\n<\/p>\n
ZenML<\/a> wants to be the glue that makes all the open-source AI tools stick together. This open-source framework lets you build pipelines that will be used by data scientists, machine-learning engineers and platform engineers to collaborate and build new AI models.<\/p>\n The reason why ZenML is interesting is that it empowers companies so they can build their own private models. Of course, companies likely won\u2019t build a GPT 4 competitor. But they could build smaller models that work particularly well for their needs. And it would reduce their dependence on API providers, such as OpenAI and Anthropic.<\/p>\n \u201cThe idea is that, once the first wave of hype with everyone using OpenAI or closed-source APIs is over, [ZenML] will enable people to build their own stack,\u201d Louis Coppey, a partner at VC firm Point Nine, told me.<\/p>\n Earlier this year, ZenML raised an extension of its seed round from Point Nine<\/a> with existing investor Crane<\/a> also participating. Overall, the startup based in Munich, Germany has secured $6.4 million since its inception.<\/p>\n Adam Probst and Hamza Tahir, the founders of ZenML, previously worked together on a company that was building ML pipelines for other companies in a specific industry. \u201cDay in, day out, we needed to build machine learning models and bring machine learning into production,\u201d ZenML CEO Adam Probst told me.<\/p>\n From this work, the duo started designing a modular system that would adapt to different circumstances, environments and customers so that they wouldn\u2019t have to repeat the same work over and over again \u2014 this led to ZenML.<\/p>\n At the same time, engineers who are getting started with machine learning could get a head start by using this modular system. The ZenML team calls this space MLOps \u2014 it\u2019s a bit like DevOps, but applied to ML in particular.<\/p>\n \u201cWe are connecting the open-source tools that are focusing on specific steps of the value chain to build a machine learning pipeline \u2014 everything on the back of the hyperscalers, so everything on the back of AWS and Google \u2014 and also on-prem solutions,\u201d Probst said.<\/p>\n The main concept of ZenML is pipelines. When you write a pipeline, you can then run it locally or deploy it using open-source tools like Airflow or Kubeflow. You can also take advantage of managed cloud services, such as EC2, Vertex Pipelines and Sagemaker. ZenML also integrates with open-source ML tools from Hugging Face, MLflow, TensorFlow, PyTorch, etc.<\/p>\n \u201cZenML is sort of the thing that brings everything together into one single unified experience \u2014 it\u2019s multi-vendor, multi-cloud,\u201d ZenML CTO Hamza Tahir said. It brings connectors, observability and auditability to ML workflows.<\/p>\n The company first released its framework on GitHub<\/a> as an open-source tool. The team has amassed more than 3,000 stars on the coding platform. ZenML also recently started offering a cloud version<\/a> with managed servers \u2014 triggers for continuous integrations and deployment (CI\/CD) are coming soon.<\/p>\n Some companies have been using ZenML for industrial use cases, e-commerce recommendation systems, image recognition in a medical environment, etc. Clients include Rivian, Playtika and Leroy Merlin.<\/p>\n The success of ZenML will depend on how the AI ecosystem is evolving. Right now, many companies are adding AI features here and there by querying OpenAI\u2019s API. In this product, you now have a new magic button that can summarize large chunks of text. In that product, you now have pre-written answers for customer support interactions.<\/p>\n \n\t\t\t\u201cOpenAI will have a future, but we think the majority of the market will have to have its own solution\u201d\t\t\t\t\t\t\tAdam Probst<\/cite>\n\t\t\t\t\t<\/p><\/blockquote><\/div>\n But there are a couple of issues with these APIs \u2014 they are too sophisticated and too expensive. \u201cOpenAI, or these large language models built behind closed doors are built for general use cases \u2014 not for specific use cases. So currently it\u2019s way too trained and way too expensive for specific use cases,\u201d Probst said.<\/p>\n \u201cOpenAI will have a future, but we think the majority of the market will have to have its own solution. And this is why open source is very appealing to them,\u201d he added.<\/p>\n OpenAI\u2019s CEO Sam Altman also believes that AI models won\u2019t be a one-size-fits-all situation. \u201cI think both have an important role. We\u2019re interested in both and the future will be a hybrid of both,\u201d Altman said when answering a question about small, specialized models versus broad models during a Q&A session at Station F<\/a> earlier this year.<\/p>\n There are also ethical and legal implications with AI usage. Regulation is still very much evolving in real time, but European legislation in particular could encourage companies to use AI models trained on very specific data sets and in very specific ways.<\/p>\n \u201cGartner says that 75% of enterprises are shifting from [proofs of concept] to production in 2024. So the next year or two are probably some of the most seminal moments in the history of AI, where we are finally getting into production using probably a mixture of open-source foundational models fine tuned on proprietary data,\u201d Tahir told me.<\/p>\n \u201cThe value of MLOps is that we believe that 99% of AI use cases will be driven by more specialized, cheaper, smaller models that will be trained in house,\u201d he added later in the conversation.<\/p>\nPrivate, industry-specific models<\/h2>\n