Noam Shazeer Leaves Google for OpenAI as AI Architecture Lead
Gemini co-lead and Google VP engineering Noam Shazeer is joining OpenAI to lead AI architecture research.
Shazeer Moves to OpenAI
Noam Shazeer, vice-president of engineering at Google and a Gemini AI model co-lead, will leave Google to join OpenAI. At OpenAI, Shazeer will act as the organisation’s lead for AI architecture research.
A History of Foundational Work
Shazeer co-authored a pioneering paper on AI in 2017, exploring the transformer architecture that defines the generative AI of today.
The move comes after Shazeer left Google in 2021 to found Character.AI. He returned to Google in 2024 with a team of researchers in a deal reportedly valued at roughly $2.7 billion.
Source: Silicon Republic
Developments since publication
-
Noam Shazeer, co-author of the 2017 'Attention Is All You Need' paper that introduced the Transformer architecture, co-authored the Sparsely-Gated Mixture of Experts paper in 2016, and designed Multi- Source
-
Noam Shazeer announced his departure from Google on June 18, 2026 and is joining OpenAI as Lead for Architecture Research. Source
-
Google paid approximately $2.7 billion in August 2024 to bring Noam Shazeer back from Character.AI in a licensing arrangement that lasted less than 22 months. Source
-
The per-month cost of Google's $2.7 billion arrangement with Shazeer, viewed from the outcome of his departure after 22 months, is approximately $122 million per month. Source
-
Claude Fable 5 achieved a score of 70% PASS@1 on the DeepSWE benchmark, confirmed by Datacurve on approximately June 19, 2026, three percentage points ahead of GPT-5.5. Source
-
Amazon CEO Andy Jassy personally called the White House, which instigated the order for Anthropic to suspend Fable 5 and Mythos 5, according to The Wall Street Journal as cited by TechPolicy.Press. Source
-
Anthropic updated its privacy policy effective July 8, 2026 to include collection of government-issued ID and biometric data including facial recognition. Source
-
China announced a national AI infrastructure investment plan targeting 2 trillion yuan—approximately $295 billion—over five years for interconnected national AI data centers. Source
-
Google made Gemini 3.5 Flash the default model across all Gemini consumer and enterprise products on June 18, 2026, retiring Gemini 2.5 Flash from the default position. Source
-
The Meituan LongCat team released General 365, a benchmark for evaluating AI reasoning capabilities across 26 mainstream models, with Gemini 3 Pro achieving only 62.8% accuracy. Source
-
The vast majority of the 26 models tested on the General 365 benchmark failed to reach 60%, which is traditionally considered a passing grade. Source
-
Meituan LongCat team introduced WBench, a systematic multi-round evaluation benchmark for interactive video world models, designed to evaluate models across diverse scenarios including lunar environme Source
-
Meituan's technical team introduced a groundbreaking approach to managing AI-assisted development, focusing on the refactoring of 310,000 lines of code using agent evaluation methods. Source
-
Meituan released LongCat-Video-Avatar 1.5, which transitions digital human technology from experimental state-of-the-art models to robust, commercial-grade applications with support for multi-person i Source
-
Google Research introduced TimesFM (Time Series Foundation Model), a pretrained foundation model specifically engineered for time series forecasting. Source
-
OpenMontage has launched as a pioneering open-source project, consisting of 12 specialized pipelines, 52 integrated tools, and a library of over 500 distinct agent skills. Source
-
Meituan LongCat team released LARYBench (Latent Action Representation Yielding Benchmark), showing that general-purpose vision models significantly outperform specialized embodied expert models in act Source
-
Codebase-memory-mcp, an MCP server developed by DeusData, achieves millisecond-level processing per repository and sub-millisecond query speeds while supporting 158 programming languages. Source
-
Samsung and SK Hynix profit forecasts are surging due to a persistent global memory shortage, with suppliers strategically shifting production focus toward server DRAM. Source
-
Apertus released Apertus Mini, a collection of 16 small language models designed to advance Sovereign AI through distillation and quantization techniques. Source