Insights from an AI Milestone

“The generative AI era is upon us, the iPhone moment if you will...” -- Jensen Huang, CEO of NVIDIA, at the SIGGRAPH Conference in Los Angeles on 8/8/23?

What an exciting time to be in technology!?

A victory lap. I'm thrilled that this week Truckstop announced our new pioneering generative AI application, the “RMIS Assistant”. This is the first generative AI tool in the logistics compliance sector and stands as one of the very first in the logistics industry. This achievement was forged from relentless effort and determination by my teams and myself over the past weeks and months. We wanted to harness generative AI in a tangible business context, aiming to deliver real value for our customers, and we’ve achieved just that.?

Let me share what we learned.?

1. Behind the AI Magic, Development is Complex and Hard: Generative AI, while revolutionary, is challenging to integrate and apply. It's one thing to find that “AI can address this particular business need or challenge,” and an entirely different ball game to make it operational. Navigating the uncharted engineering terrain of AI calls for patience, persistence, and creativity. The journey from conceptualization to actualization is harder than you think.?
2. Your Data is Everything: Your company's data remains your greatest asset, particularly when enhancing it using generative AI models like ChatGPT and Claude. These models still tend to be wild and indeterminate without proper grounding by binding them with your data. This is true whether using?static data like knowledge bases, or dynamic data like real-time customer information.?Combining the large language models (LLMs) with your data unlocks and amplifies the value of your data and also ensures the AI models’ reliable performance.?
3. Security: More Critical than Ever: Merging your and your customers’ proprietary data with AI models underscores the need for robust security measures. For now, the most powerful LLMs are the commercial models hosted by companies like OpenAI, Microsoft, and Anthropic. We must make informed decisions regarding the extent we expose our data to be visible to AI. We also need to decide how and when we let our data, or our customer’s data, transit outside our company’s walls. We need to seriously evaluate the security and data privacy commitments of our vendors. This includes our AI and cloud providers, as well as other third-parties in our data path.?
4. There Aren’t Many Experts Yet: While the researchers at the big AI labs are the experts on how AI models?are built, there's a scarcity of?experts who can?apply generative AI to practical business problems. I was talking to a business owner friend who desperately wants to try generative AI to solve his business problems. He lamented that there was not yet a pool of talent that he could call on for help. This will change over time, of course. But?I’m excited that the work that my teams and I have done has put us in the position to now be among those experts.?
5. Continuous Fine-Tuning: The launch of successful LLMs like ChatGPT required the combination of extensive training on massive amounts of public data plus a huge amount of manual fine tuning. OpenAI extensively employed Reinforcement Learning from Human Feedback (RLHF) to optimize ChatGPT. Without this fine tuning, the AI models can produce unpredictable and, at times, highly inappropriate responses. As we introduce our own data to these models, the need for further fine tuning starts over for our own applications.?It’s vital for businesses to recognize the need for human fine tuning which will likely be a continual aspect of the application’s maintenance during its lifetime.?

I'm immensely proud of my teams' accomplishments in pioneering Truckstop's first generative AI application and bringing it to market.?Kudos to the company's leadership for their visionary investment in innovation and cutting-edge?tech.?Our research, experimentation, and iterative development now cements the position that we are now the experts!?