Expectations are mad in the world of AI.

As highlighted by Sequoia’s article, the initial excitement has led to a firestorm of innovation and confusion. Many first-time companies are missing the mark, and there’s a shift toward focusing on vertical applications where generative AI (genAI) is a feature rather than a standalone product. The practice is also evolving rapidly: last year was about fine-tuning models, and this year, it’s about RAG, not to mention the new language models appearing every two weeks.

Unsurprisingly, there’s a widespread confusion about what AI can and cannot do.

This general disenchantment is not just a misunderstanding. While something might work flawlessly during testing, putting it into production often reveals challenges: model-bound prompt engineering,

• Prompt engineering is bound to their language models.
• QA is difficult.
• Complex chains increase the chances of errors, invalidating the whole output.

Even the fundamentals are still nascent, RAG is becoming an art and architectures like Mixture of Experts still have much development ahead.

We are not there yet.

However, there's already so much fantastic potential in genAI that's often overlooked. It's not just about writing like a (weird) human!

There is real power already in:

• Connecting topics or concepts that are loosely related.
• Extracting valuable information from large raw inputs.
• Categorizing data effectively.

And maybe my favorite:

• Machines aren’t limited by brainpower or time; they can explore hundreds of hypotheses and consolidate them, while an analyst might explore a few, hoping their intuition is good.

References

1. Huang, S., Grady, P., & GPT-4. (2023, September 20). Generative AI’s Act Two. Sequoia.
2. Wang, X., Wang, Z., Gao, X., Zhang, F., Wu, Y., Xu, Z., Shi, T., Wang, Z., Li, S., Qian, Q., Yin, R., Lv, C., Zheng, X., & Huang, X. (n.d.). Searching for Best Practices in Retrieval-Augmented Generation.
3. Wang, Z. (Andy). (n.d.). Mixture of Experts: How an Ensemble of AI Models Decide As One. Deepgram.