TRACE

Laminar, a startup from Y Combinator's Summer 2024 batch, has launched a reinforcement-learning-trained model designed to automatically detect errors in AI agent traces, the step-by-step logs of what an agent did. The launch is drawing attention from developers building AI agents, who face growing difficulty debugging and evaluating multi-step agent behavior in production.

A detailed technical explainer traces how Intel's 8087 math coprocessor performs floating-point addition and subtraction, using its internal microcode. The chip, released in 1980, brought hardware floating-point arithmetic to IBM PCs and shaped the x87 instruction set that influenced computing for decades. Engineers and computing history enthusiasts are sharing the piece and debating the design choices of this landmark silicon.

A lecture document from Peking University traces the history of curves and surfaces in computer aided geometric design, covering the mathematical foundations behind splines and related techniques used in engineering and graphics. Readers are sharing the text as a accessible deep dive into how modeling geometry developed, appealing to those interested in mathematics, CAD software, and the origins of tools still used across industry today.

Researchers have developed a hybrid chemical sensing method that combines the natural olfactory abilities of locusts with a surface-enhanced Raman scattering substrate. A highly porous collagen foam is applied to the insects' antennae and wings, sharpening their ability to detect trace chemicals and improving sample analysis. The work points to new uses for insect-based sensing in detecting hazardous or low-concentration substances.

The Globus team has published a retrospective on GridFTP, the data transfer protocol developed to move large scientific datasets quickly across research networks. The piece traces the protocol's history and its role in physics collaborations and other big-science projects, prompting discussion among researchers and infrastructure engineers about its lasting influence on high-speed data movement.

Open-weight AI models, whose parameters are freely available for download and deployment, may carry hidden supply-chain risks, according to a report by Data Protection Report. Because these models are often built on shared datasets, pre-trained components and third-party tooling, organisations adopting them can inherit vulnerabilities or compliance problems that are difficult to trace, raising concerns for businesses and regulators weighing the benefits of openness against security and data protection obligations.