Making AI work, for real
A series of tutorials, from sequential deep neural networks to GenAI to make your AI work, for real, so all that compute does not go to waste when the rubber hits the road.
Francois Belletti PhD
ex Google DeepMind
CTO and cofounder at Levenlight AI
Targeted AI failure modes
Resource consumption and latency.
Energy and compute budgets in our applications are orders of magnitude smaller in our applications while operating under very stringent latency constraints.
Here we give some insights on how to make things work in spite of such challenges with some insights on models ranging from state space to transformers and settings spanning sequential prediction to data generation.
Failures to adapt to real world deployment.
A lot of models work well, in the lab, and fail for various reasons in the real world.
Spending more money to acquire better data and increasing model update frequencies still leave some crucially blind spots unaddressed.
Here we talk about automated failure prevention and model patching to address some of these issues.
Aiming for efficiency
Making models more efficient is crucial to one’s understanding of deep learning beyond just enabling lower latency and serving costs.
Stripping a model down to the essential is a key exercise to understand what makes a model truly work. We’ll have tutorials dedicated to this here.