Sam Witteveen

Designing for autonomy isn’t a tweak—it’s a transformation. In “Designing for Autonomy: How Agentic AI Will Reshape Enterprise Architecture,” we explored how businesses must reimagine their tech stacks to support agents that can observe, decide, and act without human prompting. - Armand Ruiz, VP of AI Platform at @IBM - Moderated by Sam Witteveen, AI Agent Developer This conversation unpacked what it truly means to build enterprise systems that are not just intelligent, but autonomous. From architectural shifts to governance frameworks, Ruiz emphasized the need for infrastructure that supports continuous learning, trust, and real-time decision-making at scale. #VBTransform #AgenticAI #EnterpriseArchitecture #AILeadership #AutonomousSystems #AITransformation

We’ve been building RL gyms for frontier labs and we want to bring similar environments to the open-source community. One benchmark we’re excited about is Toolathlon, developed by the HKUST NLP Group. It’s used by @OpenAI GPT-5.4 to evaluate long-horizon MCP tool-calling, and we think benchmarks like this will become increasingly important for agent research. But evaluating LLM agents on realistic tool use is still messy. Most datasets today are: - Relatively small - Dependent on live APIs that change over time - Slow in running and limited for customization So we built Toolathlon-GYM. A fully local, reproducible environment for long-horizon tool-use agents: - 503 tasks + verifiers - 25 fully mocked MCP servers - Rich mock database - No external API calls Everything runs locally, making experiments stable, reproducible, and easy to compare. It can be used for either training or evaluation. We hope this can make agent research a bit easier for the community.

x.com/i/article/2028…

Google's Opal just quietly showed enterprise teams the new blueprint for building AI agents venturebeat.com/ai/googles-opa…

1/6 Introducing Unified Latents: what if your diffusion model's latents were measured in bits? Instead of relying on dimensionality reduction, we learn a latent AE with explicit bitrate control. Paper: arxiv.org/abs/2602.17270 @emiel_hoogeboom, @TimSalimans

Honored to be presenting at the sold out @CalCities Public Works Officers Institute: "From Concept to Impact: AI Strategies for Public Works Leadership" following a training I'm conducting next week on the same topic for one of the largest cities in the US, via @Ardurra. #CAai