AI
OpenAI launches GPT-Realtime-2.1-mini with reasoning and tool use, cuts voice latency 25%
OpenAI released GPT-Realtime-2.1-mini to its API, adding reasoning and tool-use capabilities to the Realtime mini lineup at no price increase over its predecessor. Alongside the model launch, the company announced it has reduced p95 latency by at least 25% across all Realtime voice models through improved caching, a meaningful performance gain for developers building voice applications.
Anthropic's 'J-space' paper shows models can detect mid-reasoning interventions, raising interpretability stakes
Anthropic published research on what commentators are calling its J-space paper, demonstrating two significant findings: the lab can perform targeted interventions into a model's reasoning chain to redirect its thinking mid-stream, and the model itself is capable of detecting what kind of intervention was performed. The second finding is drawing particular attention from AI researchers, who see it as closely related to model self-evaluation and a step toward more transparent, auditable reasoning systems.
Google DeepMind launches Predicting the Past skill, letting historians query ancient Greek and Latin texts in plain English
Google DeepMind unveiled a new Gemini-powered tool called Predicting the Past, built into Google Antigravity, that allows historians and researchers to analyze ancient Greek and Latin inscriptions without writing code. The skill grounds Gemini in DeepMind's specialized expert models Aeneas and Ithaca, enabling plain-English queries, custom visualizations, and cross-source pattern mapping across historical texts — capabilities the team validated through three case studies developed with epigrapher Thea Sommerschield.
This briefing is part of the archive.
Get the app for full archive access — every daily briefing, sourced and attributed.