Agent oversight moves into silicon as chip demand outruns forecasts

Nvidia is putting agent monitoring directly on its BlueField-4 DPU, Devin says it is now up to 40% more cost-efficient, and a 2024 forecast for chip industry growth already looks far too low. CERN also opened up a portable FPGA library.

Running agents cheaper and under watch

Two of today's items deal with the operating side of AI agents rather than what they can do. Nvidia's open agent safety platform is a reference for continuous in-silicon agent monitoring, with the monitoring running on the BlueField-4 DPU instead of the host. Keeping oversight on separate silicon is meant to protect against compromises that could otherwise spread to host systems, so the agent is watched from outside the machine it runs on. On cost, the Devin team says the agent is up to 40% more cost-efficient, a change that could make it easier for companies to justify agent spending against the work it produces.

Chip demand and AMD's World Labs deal

Semiconductor demand appears to be running well ahead of recent expectations. Semiconductor Engineering reports that TSMC's 2024 projection of a $1 trillion chip industry by 2030 was already $700 billion too low. If growth tracks that gap, supply chains and chip pricing could feel the pressure well before 2030. Separately, AMD has announced it is acquiring World Labs, a move expected to add to AMD's AI and machine learning capabilities as it pushes further into that market.

New options for embeddings and model hosting

For teams building search, retrieval or other language features, Cohere has launched its Embed 5 models, which the company positions as more capable and versatile for natural language processing and related tasks. On the deployment side, Baseten has announced a partnership with OpenAI to integrate OpenAI's technology into its platform, with the stated aim of simplifying model deployment and making those capabilities easier for developers to reach from their applications.

Portable building blocks for FPGA designs

CERN has released colibri, a vendor-independent VHDL common library for FPGAs. Because it is not tied to one vendor, designs built on it should be more portable and interoperable across devices, which matters for scientific instruments as much as for commercial hardware. For engineers who move FPGA work between toolchains, a shared, openly available library from CERN may cut down on vendor-specific rework.