Kant and Warner, however, became fast friends.

“We kind of bonded over neural networks and their applicability to software,” Warner said.

Warner approaches software development from an SDLC perspective. Kant approaches it based on working with neural networks.

“We’re kind of meeting in the middle of how we apply this for developers because it’s the one common area between us,” Warner said.

Warner said he has obsessed over developing a single flow, from when a developer touches a keyboard to what happens in production. For him, the opportunity comes in making the systems work better, even by using neural networks to find the relevant connections, fitting the code into place and testing for vulnerabilities.

As it is now, he said, developers waste time writing YAML code and configuration environments.

The SDLC is the last legacy waterfall system, Warner said. “And  if you think about that, we’ve got to put a piece of software in; it’s got to go through this gate, it’s got to go through this check, somebody else has to approve it, all that sort of stuff.”

“And there’s really good reasons for that. But it doesn’t mean it needs to stay that way. I think neural networks can help on this. Not only help on this: I think they can advance us past where this is.”

How Poolside Filters Data Sets

Poolside filters to the highest quality data sets that go into their models, Warner said. Also, “we’ve actually removed what we would consider non-business-friendly license types from the pre-training data. A lot of concerns that enterprises have is on the generative nature of non-business-friendly license types, the virality nature of those license types.” Therefore, he said, those licenses are removed from the training data so that the customer can use generative code assist or code completions without repercussions.

In addition, “Poolside installs inside an environment,” Warner said. “So inside the enterprise. We do fine-tune and RAG all that information, but we never see that data coming back to us. That’s the customer’s data. We have no right to ask for that to come back to us.”

Poolside has developed a technique to generate synthetic data and expand its model’s capabilities.

“This is something that Kant has been working on for a long time,” Warner said, “And now we’ve done it at scale and proven out the technique. We use the seed data of the open source world to generate a massive amount of synthetic data.

“We effectively get open source projects to run inside of our massive runtime environment. And we use Poolside to solve tasks, and reinforce, solving tasks inside that environment. And from that, a lot of synthetic data comes out. This is patterned after AlphaGo, if you’re familiar with that, from Google DeepMind in 2016.”

Poolside also signed an agreement to work as a first-party offering in Amazon Web Services, a clear sign of the competition among the largest service providers. Poolside looks like any other native AWS service. S3, EC2, whatever it may be. For example, a customer can buy Poolside by burning down its AWS credits.

“And it also means that we’re on AWS paper. So [master service agreements] and all those sorts of clauses. So, if a customer is already familiar with AWS, they can buy Poolside under the same structure as any other AWS service. A big deal for us. What it also means is that we are anywhere AWS is. You can install it inside any AWS infrastructure.”

AWS and Poolside

Rahul Pathak, AWS vice president, data and AI go-to-market, told The New Stack at re:Invent that he works closely with Warner. For AWS, Pathak said, “The value is simple. We think our customers need access to a broad and diverse partner ecosystem, as well as a broad and diverse set of capabilities that we offer in order to achieve their objectives.

“So, for Poolside, they’ve taken an interesting approach to building models that are specific to software development. And so, from our perspective, having them run on AWS gives our customers more choice, gives them options in Bedrock. And we think that model choice is key to achieving incentives. And so from our perspective, it’s more options for customers when they come to AWS for whatever their use case happens to be.”

Is Poolside’s synthetic data approach a differentiator?

“I think they’ve taken a very interesting approach to developing code models with code synthesis,”  Pathak said. “Synthetic data is also available in the Llama models, as an example. So, I think synthetic data as a mechanism for training and optimizing [large language models] is a proven technique. I think their application to code completion models and code generation models is interesting.

“Poolside is focused on developing great models that help with coding,” Pathak said. “I think they’ve also built a strong coding assistant framework to help enterprises have a coding system that’s adaptive, that adapts to their code bases, and they can operate within their VPCs. They’ve been very thoughtful about how they’re delivering those capabilities.

“And then, I think, it’s also a very strong team. Obviously, it’s the team that built GitHub back in the day. And think they’ve clearly shown that they know that space really well.”

But what about the tradeoffs to the Poolside approach?

“If we lived in a world of infinites, infinite energy, infinite compute and infinite data, we’d all be doing the exact same thing from an AI frontier lab perspective,” Warner said.

Poolside could pursue a broader approach. he insisted, however,  that his company’s software approach applies in a world where infinities do not exist.

“We only have so much computing, we only have so much data,” Warner said. “So we’ve decided to spend our energy focusing on software.”