Google DeepMind Restructures Leadership as Demis Hassabis Steps Back and Jeff Dean Departs to Launch Startup
Google's artificial intelligence division is undergoing a profound structural and operational transformation under its newly appointed leader, Senior Vice President Koray Kavukcuoglu. Kavukcuoglu, a 13-year DeepMind veteran who led pioneering breakthroughs like WaveNet and DQN, has taken over day-to-day operations of Google's main AI research and product teams, reporting directly to Alphabet CEO Sundar Pichai.
This shift occurred as co-founder Demis Hassabis stepped back from day-to-day management to become Chairman of Google DeepMind and Alphabet's Chief Scientist. Concurrently, legendary Google pioneer Jeff Dean departed after 27 years to launch Discovery Loop—a public benefit corporation focused on using AI to automate scientific and engineering research. Discovery Loop is co-founded by senior Google AI giants Oriol Vinyals, Sanjay Ghemawat, and Quoc Le, and is backed by Alphabet itself as a founding investor and cloud computing partner, alongside co-leads Radical Ventures and Khosla Ventures.
Flagship Delays and the Audio Consolation Prize
This massive leadership overhaul is taking place at a critical juncture, as persistent delays of Google's highly anticipated flagship Gemini 3.5 Pro model threaten its competitive edge against OpenAI's GPT-5.6 family and Anthropic's Claude 5. While Google has been actively testing Gemini 3.5 Pro with select enterprise partners, it has delayed the official launch once again, leaving Google with an empty flagship tier:
"Google has updated Gemini Audio with new transcription capabilities... and comes as we’re still waiting for Google to release the Gemini 3.5 Pro model that it promised to roll out in June."
To fill this flagship gap, Google has rolled out Gemini 3.5 Transcribe and Gemini 3.5 Transcribe Live (as part of the new Gemini Audio family alongside Gemini 3.5 Live and Live Experimental).
Gemini 3.5 Transcribe is Google's most precise speech-to-text model to date, boasting an exceptionally low 2.6% error rate (AA-WER) and 4.0% for Transcribe Live (streaming at 0.40s latency). Unlike conventional speech recognition models that struggle with disfluencies, background noise, and jargon, Gemini 3.5 Transcribe converts raw, unstructured audio directly into accurate, polished, formatted text.
It automatically edits out filler words ("ums", "ahs") and handles real-time self-corrections (e.g., "let's meet Tuesday—no, Wednesday"), as well as supporting over 85 languages. It is already powering Gboard's Rambler feature and is slated to launch in Chrome.
While Gemini 3.5 Transcribe represents a significant technical leap in real-time voice API capabilities, Google's ability to maintain its position at the absolute frontier of AI reasoning depends heavily on how quickly Kavukcuoglu can resolve the delays surrounding Gemini 3.5 Pro or its rumored successor, Gemini 4.