Google DeepMind has launched Gemini 3.1 Flash Live, a real-time speech/audio model designed for low-latency conversations and voice agents. The company claims it scores 90.8% on ComplexFuncBench Audio and 36.1 in the 'thinking' mode of Scale AI Audio MultiChallenge, while maintaining conversation context longer than its predecessor. The model is available via the Gemini Live API in Google AI Studio and is also used in Search Live, Gemini Live, and enterprise customer service solutions. All generated audio includes SynthID invisible watermarking.

The European Parliament has adopted its negotiating position on the 'AI Omnibus/Simplification Package,' proposing to delay key provisions of the EU AI Act: the effective date for high-risk AI systems will be postponed to December 2, 2027, and compliance deadlines for systems covered by sectoral legislation extended to August 2, 2028. Requirements for watermarking AI-generated content must be implemented by November 2, 2026. The Parliament also supports banning non-consensual sexual deepfake systems such as 'undressing/nudification' tools. The proposal still requires trilogue negotiations with the EU Council and Commission, leaving businesses facing policy uncertainty.

U.S. defense AI company Shield AI has announced a $1.5 billion Series G round, achieving a post-money valuation of $12.7 billion. The funding includes $500 million in preferred equity and a $250 million delayed draw facility. The company plans to acquire Aechelon Technology, a high-fidelity simulation and synthetic reality software provider, to accelerate development of its Hivemind 'AI pilot' software and X-BAT program. Hivemind has reportedly completed flights across 26 aircraft types, including the F-16, and has been selected by the U.S. Air Force as one of the autonomy providers for Collaborative Combat Aircraft (CCA) missions.

The ARC Prize Foundation has released ARC-AGI-3, a new benchmark for general intelligence agents, introducing an interactive, turn-based environment that requires models to explore, model, and plan actions without explicit instructions—reducing risks of data contamination and overfitting. Public reports indicate humans achieve 100%, while the best frontier models score only 0.37%, with others similarly below 1%, highlighting poor generalization in novel environments. The benchmark has launched a $2 million prize competition on Kaggle to encourage more effective general problem-solving strategies and training paradigms.

According to The Information, Apple is using 'knowledge distillation' to deeply reverse-engineer Google's Gemini large model, compressing its knowledge and reasoning capabilities into smaller local models suitable for on-device operation in devices like the iPhone 17. The report states Apple has full internal access to Gemini within its data centers and can modify and use its outputs to train proprietary micro-models. Future Siri may route complex queries to the full, un-distilled Gemini, while local models handle high-frequency, low-latency tasks—improving response speed and enhancing privacy. This effort is led by Apple's Applied Foundations Model (AFM) team.

Mistral has released Voxtral TTS, an open-source text-to-speech model supporting nine languages including English, French, German, and Spanish, positioned as a real-time interactive speech generation component. It reportedly achieves a Time to First Audio (TTFA) of 90 milliseconds and a real-time factor (RTF) of approximately 6x, generating 10 seconds of audio in about 1.6 seconds. The model enables voice customization from as little as 5 seconds of sample audio, preserving accents and intonations. Emphasizing compact size, it is designed for deployment on edge devices such as smartphones and wearables, competing with voice products from ElevenLabs and OpenAI.

Cohere has released Cohere Transcribe, an open-source speech-to-text model with 2 billion parameters based on the Conformer architecture, designed to run on consumer-grade GPUs and support 14 languages. It achieves an average Word Error Rate (WER) of 5.42 on the Hugging Face Open ASR leaderboard and reportedly wins 61% of comparisons in human evaluations. The model takes raw audio waveforms as input and outputs punctuated text; language must be explicitly specified, with no support for automatic language identification or speaker diarization. Licensed under Apache 2.0, it can be deployed via Transformers or vLLM, suitable for offline transcription and production transcription pipelines.

Zendesk has announced the completion of its acquisition of AI agent platform Forethought, launching 'Forethought AI Agents by Zendesk' as part of its Resolution Platform roadmap. These agents are capable of operating both within and outside the Zendesk platform, automating routine tasks across chat, email, and voice channels, and integrating with existing workflows and tech stacks. Zendesk emphasizes that the agents continuously improve from each interaction, enhancing problem resolution efficiency and service quality. The company initially disclosed the acquisition intent on March 11, 2026, and now confirms the deal has closed following standard closing conditions and regulatory approvals.

The governor of Washington State has signed HB 2225, making it the first U.S. state to enact legislation specifically protecting minors from AI chatbots. The law requires platforms to detect potential self-harm signals during interactions with underage users and provide crisis support resources, while restricting inappropriate or exploitative content. It also bans manipulative 'dark patterns' that exploit emotions such as loneliness, guilt, or fear of abandonment to prolong teen engagement. Supporters argue the law fills gaps in platform self-regulation, while the tech industry warns it may impose excessive constraints on broad AI tools. The law is expected to take effect early next year.

The University of British Columbia (UBC), in collaboration with Sakana AI, Vector Institute, and Oxford University, has developed an 'AI Scientist' system that can end-to-end automate research: generating ideas, designing experiments, writing code, analyzing data, writing papers, and self-reviewing. The team reports submitting fully AI-generated papers to an ICLR workshop that passed peer review, and they have built an automated review system whose acceptance rate predictions closely match human evaluations. The work was published in Nature. The team acknowledges current limitations, including immature idea generation and inaccurate citations, with the system primarily applicable to computer science today.