AI models can fundamentally change how we interact with technology, but there’s still a huge gap between these models and what makes them useful in end products. We started Sieve to bridge this gap.
Programming, content creation, and search are all examples of areas where AI has made a mark. A lot of this progress, however, hasn’t been as simple as an OpenAI API call. It’s months of deep, applied R&D that involves evaluating dozens of models, setting them up on complex cloud infrastructure, and composing them into production-ready pipelines — all things only a few teams in the world know how to do.
Take ElevenLabs Dubbing Studio, for example, which combines speech transcription, LLM translation, text-to-speech, and audio separation, with nuanced logic that enables quality, automated AI dubbing. Or Perplexity AI which combines an LLM with search results to produce beautiful citations along with its summarized answers to questions — all in a few seconds.
If we want more teams to ship great AI products, models shouldn’t be the only tool in the developer toolbox. Developers should have direct access to tasks that models enable such as dubbing, content moderation, translation, etc. A task is a specific use case that a model or set of models might enable. It handles the domain-specific legwork through prompting, relevant pre/post-processing steps, hand-coded logic, fine-tuning, and precise multi-model pipelining. It is packaged with a set of options designed to help developers customize and make tradeoffs without doing the legwork themselves. For example, translation is a task, while GPT4 or SeamlessM4T are models that enable it.
Why reimplement and maintain bank API integrations when you could use Plaid? It’s not that you couldn’t learn how to. It’s that spending the months in the weeds of bank APIs isn’t core to your differentiating product experience and would take many months to execute well.
Developers could integrate these tasks directly, make cost/quality/speed tradeoffs, or combine them to make even more complex tasks. Complex use cases would become trivial to string together, improvements in base models would constantly propagate through the stack, and meeting a certain set of constraints wouldn’t mean having to keep up with every new model release or model provider.
That’s what Sieve is, and it’s why building AI products will become incredibly simple.
Being the interface a developer calls to run AI tasks is valuable because it becomes the distribution mechanism by which any new model or related tool is used.
Regardless of whether open or closed models become more useful for production applications, this allows us to partner with companies that build closed models or efficiently serve open models ourselves in order to add them as backends for the task at hand. In closed-source scenarios, we provide distribution for the model, and in exchange, can take a cut of revenue without having to pay the cost of model inference. In open-source scenarios, we can serve the best ones ourselves at scale and make a profit on top of compute.