Microsoft’s ‘Open-source ChatGPT’ - The next big thing in Enterprise Adoption of AI

Big news coming out about Microsoft launching (or about to launch) an open-source version of ChatGPT for enterprises.

Context: OpenAI / ChatGPT as a service has been available on Azure since March 2023. Here we are talking about a different, ‘private’ version that enterprises can use.

Why this could be the most significant development in the Enterprise Adoption of AI

Enterprises have been approaching Generative AI with an eye of caution. There are obvious issues and concerns in the adoption.

1. Data Security: Every question, code snippet and context document given to an AI system being used in training or sent to OpenAI or a similar player is a source of concern. Read about Apple’s ban on using ChatGPT internally.

2. Effectiveness and Correctness: Hallucination, boilerplate replies, information without ‘life’ - All these are issues that would make an enterprise concerned. What if a wrong piece of information goes out to the public via the enterprise blog because of a lack of proper human oversight? Read: Lawyers cited bogus case laws ‘invented’ by ChatGPT and got into trouble.

3. (Lack of) Control: Enterprises lose control over the data that is being shared. It’s scary in the situation when AI is being used for use cases ranging from writing a blog to pharmaceutical research. The secret sauces, code, and other proprietary information getting into the hands of an infinitely intelligent, learning system is something enterprises do not want to happen!

It’s not a surprise that Microsoft partnered with Meta to launch LlaMa 2 as an open-source model that enterprises could host and control.

Now comes the next big news!

Microsoft is gearing up to launch an Open Source version of ChatGPT as an Azure service, as reported by MS Power User.

As of the date this article is written, we are unclear about the details, but here is a wishlist from my interactions with the enterprises. Also, the GitHub repo that Microsoft published seems to be removed.

All potential layers could be built on top of the open-source system at hand.

1. The obvious - Opensource, self-hosted version: The ability to self-host the LLM and keep the data flow within the systems is going to be key.

2. ‘Finetuning Packs’: The quality of finetuning makes or breaks the AI system. It also determines the effectiveness of a system for specific use cases. Finetuning packs could be preset fine-tuning principles for specific use cases.

3. The Access Control Layer: A control layer for the enterprise to enable control of information within the organisation. Think of it as Octa or MD Active Directory for AI.

4. The Policy Layer: Enforce policies of the enterprise to the AI usage. Think of AI as extra members of the workforce and deal with the usage, learning, data access and more to be in line with the policies.

5. Analytics Layer: Usage analytics, efficiency, correctness etc could be huge for enterprise systems.

6. RLHF simplified: How can reinforcement learning happen in private enterprise systems to enable continuous improvement in the results? There is scope for simplifying RLHF for enterprises.

PS: This is a developing story. We will keep you updated on this.

What could be built using Private LLMs? Check out these articles »