analysis Databricks has announced that it will acquire generative AI startup MosaicML for $1.3 billion in a deal to make it easier for private companies to train and run their own custom machine learning models.
The deal appears to be a logical step to enable growth for both parties. Databricks’ core platform helps customers store and sort data received from various sources into their own cloud clusters. MosaicML, on the other hand, provides tools to launch custom AI models at low cost.
The two companies have the combined technical infrastructure and expertise to attract large companies looking to train and deploy generative AI systems using their own data. Many small businesses want to adopt machine learning, but are unsure of relying on off-the-shelf models built by private companies.
Most of all, they don’t want to share confidential information with potential competitors and feel uncomfortable using a model if they don’t know exactly how it will work. For example, neither OpenAI nor Google disclose exactly what data is being used to train their models, beyond the possibility that they behave in ways that are difficult to predict.
“Whatever data was used to train a model, that model is going to represent that data, and wherever the model weights go, the data is gone. To actually start deploying , model ownership needs to be respected in order to respect the data “privacy balance,” MosaicML CEO and co-founder Naveen Rao previously said. register.
Databricks and MosaicML reliably indicate what data is used to train the model. The acquisition will give Databricks better AI resources, while giving MosaicML access to better data to build custom private models.
Over the past few months, MosaicML has released two open source large scale language models, MPT-7B and MPT-30B. It claims to demonstrate that the training required by alternative models can cost hundreds of thousands of dollars instead of millions of dollars. -Maker.
Smaller models typically don’t have features like GPT-4, but businesses don’t always need them. Many people want a system that performs well for a particular task and are happy to go for a smaller model if it gives them more control over the system and reduces development costs.
“The economic situation has to be favorable. Ultimately it comes down to how we can optimize it,” Rao said. The Leg.
“Every time you type into ChatGPT, an inference call is made and a word is spewed out. They basically run on a web server with eight GPUs. It costs a lot, which means it costs a lot,” Rao said.
“It’s about really optimizing your stack, hacking multiple requests together and making efficient use of your hardware. More importantly, don’t waste a small amount of GPU time per request unnecessarily. So it’s important to make the stack super efficient.”
The custom AI market is heating up. Working under Databricks will enable MosaicML to reach larger customers in more diverse industries.
Databricks already claims to work with over 10,000 organizations worldwide. With the addition of MosaicML’s tools, we now have a better-than-seeming machine learning pipeline to attract companies from our competitors and retain customers who are increasingly investing in AI.
Databricks said it hopes to complete the transaction in July and that all 64 MosaicML employees will transition to the combined business. ®
