Some predictions for 2024

Augmentation, synthetic data, and the return of neurosymbolic AI

The following list was first shared on LinkedIn on December 19th. I am reposting it here for my non-LinkedIn readers.

---

1. The (re)birth of Neurosymbolic AI
   A hybrid between symbolic AI (which represents human expertise using logic, structure, and interpretable symbols) and deep learning (good at pattern recognition and capturing latent representations), neurosymbolic AI offers a promising path towards intelligence, explainability, and modularity. Recent advances in language models, data annotation, and reasoning offer fertile grounds for these models to boom in the coming months.

2. Augmentation > automation
   A year into the mass adoption of some of the most advanced AI models ever built, it is becoming clear that one-click solutions don’t work. We are moving towards an ecosystem of specialised, context-aware AIs, each working hand-in-hand with the user to augment them in different ways. Copilots, not autopilots.

3. The rise of high-quality synthetic data
   As the demand for data outstrips supply, high-fidelity replicas of existing datasets are going to arise. With the right design (especially around bias, fairness, and quality of the underlying signal), these are already proving invaluable for LLM training, counterfactual modelling, and scenario simulation. Not to mention the many privacy issues these datasets will help circumvent.

4. Better planning with agents
   Where conversational interfaces offer 2-dimensional engagement (send and receive information), agent-based systems open a third dimension of context-aware decision-making and action through complex, ongoing interactions with humans. I’m excited to witness the rise of specialised APIs and emergent operating systems to enable this new wave of agents.
   

5. Data curation as the new moat
   Compute is getting cheaper, AI models are going open source. The only remaining competitive advantage for businesses large and small lies in the quality of their training data. And unlike compute or models, which only gain value once released into the world, there is little incentive to release the underlying data. In navigating the troubled waters of signal and noise, the best curators are bound to thrive.