DeepL vs Google DeepMind

DeepL Voice now preserves your voice in real-time multilingual conversations

DeepL announces that DeepL Voice now preserves your voice in real-time multilingual conversations, as highlighted in the newsroom with a dedicated announcement dated September 15, 2026.

DeepL Voice now preserves your voice in real time across languages

Real-time voice translation now preserves each speaker's tone and voice across languages, with a unified desktop app for Zoom, Teams, and Google Meet.

Google DeepMind launches institute to widen AGI debate

Google and Google DeepMind researchers launched the DeepMind Institute to advance the conversation around artificial general intelligence (AGI). The institute lists DeepMind co-founder Shane Legg, Google executive James Manyika, and Google DeepMind chair Demis Hassabis as directors, with Legg serving as managing editor. The institute aims to surface differing views between Google, Google DeepMind, and the broader global research community around AGI. The inaugural collection of four essays covers topics such as economic policies for managing potential AGI disruption, preserving human-readable model reasoning, principles for human flourishing, and a framework for evaluating frontier AI models. One essay, by DeepMind safety researchers Rohin Shah and Anca Dragan, argues that AI's shrinking window of transparency is not inevitable and suggests limiting opaque serial depth. Another essay by Hassabis proposes a U.S.-led frontier AI standards body to evaluate advanced AI models, potentially including a coordinated slowdown among frontier AI developers.

• Google DeepMind launched the DeepMind Institute on September 17, 2026.
• Shane Legg, James Manyika, and Demis Hassabis are listed as directors.
• The inaugural collection includes four essays on AI topics.

Z.ai Says AI Model Built Its Own Inference Infrastructure

Chinese AI company Z.ai (formerly Zhipu AI) published a research paper detailing how its GLM-5.3 model, via an Infra Agent, built and optimized the production inference infrastructure on a cluster of over 100,000 Chinese-made AI accelerators. The process from model adaptation to production readiness took under two weeks, with end-to-end throughput tripling. The company reports performance comparable to Nvidia GPUs and introduced 'Dense Feedback,' where an AI agent uses system metrics to autonomously identify and fix bottlenecks, such as reducing a parallelism bottleneck from 20% to under 1%. While not yet achieving full recursive self-improvement (RSI), Z.ai sees early forms of it. Unconfirmed rumors suggest Google DeepMind may have reached RSI, but Google has not commented. The event occurred in September 2026.

• Z.ai (formerly Zhipu AI) published a research paper on GLM-5.3 achieving near RSI by building its own inference infrastructure.
• The inference system runs on over 100,000 Chinese-made AI accelerators, with performance comparable to Nvidia GPUs.
• From model adaptation to production readiness took under two weeks, with end-to-end throughput tripling.

Study: AI Models Learn to Refuse Answers When Uncertain

Researchers at Google DeepMind conducted a study on large language models (LLMs) including GPT-4o, Gemma 3 27B, Deepseek-V3, and Qwen3-Next-80B-A3B-Instruct to investigate how these models decide whether to answer a query or abstain due to uncertainty. Using an experimental paradigm with four phases, they found that models apply implicit confidence thresholds, and that steering their internal confidence levels causally affects abstention rates. The findings suggest that models can be made to refuse answers when their confidence is low, potentially reducing hallucinations. This ability is considered crucial for autonomous AI agents that must recognize their own uncertainty. The study was published in Nature Machine Intelligence.

• Google DeepMind researchers studied GPT-4o, Gemma 3 27B, Deepseek-V3, and Qwen3-Next-80B-A3B-Instruct.
• The study introduced a four-phase experimental paradigm to test model abstention behavior.
• Phase 3 used 'Activation Steering' to causally link confidence levels to abstention rates.