HelixML has announced that their Helix platform for Generative AI is production-ready at version 1.0. Described as a "Private GenAI Stack," the platform provides an interface layer and applications that can be connected to a variety of large language models (LLMs). It can be used to prototype applications, starting with just a laptop, with all components version-controlled to ease subsequent deployment and scaling of apps that prove useful. There’s also a heavy emphasis on evaluations (evals) as the substitute for tests in the non-deterministic domain of LLMs.
Helix was launched in December 2023 and found some initial traction with a German company concerned about compliance with European Union regulations. In a "vision" blog post titled "The Open Source AI Revolution: How Regulation Is Reshaping Enterprise GenAI" HelixML co-founder and CEO Luke Marsden explains that "Open Source models can be run locally to avoid sensitive data being shared with US based service providers". He goes on to note that such models now offer similar capabilities to those of frontier proprietary models a short while ago, so for many use cases it's possible to achieve similar performance without having to sacrifice control over data.
Having set out to facilitate fine-tuning of models, the HelixML team discovered that it was of less interest to customers, mainly due to retrieval-augmented generation (RAG) gaining traction as a quicker way to accomplish the same outcomes. So Helix 1.0 ships with RAG capabilities. Though Marsden notes:
I believe fine-tuning will make a comeback when people get further down the line and want to optimize specific systems that use large general purpose LLMs (e.g 70B) and "distill" a fine-tuned LLM in 3B so they can scale to production traffic without much cost. We still support fine-tuning in the product, and we’ll get there…
Marsden has also published a "product" blog post "Announcing Helix 1.0 – Secure Local GenAI for Serious People" that runs through the implementation details of the platform: its architecture, the interface layer, applications, and choice of underlying LLMs. The platform has been written in Golang, and is deployed as a set of containers, so it’s sympathetic to well-established industry norms around deployment and operations.
By building what amounts to integration middleware for GenAI-based applications, Helix doesn’t have to involve itself in the high-cost aspects of the domain - building infrastructure and training models. But they don’t have the space to themselves, with competitors like Griptape selling a similar story. Marsden concludes his "vision" post with a sober forecast for 2025 that holds a final note of optimism for the longer term:
Yes, there’s a trough of disillusionment coming for GenAI. But through the trough of disillusionment comes the plateau of productivity. GenAI is just mathematical models, and models don’t generalize beyond their training data. This stuff is not going to take over the world. The capabilities will plateau. Nevertheless, super-human scale knowledge processing capabilities will change the business world for good.