I attended a chat event with Srinivas Narayanan, organised by T-Hub. A few points stood out for me, as they connect closely to what we are doing at Actalyst.
Data security
Right now, if you need frontier intelligence, you still have to send your data to one of the labs. Enterprise contracts explicitly say no training on your data. But are enterprise leaders ready to send sensitive data outside their own cloud?
This is a question we regularly encounter at Actalyst. We work with large enterprises whose production and sales data is highly sensitive.
Srinivas suggested that the next phase would involve conversations around:
Not being tied to a single model. Using a multi-model solution.
Choosing open-source models(which are equally good in most tasks) for such tasks, and closed-source models for other tasks.
The next phase of model development
The talk also discussed how people will use AI. Two points reinforced what we have seen at Actalyst.
Chat as an interface might change. We hear this often from our users. They don’t want to sit and chat to get insights. They want AI to work in the background and bring insights to them. Agents and orchestration have made this possible to some extent.
Less need for fine-tuning. We arrived at this quite early, by 2024. Model intelligence was improving faster than we could keep pace with fine-tuning. For general tasks, frontier models were already capable. Fine-tuning is still needed for certain specialised tasks, for cases where training data is skewed.
Where enterprise AI is going
Each enterprise is unique, and context about their data is still missing. What data should the AI check? Which field should it consider? There is knowledge behind these decisions that sits outside the model layer. Building this context layer is important for enterprises.
We also discussed how AI engineering teams should be structured for these enterprises. Srinivas suggested a McKinsey + Palantir approach. Create the strategy and deliver it using AI.