AI News: OpenAI's Jalapeño Chip, Agent Harnesses, and Local Stacks

OpenAI released preliminary benchmark details for its custom inference chip, Jalapeño, claiming significantly higher work-per-watt efficiency and lower latency than NVIDIA GB200 systems. The company utilized AI models to write and optimize low-level code for the chip, reducing the usual throughput-versus-latency trade-off and signaling a shift in hardware infrastructure development.

In agent engineering, performance heavily relies on the surrounding software harness rather than just the underlying model. Research demonstrates that swapping harnesses drastically alters system scores, prompting new standards like Harness Card disclosures. Meanwhile, memory systems are shifting away from compressed chat histories toward programmable states using append-only logs and persistent Python kernels.

Local-first computing and on-device AI tools saw major updates, led by Perplexity launching its Portable Computer on NVIDIA DGX Spark to run fully local orchestrators and subagents. Apple integrated local clustering tools like exo on Mac hardware using Thunderbolt 5, while tools like Ollama added gateway integrations for desktop applications. Additionally, Alibaba introduced Qwen 3.8 models, which gained widespread open-source adoption, and Figure unveiled Index, a massive robotics dataset designed to overcome physical AI data bottlenecks.

These developments matter because they highlight a broader industry transition toward custom inference hardware, more reliable agent evaluation frameworks, and accessible local-first computing stacks that reduce reliance on cloud infrastructure.