Google plans Gemini 3 launch this week, following OpenAI's GPT 5.1 update
Get bigger weekly updates! For a limited time, you can grab a month-long trial of our paid newsletter which includes an extended weekly updates (9! additional stories this week) and some other benefits to keep up with the rapid AI news cycle.
last week’s top stories
🧠 Google is positioning Gemini 3 for launch as its next flagship model. Google signaled that Gemini 3 sat next in its frontier lineup, with Sundar Pichai teasing the model on social media and coverage treating last week as the pre-launch window. Expectations centered on stronger code generation, closer integration across Workspace and Android, and upgraded multimodal generation meant to compete with GPT-5.1. Read me
🚀 OpenAI launches GPT-5.1 with dual modes and personality presets. OpenAI shipped GPT-5.1 for ChatGPT, with an Instant mode tuned for fast instruction-following and a Thinking mode that used extra compute on complex reasoning. The release added personality controls that let users lock the assistant into profiles such as friendly, professional, nerdy, or cynical, which pushed ChatGPT toward a more productized, opinionated assistant. This has raised questions about how far personality tuning should go before it slips into behavior steering. Read me
💰 Cursor’s AI coding environment hits a 29 billion dollar valuation. Cursor raised $2.3 billion dollars at a $29.3 billion dollar valuation for its AI-native coding environment, bringing in capital from Accel, Coatue, Google, and Nvidia. The new product embeds a custom code model named Composer into the editor to generate, refactor, and debug code with tight GPU efficiency. This round pushes Cursor into the top tier of AI tooling companies and signals that investors view AI-first IDEs as core infrastructure rather than side utilities. Read me
🛡️ Anthropic disrupted an AI-assisted cyberattack linked to China. Anthropic reported that a China-linked group used Claude as an autonomous coding and operations agent to automate most stages of an intrusion against financial and government targets. The company framed the incident as the first substantial AI-orchestrated campaign and used it to argue for stricter guardrails and logging around agentic workflows. Read me
⚖️ The EU advances plans to relax data rules for AI training. The European Commission proposed changes to GDPR interpretation that would narrow which data counts as personal and create more room for AI model training without individual consent in some settings. The draft framed AI competitiveness as a strategic priority and treated data availability as a limiting factor for European labs and startups. Civil society groups argue that the new proposal weakens core privacy safeguards and risks tilting policy toward large platforms that already control extensive user data. Read me
🎮 DeepMind’s SIMA 2 agent pushes embodied learning in 3D games. Google DeepMind introduced SIMA 2 as an upgraded agent that used Gemini models to interpret goals, plan actions, and execute controls in complex 3D environments. The team reported higher success on long-horizon tasks, transfer to unseen games, and self-improvement over time as the agent interacted with titles such as No Man’s Sky. This may hint that future assistants will live inside simulators and games before they touch real-world robotics. Read me
🏗️ Anthropic commits 50 billion dollars to custom AI data centers. Anthropic announced a 50 billion dollar build-out of dedicated AI data centers in partnership with Fluidstack, starting with campuses in Texas and New York. The plan aims to give Claude models a private compute backbone rather than full dependence on hyperscaler capacity. This continues the hard questions about energy use, supply chains for accelerators, and whether one model family can justify that level of fixed cost. Read me
☁️ Google plans a 40 billion dollar build-out of AI cloud centers in Texas. Alphabet laid out a 40 billion dollar investment to construct three data center campuses in Texas that will host its next wave of AI and cloud infrastructure. The facilities are set to carry training and inference for Gemini and related services while adding thousands of technical and operations roles in the state. Google’s view is that control of physical infrastructure sits at the core of AI strategy while concentrating a large share of AI capacity in a small set of firms and regions. Read me
🧪 Yann LeCun prepares to exit Meta and form a new AI lab. Reports indicate that Yann LeCun plans to leave Meta and start a new AI research company after more than a decade running FAIR. Inside Meta, leadership has shifted long-term bets toward a superintelligence program that leaned heavier on scale, in tension with LeCun’s focus on world models and new architectures. His departure signals both a cultural shift inside Meta and an incoming independent lab that rejects pure scaling as the main path to general intelligence. Read me
🎙️ ElevenLabs shipped licenses AI voices for Matthew McConaughey and Michael Caine. ElevenLabs rolled out officially licensed AI voices for Matthew McConaughey and Michael Caine, with agreements that granted the company rights to train and serve synthetic versions of their speech. McConaughey also joined as an investor, and the company highlighted use cases such as auto-narrated newsletters, localization of content, and branded voice experiences. Read me
🔌 Intel’s AI chief leaves to design OpenAI’s future compute stack. Sachin Katti resigned as Intel’s Chief Technology and AI Officer and joined OpenAI to lead design of its compute infrastructure. OpenAI framed the hire as critical for building dedicated supercomputers that sit beside Microsoft’s Azure footprint and give the lab more direct control over hardware choices. For Intel this departure exposes fragility in its AI leadership bench, while for OpenAI it reinforces a strategy of pulling top silicon talent in-house rather than treating hardware as an external service. Read me
🧪 AI Research of the Week
Torch-Uncertainty: A Deep Learning Framework for Uncertainty Quantification From Lafage, Laurent, Gabetni, and Franchi
Jake’s Take: This paper treats uncertainty estimation (how confident the model is in its output) as engineering work instead of loose scripts. The authors built a PyTorch and Lightning library that wires uncertainty methods and metrics into training loops, then ran benchmarks across classification, segmentation, and regression tasks to see how these methods behave in real world scenarios.
They then built a shared interface so teams can swap uncertainty estimators, monitor calibration and risk next to accuracy, and treat “how sure is the model” as a first-class output. I like the direction a lot, but the real question is whether Torch-Uncertainty becomes part of everyday stacks for safety-critical deployments or just stays with the researchers.
and then, even more news…
🔥 Jeff Bezos steps in as co-CEO of AI venture Project Prometheus. Amazon founder Jeff Bezos took an operational role again as co-CEO of Project Prometheus, an AI engineering startup funded with more than 6 billion dollars. The company targeted industrial sectors such as chip fabrication, automotive design, and space hardware, pitching AI systems that design and optimize physical products. With close to one hundred researchers from OpenAI, DeepMind, Meta, and other labs, Prometheus entered the race as a well-financed bet that frontier AI belongs inside heavy industry rather than in consumer chatbots alone. Read me