EN · International edition · 30/10/2026
Inferama daily · 30/10/2026
The most important AI developments since the previous newsletter, selected for editorial relevance.
News coverage period: 28/10/2026 08:30 → 30/10/2026 08:30 (Europe/Madrid).
Google DeepMind introduces Genie 4, an experimental model for interactive worlds
Genie 4 generates explorable environments with audio and lets users modify them through prompts during a simulation. Google DeepMind has released it as a research preview and acknowledges limits in long-term consistency, physics, audio, and legible text.
Read article →
NVIDIA releases Nemotron 3.5 Transcribe for recorded audio, with transcription, translation and speaker attribution
Nemotron 3.5 Transcribe combines multilingual transcription, translation and speaker attribution in an open-weight model for recordings up to one hour long. NVIDIA also released Granary v2 and adaptation tools; the performance results available are those reported by the company.
Read article →
OpenAI updates the Model Spec, but specific changes remain unconfirmed
OpenAI announces an update to its Model Spec to guide the judgment of more capable models, human agency, and faithfulness to intent. Available documentation does not allow verification of the five announced values or their practical application.
Read article →
Anthropic says Claude found five exploit chains and critical vulnerabilities
Anthropic describes using an unreleased internal research model to identify flaws in Chrome, Safari, Android, Windows, and Linux. The company says it reported the findings and that they were fixed, but its account is not independent validation of the results.
Read article →
Epsilon automates UI testing with Nova Act in AWS-managed browsers
AWS describes Epsilon’s UI testing platform, which turns natural-language instructions into tests run in managed browsers. The available information outlines the architecture but provides no reproducible Epsilon-specific metrics or details about human oversight or isolation between customers.
Read article →
AWS Batch adds shared quotas and preemption for SageMaker training jobs
AWS Batch can now divide compute capacity among teams, lend unused capacity, and prioritize SageMaker training jobs. AWS says each job queue can have up to 20 shared quotas, with availability in regions where AWS Batch is available.
Read article →
In an Anthropic experiment, analysts followed their AI agent’s assigned lean in 86% of four decisions
In a causal-analysis task, 50 professionals who used Claude Code chose the option favored by their agent in 86% of four methodological decisions. The result comes from an experiment that deliberately strengthened the agents’ preferences; it is not a general rate of AI compliance.
Read article →