Ai2 released AstaBrief 8B, an open-weight model that turns research questions and retrieved literature excerpts into cited reports. It is available as Fast mode in Asta, alongside the existing Thinking mode; Ai2 also released training data and an example workflow for local PDFs. The team reports average end-to-end times of 51.1 seconds for Fast mode and 178.5 seconds for Thinking mode. That is a useful starting point for research assistants where waiting time matters. There is an important limitation: much of the evaluation was completed in 2025, and Ai2 says it has not rerun the full comparison against current frontier models. Learners should inspect citation support and whether conclusions preserve the original study’s scope, rather than treating a polished report as verified science.
AI news
Ai2 opens AstaBrief while Google makes federated training more auditable
Google Research announced a federated-learning system that uses trusted execution environments to make server-side processing more verifiable. Devices encrypt training examples and authorize permitted computations; access policies are published in a transparency log, while the protected workloads release anonymized results. Google says Gboard has already used the system for English and Japanese next-word prediction, improving accuracy and training speed. The architecture moves more computation to servers, so the privacy story depends on the stated hardware assumptions and safeguards, rather than simply keeping every operation on a phone. For builders handling sensitive training data, the practical lesson is to examine which code can access it, how that code is verified and what outputs may leave the protected environment. For learners, this is a concrete case study in combining differential privacy, encryption and auditable execution, with current trusted-hardware limitations still relevant.