AI models are now exceeding human expert performance on comprehensive benchmarks like MMLU, which measures intelligence across various subjects.
The transformer architecture enabled a major shift from training on limited, human-labeled data to pre-training on vast amounts of unlabeled internet text using next-token prediction.
Pre-trained models are made useful for specific tasks like chatbots through post-training methods such as supervised fine-tuning and reinforcement learning from human feedback (RLHF).
Reasoning models improve accuracy by using test-time scaling, a process where the model prompts itself to double-check facts and logic before providing a final answer.
Developers can build AI applications by starting with an API, using structured prompt engineering, and evaluating models in context rather than relying solely on benchmarks.
Protect your AI application from manipulation and misuse by implementing guardrails, detailed system prompts, and specialized guard models to enforce desired behaviors.
Agentic applications use a modular architecture where each agent can use specific tools, often defined with natural language prompts, to perform complex tasks.
This Q&A covers why LLM responses are non-deterministic, how synthetic data is used for model distillation, and strategies for preventing hallucinations.
31:05Ankit Patel, Matt White, Philipp Schmid, Lucie-Aimée Kaffee & Andreas Blattmann
28:05Toju Duke
25:17Scott Hanselman
32:26Krzysztof Cieślak
22:07Ankit Patel
27:26Cassie Kozyrkov
31:50Cedric Clyburn & Legare Kerrison
26:25Stephan Gillich
Jobs that call for the skills explored in this talk.
Deloitte Ag