AI Journey: Increase Value & Impact - Part 2

0:00 - Kickoff of part 2
0:41 - Manifesto - don't build your own LLM
2:11 - Cautionary tales of companies building LLMs
3:00 - Building the whole chain from prompts to finetuning to RAGs
4:24 - Agents as the future of GenAI
6:20 - The tradeoffs between cost, resource utilization and results
7:50 - Finding big enough problems to solve with GenAI
9:30.- Bigger is not always better in LLMs
11:50 - How Google created the opportunity for Perplexity to exist
12:05 - Adobe and Microsoft reducing friction in GenAI usage
15:10 - GenAI repeated the music licensing wars much faster
16:50 - What we learned from our 4E poll
18:00 - How Delvin AI may change the coding game
18:43 - The massive productivity increase of Github Copilot
In our latest episode of Digital Value Creation, my brother and I continued our discussion on the personal AI journey from entertainment through education to efficiency and effectiveness. How to increase value & impact. We started with a manifesto: don't roll your own large language model (LLM)! With over 500,000 open-source LLMs available, we believe most businesses don't have problems complex enough to require developing a brand new model from scratch. Instead, companies should invest their time and resources into fine-tuning existing models and optimizing prompts, search, and other parts of the AI pipeline.
We emphasized that the real edge for businesses is figuring out how to integrate AI outputs into their unique workflows and processes through the use of agents. Rather than just talking to us, AI needs to actively help accomplish tasks and solve real business problems in an economical way. This requires carefully considering the cost-benefit tradeoffs in terms of compute resources, context window sizes, precision, and more.
Usability and seamless integration are key - adding friction defeats the purpose of AI automation. We discussed the fast-moving legal landscape around AI and how some companies are proactively addressing it, such as Adobe training on licensed images. The battle between AI companies and content owners is quickly evolving into a licensing model.
Finally, I shared results from a poll showing nearly 30% of respondents are already using AI for business effectiveness, with coding as the top use case. While there are open questions around IP protection, AI is proving valuable for internal code refactoring, testing, and documentation. Exciting new innovations like AI software agents show the potential for AI to resolve real-world coding issues at increasing rates.
The AI journey is evolving rapidly, and we'll continue to discuss the latest developments. Please subscribe, like, and reach out with your own thoughts and experiences!