It’s boom time for Large Language Models (LLM).

The early movers, such as GPT-3 and GPT-4, and Google’s LaMD and Bard, have become very popular and unlocked several cutting-edge applications.

The latest LLM off the block is Databricks RX or DBRX.

How is DBRX Similar to Other LLMs?

First the similarities.

Like all other LLMs, DBRX power gen AI applications that understand queries and generate answers in human text. It creates content, answers queries, translates languages and generates code. DBRX does everything that GPT, Bard, LaMD and other LLMs deliver.

DBRX uses established techniques as well. It uses transformers, attention mechanisms, and mixture-of-experts (MoE) architecture. These techniques already find use in many other LLMs.

The Key USPs of DBRX

So how is DBRX similar or different from the other LLMs already entrenched in the market?

Openness and Transparency

The key USP of DBRX is its open model. DBRX makes available AI development capabilities for open-source users.

Closed model LLMs, such as GPT-4 operate like a “black box”. Developers and users cannot see how the model works, or modify it. This creates trust issues and becomes a key stumbling block for the widespread adoption of gen AI.

DBRX comes with an open licence, making it transparent. Developers can see the code and training data to understand the model’s inner workings. With developers in control of the model, there are no trust issues.

The open model of DBRX means greater community involvement. Open source contributions boost innovation and faster improvements.

The open nature of DBRX also makes it the best fit for responsible AI (RAI).

DBRX is not the first open-source LLM though. Meta’s Llama2, Mistral AI’s Mixtral, and Bloom are also open-source LLMs. But DBRX outperforms these LLMs in benchmark comparisons. It outperforms Gork-1, Mixtral and LLaMA2-70B in programming and mathematical reasoning.

Improved Customisation

Closed-sourced LLM, trained on public data, gives generic answers. These models do not have knowledge of specific businesses, and cannot generate business-specific replies.

The open model of DBRX allows enterprises to customise the model to suit their specific needs. Businesses can fine-tune the model on specific datasets, to make the model understand specialised knowledge and language patterns, and answer business-specific queries.  

For instance, a retail firm could fine-tune DBRX to make the application project the retail outlet’s sales over the next month. 

Closed source models also allow fine-tuning, but DBRX offers more flexibility and control. DBRX users can go in for unhindered customisation, without being hindered by the closed source models terms and conditions or opacity. 

Databricks customers can pre-train their models from scratch or continue training on top of prebuilt models. They can create custom gen AI apps without vendor lock-in, using proprietary data to train the model.

Superior Performance

DBRX claims super-efficiency as another USP.  DBRX does the same things older AI models do, using four times less computing power.

The behaviour of any LLM depends on its parameters. The creators train the model on word meanings, grammar rules, patterns, and relationships. The LLM stores such knowledge in the parameters. When the user inputs a prompt, the LLM rely on the parameters to generate the response.

Databricks has designed DBRX ground-up for efficiency. It adopts a fine-grained approach to deliver superior performance. The model uses fewer active parameters compared to what similar models use.

DBRX uses 132 billion parameters to process information and generate answers. To answer any query, it uses 36 billion parameters.

In comparison, GPT-3 uses 175 billion parameters, and uses all when answering any input.

Improved Efficiency through MoE Architecture

DBRX adopts a mixture-of-experts (MoE) architecture that improves LLM efficiency and accuracy.

The MoE architecture deploys many smaller AI models, each specialising in different things. For instance, one model may be adept at translation, another in law, another in writing code, and so on.

A “gating network” acts like a traffic controller, deciding which “experts” or smaller AI model best suits a specific task. For instance, a legal query will redirect to the AI model trained on law.

The MoE model activates only the necessary resources for each task, instead of using the complete resource for every task. Such an approach saves energy and makes the LLM faster.  

The MoE model also makes the LLM scalable and resilient. Developers can add new “experts” instead of re-training the algorithm.

To put things in perspective, DBRX has 16 experts and chooses four when processing replies. In contrast, Mixtral and Grok-1, other LLMs with MoE architecture, have eight experts and choose two when processing replies. DBRX offers 65x more possible combinations of experts.

Improved Accuracy and Consistency

DBRX ensures high model quality through more parameters and training data.

DBRX comes pre-trained with 12 trillion tokens of text and code data. These curated data come with a maximum context length of 32,000 tokens. This is 2x better than other comparable LLM models. The standard GPT-3-5-turbo has a token limit of 4000. 

The larger context window, that manifests through a larger number of tokens, allows the model to process and remember large chunks of text or code. This makes the model more competent to handle complex tasks, such as processing longer documents, writing extensive code, or generating detailed summaries. Many LLMs falter or hallucinate over time. But DBRX maintains coherence and consistency over longer stretches.

Speed

There is often a trade-off between the AI model’s accuracy and the speed at which it delivers the output. Often, speed comes at the cost of accuracy.

More parameters and training data make the model complex, and slow down query responses. But DBRX makes perfect trade-offs to generate text twice as fast as other comparable AI models. 

When hosted on high-tech delivery services such as Mosaic AI Model Serving, DBRX creates text at the speed of 150 words per second for each user! Such high speed makes viable real-time chatbots, interactive storytelling, and on-the-fly content creation. In comparison, GPT-3.5 Turbo creates text in the range of 40 to 60 words per second.

Open-source LLMs, while allowing customisation. often lag closed-source LLMs in accuracy, speed, and efficiency. DBRX alters this paradigm.

Databricks estimates the new LLM to surpass the capabilities of OpenAI’s GPT-3.5 and considers Google’s Gemini 1.0 Pro as its competition.

With DBRX, every enterprise can build AI systems easily.