Model optimization

Oct 06, 2026

Is Speculative Decoding Worth It? Profiling vLLM on NVIDIA Blackwell — Akamai

Sheilah Kirui explains speculative decoding for LLMs, detailing how a small draft model proposes tokens for a larger target model to verify in a single pass, accelerating the decoding phase. She covers the memory costs of hosting two models, criteria for choosing an effective draft model, and demonstrates its benefits for structured output (1.6x faster with high acceptance) versus its limited gains for creative tasks. The talk concludes by outlining when speculative decoding is most advantageous, considering VRAM availability, batch size, workload type, and context length.

Sep 17, 2026

How Real-Time AI Video Is Changing How Creators Work

fal's H3 Max significantly accelerates generative video, achieving real-time capabilities and substantial cost reductions through advanced model post-training and system optimizations. The discussion covers its technical "secret sauce," the emergence of continuous, controllable video experiences like H3 Max Director, and fal's strategic shift to focus on granular controllability and quality for professional Hollywood workflows, including precise camera, lighting, and motion controls.

Sep 16, 2026

Your Agreements Are a Database You Can't Query — Hiral Shah, Docusign & Sean Sodha, NVIDIA

The talk by Hiral Shah (DocuSign) and Sean Sodha (NVIDIA) addresses the critical challenge of extracting valuable, negotiated data from unstructured agreements, highlighting how generic AI models fail at complex tables. They introduce DocuSign's Intelligent Agreement Manager powered by NVIDIA's purpose-built, small-parameter NeMo Parse VLM, designed specifically for efficient and accurate table extraction, demonstrating 20x faster processing, lower cost, and its impact on unlocking billions in trapped enterprise data.

Jul 23, 2026

The Unreasonable Effectiveness of Separating the Task from the Model — Maxime Rivest, DSPy

DSPy emphasizes separating task definition from model implementation using a "Signature" (inputs/outputs) to enable flexible, optimizable, and scalable AI programs. The framework relies on three pillars—instructions (specs), hard constraints (code), and examples (evals)—to fully specify tasks. DSPy 4.0 introduces DSPy Flex for learning program harnesses and Qualitative Learning for automated, feedback-driven evaluation refinement, offering significant benefits for enterprise applications and addressing "last-mile learning" for future AI systems.

Jul 18, 2026

Your Agents Need a Save Button - Hamza Tahir, ZenML

Most AI agents lack a "save button" for persistent state, making debugging and optimization challenging. This talk introduces the concept of a durable runtime with checkpointing, allowing developers to replay agent executions, perform "what-if" analyses (e.g., swapping models, mocking tools), and conduct cohort-based evaluations using production data. The open-source tool Kitaru is demonstrated as a solution to build, debug, and improve agents at scale by capturing full execution state, not just traces.

Jul 11, 2026

State of the Union: Why Local, Why Now — NVIDIA, Osmantic, Roboflow, EXO Labs, @matthew_berman

This panel discussion explores the inflection point of Local AI, driven by advanced models, improved hardware, and a robust ecosystem. Experts discuss how this shift addresses critical concerns around privacy, cost, sovereignty, and resilience, emphasizing the pivotal role of open-source AI and specialized models. They delve into technical optimizations, the evolution from generalized to specialized AI, and the challenges of making local AI accessible and performant for both enterprise and individual users.