Hello World! It’s me back from my writing hibernation. Blame it on my writer's block or being engulfed in a never-ending journey of learning Generative AI, ChatGPT, AutoGPT, Vector embeddings, Langchain, and a slew of new developments that come up every day. I thought it was insane when Satya Nadella and Bill Gates predicted Generative AI space to be bigger than the Internet and Mobile, but Sundar Pichai blew everything when he compared it to electricity. This all could be an exaggeration, but the pace of innovation in GenAI space is nothing like what we saw in the past.
Think of significant innovations post iPhone launch - every new innovation took months if not years - be it appstore or Geo-tracking apps etc. But compare it to innovations since the launch of ChatGPT 3.5 - AutoGPT, Langchain and the list goes on and all these were launched in months if not weeks. GenAI is the new buzz in the startup world and does anyone even talk about Crypto, Metaverse, and Web3 anymore? Every Enterprise is on high alert to make sure they don’t become the next Blockbuster or Nokia. If you are an Enterprise executive, I’m pretty sure you received a memo from your C-Suite to submit your GenAI plan.
If you are an Enterprise executive, it's easy to get overwhelmed on where to start. As always, go to your basics and first identify the use cases you want to target. Then decide on the fastest way to market with minimal risk. The fastest way is the vanilla chatGPT implementation by calling out the foundational model APIs such as OpenAI. This works great if you are an SME, However, if you are a large enterprise dealing with sensitive patient data or financial data, you will have a lot of things to worry about
How do I ensure my sensitive data doesn’t end up with a vendor or even worse end up in bad hands? Remember the chatGPT Samsung incident?
How do I ensure my data security and permissions are honored?
How reliable are the responses generated by my chatbot? Can I provide it to my end customer without any human validation?
You are 100% right if you are concerned and you should be. In a world where access to AI gets increasingly democratized, the true differentiator is your data, So needless to say your GenAI system needs to be built upon trust and security. However if you choose to restrict your data with a vanilla ChatGPT implementation, you will lose out on relevance. The generic responses generated by assistants such as ChatGPT will be irrelevant or incorrect. You need to either provide context to the foundational model (via prompt engineering) or finetune the LLMs with your custom data to make it both relevant and reliable.
If you are an Enterprise trying to build your custom GenAI solution, the implementation can easily take you into a rabbit hole. As said earlier a vanilla chatGPT implementation has no context of your business nor has a memory of your conversations. So, you'll need to build a comprehensive infrastructure to augment the foundational model and also to give your chatbot long-term memory and context. There are frameworks and tools such as Langchain and Retrieval Augmented Generation (RAG) to help you, but you might end up with a massive infrastructure to maintain including vector databases, Data ingestion tools, In-memory databases, etc.
There are additional cost implications and effort required for continuous fine-tuning and retraining to provide accurate results. I haven’t even started on data security, compliance, etc. My purpose is not to plant fear, but to make sure you understand what you are getting into. Finetuning and Augmenting a foundational LLM is a viable option for organizations that have established a robust MLOps process and possess the requisite knowledge to create, customize, manage, retrain, and govern their own ML models. However, for those who are inexperienced and new to the field, relying on pre-trained models can be a risky and costly endeavor in the long run
So, How do I get started?
First things first. Any AI project begins with data. You need to first get your data in place. If your data source is spread across disparate platforms, get your data harmonized via a data management platform.
Once you identify your use cases, focus on the user experience. How am I going to deliver my AI predictions embedded in the user workflow?
Am I confident to deliver my AI-generated responses directly to my end customer or will it require a human in the loop?
Do you have a preferred application development platform? If so, what capabilities do they offer? Do they provide a secure AI Stack that can honor all your Org permissions?
There are multiple vendors who have released their own GenAI strategy and help you get to market faster. One of the best AI architecture and Enterprise AI demonstrations I saw is from Salesforce EinsteinGPT. Full disclosure - I work for Salesforce and I may be biased, but let me explain why I think so:
EinsteinGPT allows you to use any foundational model, but at the same time ensures your sensitive customer data is protected.
It lets you augment or finetune the foundational model with your very own customer data.
Security is baked in via PII data masking, toxicity/bias checking on generated responses and finally assuring LLMs don't retain or use your data for training.
You can not only use data in the Salesforce platform but use data from other data sources via Data Cloud which is a Data lakehouse native to the platform.
Finally, the most important part is the human feedback. They baked it right into the stack, for example, a Call center Rep on the phone gets an AI-generated response that is relevant thanks to finetuned data and also gets a chance to review and edit before they send it to the end customer.
As a customer, you don’t have to do any of the heavy lifting and can easily introduce Generative AI capabilities in your business without compromising on security and compliance. Good luck with your AI journey and lastly, I want to end it with a quote “I am not sure if AI will replace humans, however, I'm sure humans and organizations that know how to use AI will replace the ones that don't.” - Anonymous.
Image credit - Marketoonist.com
Kommentare