cost function
A mathematical objective an algorithm optimizes, balancing competing factors (e.g., fuel efficiency vs. performance), often encoding value tradeoffs.
As defined in David Mindell on Our Robots, Ourselves
Other definitions
In AI, a measure of how well a system is achieving its goals by quantifying dissatisfaction with predicted outcomes; LeCun identifies this as the seat of emotions and as essential for safe, controllable AI
AI: Grappling with a New Kind of Intelligence | World Science Festival
Where it comes up · 4
- Ilya Sutskever: Deep Learning | Lex Fridman Podcast #942 claims
- Yann LeCun Says LLMs Have 2 Years Left…2 claims
- AI: Grappling with a New Kind of Intelligence | World Science Festival1 claim
- David Mindell on Our Robots, Ourselves1 claim
Most important claims · 6 of 6Search all claims for “cost function” →
- Every time you drive you make tradeoffs—e.g., turning the risk knob up when late or down when carrying kids—and so does every autonomy algorithm, which works by optimizing cost functions balancing factors like fuel efficiency versus performance that are often in conflict; therefore systems should be designed so the user has transparent input into those value-laden decisions, so the car drives according to your values and priorities rather than a programmer's hidden choices.David Mindell on Our Robots, Ourselves
- Cost functions in world model systems might be inaccurate or world models might make incorrect predictions, so objective-driven systems can still make mistakes.Yann LeCun Says LLMs Have 2 Years Left…
- The cost function is a profound foundational idea in deep learning because it allows reasoning about system behavior through optimization, and it is unwise to bet against cost functions as the central organizing principle despite emergence of alternatives like GANs and self-play which use different mathematical objects (game equilibrium, self-competition) for reasoning about behavior.Ilya Sutskever: Deep Learning | Lex Fridman Podcast #94
- LLMs are intrinsically unsafe because they cannot close the gap between training error and test error—there is always a new prompt where they will do stupid things, unlike objective-driven world model systems which can enforce safety constraints by construction through cost functions.Yann LeCun Says LLMs Have 2 Years Left…
- LeCun's proposed JEPA (Joint Embedding Predictive Architecture) predicts representations of video rather than pixels, learning in abstract terms what may happen as a consequence of actions; this approach combined with world models and cost functions could enable planning and safe, controllable AI systems.AI: Grappling with a New Kind of Intelligence | World Science Festival
- All major subfields of machine learning—computer vision, natural language processing, and reinforcement learning—exhibit fundamental unity in that the same core principles (gradient descent, backpropagation, optimization of cost functions) apply across modalities, and architectures will likely eventually unify under a single framework just as vision and language have begun moving toward unified transformer-based approaches.Ilya Sutskever: Deep Learning | Lex Fridman Podcast #94