What is the state of Gen AI? Interesting discussions lately. It really depends on the team you’re speaking with as swim lanes matter as the field grows. Let me explain.
Hardware folks:
There’s a sense that the market in AI infra will continue to expand/differentiate, and companies are trying to segment it. Inference dedicated chips, and other flavors of segmentation. We’re starting to hear more about mobile-oriented AI chips / edge. And folks are trying to figure out not just the chip but the whole stack solution (PyTorch, etc).
Research folks:
The AI scientists are working on the next generation LLMs. And I don’t mean the one that’s coming out later in 2024, that’s been baked already and would be undergoing alignment testing, etc. They’re working on the AI LLMs / architectures for 2025. I think we’re going to see some pretty cool stuff. Multimodal continues to impress. Still ongoing discussion on whether SOTA latest model is better than smaller domain-specific/fine tuned. Only evals / test datasets can be the judge; go benchmark it!
Innovation folks:
Last year (2023), innovation teams in large enterprises were building PoCs of various LLM-based user experiences. Now that those are in the hands of the core tech teams, being industrialized, hardened, and deployed, that’s not the game of innovation anymore. Innovation is always ahead. These folks are now focused on agents, orchestration, digital workers, proprietary models, etc.
Enterprise CIO / CTO teams:
These folks are rolling out GenAI at scale. And now that those teams are pushing out these products, the ROI queries start rolling in from the CFO and transformation offices. What’s the impact, how did you measure it, is it going to be worth the spend in compute and change management. The orgs are maturing and the ROI ask is real.
Software / cloud vendors:
Every software company has some GenAI baked into their marketing brochures if not their products. Enterprises are trying to filter out the wheat from the chaff. How well is it implemented? How do I compare two comparable products from competitors, which do I choose, etc. Now comes the work of making it work seamlessly and affordably.
Safety:
We still haven’t solved hallucinations. There are a lot more techniques to diminish these now though. We have RAG, where the answers come from your data, we have mixture of experts, where we make sure we don’t have outlier models, we have safety modules where LLMs can be used to test other LLMs. It’s moving forward.
Knowledge of how to use this tech:
More people are getting better at creating prompts, knowing when and how to use AI, learning that the tools now can browse the web and be more capable. And we exchange notes with friends and colleagues, and enterprises have developed educational programs and developing this skill with their people.
What do you think? What swim lanes have I missed? And if you’re in one of these communities, does it resonate?
Read more about The Current State of Gen AI.