Is Databricks + MosaicML now competing with OpenAI, Vertex, Azure and Bedrock?

Unless you've been living under a rock, you'll have seen that #genai and Large Language Models (LLMs) have taken the world by storm. These models, trained on vast amounts of data, have the ability to generate human-like text, making them invaluable for a range of applications from chatbots to content creation.

One of the rising stars in the LLM space was MosaicML. Its open-source platform for training large language models and deploying generative AI tools had started to distinguished itself in the crowded LLM landscape.

Breaking News: Databricks Acquires MosaicML for $1.3B

Today, in a significant announcement Databricks, the leading data analytics platform, announced. that it has acquired MosaicML. This acquisition, valued at $1.3 billion, puts Databricks into the LLM game, while continuing its ongoing commitment to enhancing its AI capabilities and providing more value to its customers.

Side note: Matt Turck , posted having done the math, that Databricks paid $21M per employee . "$21M per employee. That's the price Databricks is paying for MosaicML — a total of $1.3B for 62 employees?"

MosaicML had carved out a niche for itself with its open-source platform that allows organizations to train large language models and deploy generative AI tools based on their own data. Just one month ago, it?launched MPT-7B?an update to its open-source MosaicML Foundation Series models. With its latest release, MPT-30B announced just 4 days ago, MosaicML further demonstrated how organizations can quickly build and train their own state-of-the-art models in a cost-effective way.

What's the Big Deal with Foundation Models (FM)?

Before we delve into how the Databricks acquisition impacts the other LLM providers here's a quick primer on Foundation Models (FMs). FMs are machine learning models trained on broad data at scale, making them adaptable to a wide range of downstream tasks. They are multi-billion-parameter models pretrained on terabytes of multimodal data and can perform complex tasks relatively independently.

FMs require model alignment, which involves guiding, fine-tuning, post-processing, and chaining of FMs to steer their behavior. Guidance includes prompt engineering, while fine-tuning involves narrowing and deepening the capabilities of FMs. Post-processing refines or filters the model's output, and chaining involves building software applications with FM actions as building blocks

Data Management impacts the performance of FMs is directly related to the availability of high-quality, domain-specific data.

FMOps is an emerging discipline that manages the operational capabilities required to manage data and align, deploy, optimize, and monitor foundation models as part of an AI system. It is a combination of traditional operations, data management, and model alignment.

Like any set of technology tools are different ways to set up an FM-based AI system. The choice depends on the specific use case, available technical expertise, and desire to avoid lock-in.

Which leads us to the topic of Open Source.

The Open-Source Advantage

Open-source as we've all experienced is a significant strength. In the context of AI, open sourced LLMs democratizes access to AI technology, allowing organizations of all sizes to leverage the power of LLMs. MosaicML's open sourced MPT-30B comes on the heels of other companies who have seen great success and adoption using this model.

Meta's LLaMA shows that Developers want more access

Meta has also made significant strides in the open-source AI space with its LLaMA (Large Language Model Meta AI) announced Feb 2023. Unlike other popular LLMs, LLaMA's weights can be fine-tuned, allowing developers to create more advanced and natural language interactions with users in applications such as chatbots and virtual assistants. This adaptability has made LLaMA a popular choice among developers in the open-source community.

Meta's decision to release LLaMA's model weights to the research community under a non-commercial license has been a game-changer. It has effectively democratized access to advanced AI technology, allowing developers to fine-tune the model to their specific needs. This is a significant departure from other powerful LLMs, such as GPT, which are typically only accessible through limited APIs.

As usual it's about flexibility and avoiding lock in

Like multi-cloud, open source is about choice. Depending on the resources and expertise in your organization, it can also be about cost.

When compared to other vendors, like OpenAI, Microsoft Azure, Google Vertex AI, and Amazon Bedrock, MosaicML's open-source approach offers unique advantages. While OpenAI and Microsoft Azure offer multiple variants of Generative Pre-trained Transformer (GPT) models, they come with their own set of challenges such as cost, complexity related to tokenization, pricing variations based on usage type, and dependency on the Azure cloud platform.

Google Vertex AI, on the other hand, offers a user-friendly interface and intuitive tools, but it has a limited but growing range of pre-trained LLM models and specific dependencies on the Google Cloud ecosystem. To Google's credit they are rapidly onboarding partners. Amazon Bedrock offers a diverse range of foundation models for various use cases and seamless integration with AWS services, but it also comes with the drawback of vendor lock-in with the AWS ecosystem, and to be honest AWS has been surprisingly behind IMHO in the marketing and positioning of their capabilities.

Why not Open Source?

As usual it's important to note that fine-tuning and large context windows are specialized skills and may not be necessary for every developer or organization. For small companies, a general tool like GPT may suffice, while larger companies may have a dedicated team member in charge of fine-tuning the models or managing large context windows as I listed earlier. So it's important not to get seduced by the allure of Open Source.

What about Responsible AI and Compliance with the EU AI Act?

There are still many questions to be answered by all LLM providers. Responsible AI is one such dilemma. As are the upcoming regulations like the EU AI Act, which I cover in a separate article here.

Interestingly Stanford Research's comparison of some of the Gen AI offerings and how they comply with the EU AI Act, acts as a nice rubric for evaluating all the vendors on their compliance and capabilities.

What does this mean for Databricks Customers?

MosaicML was a relatively new player in the LLM space. They have now been taken off the board via a strategic move by Databricks that will have significant benefits for its customers and partners. By integrating MosaicML's technology into the Databricks Lakehouse Platform, Databricks can now provide generative AI tooling alongside its existing multi-cloud offerings. This will make the Lakehouse an even more powerful platform for building generative AI and LLMs.

But this move is really about choice as well. Databricks has significant relationships with Microsoft, Google, Amazon, with all 3 having invested in Databricks' multiple rounds of funding. Databricks' support and integration with the mega vendors who offer their own LLMs will not change. MosaicLLM just provides Databricks customer's with a tighter package in the future that could be irresistibly differentiated.

Your move Snowflake

Databricks and Snowflake have been duking it out over the last few years. To be fair, IMHO there's room for both and most enterprises use them in different use cases and solutions. However there is an undeniable battle going on between the two with both investing heavily in M&A. Here's a quick comparison of the two from Macrometa 's blog post

Just last month Snowflake acquired Neeva to accelerate search in the Data Cloud through generative AI (paying about $150M in a bit of a fire sale), and also Streamlit and Myst.ai among others.

UPDATE: 6/27 Snowflake announces partnership with Nvidia See?https://lnkd.in/gTmrRuZZ

More acquisitions on the way?

This is the first major move in what I predict will be a wave of M&A in the Gen AI, LLM space. Players to keep an eye on include

Anthropic: Anthropic is a research organization focused on making AI safe and promoting the broad distribution of benefits. Its strengths include its focus on safety and ethics in AI.

Big Science Bloom: Big Science Bloom is a project by Hugging Face that aims to train a large language model in an open and collaborative manner. Its strengths include its open-source nature and its collaborative approach.

AI21 Labs Jurassic-2: Jurassic-2 is the latest generation of AI21 Studio's foundation models, a game-changer in the field of AI, with top-tier quality and new capabilities. Its strengths include its high versatility, top-tier quality, and high performance.

EleutherAI: EleutherAI's GPT-J is a large language model. Its strengths include its open-source nature, large scale, and its ability to generate coherent and creative text.

Snorkel AI: Snorkel AI equips enterprises to build or adapt foundation models and LLMs to perform with high accuracy on domain-specific datasets and use cases, by supporting the most critical component of developing models: the datasets they learn from.