Company / organization
By and about Google
More Random
Two neural network features that appear identical by standard metrics can have vastly different causal effects on model behavior, with downstream connections predicting influence better than activation patterns.
newsGoogle: Gemini 3.5 Flash
Google launched Gemini 3.5 Flash as the default model in Gemini app and Search AI Mode, scoring 76.2% on Terminal-Bench 2.1 and 1656 Elo on GDPval-AA while running ~4x faster than comparable frontier models.
newsGoogle: Gemma 4 12B and DiffusionGemma
Google released Gemma 4 (12B to 31B, Apache 2.0, natively multimodal, 256K context) for local deployment, and followed with DiffusionGemma, a 26B diffusion model generating 1,000+ tokens/sec on a single H100 by denoising in parallel instead of…
newsDecoupled DiLoCo
Google DeepMind published Decoupled DiLoCo, a distributed training method that dramatically reduces inter-node bandwidth, enabling model training across data centers and potentially globally.
newsDwarkesh Blackboard Lectures
Dwarkesh Patel's blackboard lecture series features top AI researchers explaining how frontier LLMs are trained and served, reinforcement learning from AlphaGo, and chip design fundamentals.
newsGoogle launches Omni
Sources: tweethttps://x.com/samsheffer/status/2056820022144905380
newsGoogle Releases Gemma 4 Open Models
Google DeepMind released Gemma 4, an open-source model family under Apache 2.0 with 26B MoE and 31B variants that outperform models 10x their size on reasoning benchmarks while optimized for on-device deployment.
newsGoogle Simula: Reasoning-Driven Synthetic Data
Google DeepMind and Research introduce Simula, a framework for generating synthetic training data using AI reasoning to address data scarcity and privacy concerns in multi-modal AI development.
newsAgentic AI & the Next Intelligence Explosion
Google researchers argue that AI progress depends less on scaling single models than on structured interactions—debate, verification, and synthesis—between multiple agents.
newsThe Abstraction Fallacy: Can AI Be Conscious?
An argument that computational functionalism commits the "Abstraction Fallacy" by assuming consciousness emerges from abstract computation alone, when physical substrate may be essential to generating rather than merely simulating experience.
newsMore Random: Design.md, AI Writing Tells, Palantir, Pope & More
Google releases Design.md to let designers export and import design rules across projects, while AI writing tells emerge as a recognizable pattern and optimism about AI's future wanes amid concerns about job displacement.
newsGoogle TurboQuant: 6x KV-Cache Compression with Zero Accuracy Loss
Google releases TurboQuant, a compression algorithm that reduces LLM key-value cache memory by at least 6x with up to 8x speedup and zero accuracy loss.
newsGemini Embedding 2: Natively Multimodal Embedding Model
Google released Gemini Embedding 2, a multimodal embedding model that maps text, images, video, audio and documents into a single embedding space for cross-media retrieval and classification.
newsThinking to Recall: How Reasoning Unlocks Parametric Knowledge in LLMs
Google Research finds that enabling reasoning in LLMs helps them recall simple facts through two mechanisms: extra tokens provide computational buffer time, and intermediate facts prime the correct answer—but hallucinated intermediate facts sharply…
newsGoogle Gemini 3.1 Flash-Lite
Google released Gemini 3.1 Flash-Lite, a distilled lightweight model designed for speed and cost-efficiency in tasks like PDF extraction, OCR, and summarization.
newsOpenAI, xAI, and Google Sign DoW Defense Agreements
OpenAI, xAI, and Google have signed defense agreements with the U.S. Department of War, with OpenAI committing to a $120M contract while xAI lifted AI safety restrictions for classified work.
newsGoogle Gemini 3.1
Google released Gemini 3.1 with 1M+ token context, stronger reasoning performance on ARC-AGI benchmarks, and native support for text, image, audio, video, and code across its app, Workspace, and Vertex AI platform.
newsApple — Think Different on AI Capex
Apple is skipping the $100B+ data center arms race, instead deploying its $150B cash pile on inference through Google's Gemini while rivals race to build their own AI infrastructure.
newsGenomic Models: Bytedance SeedFold and Google AlphaGenome
ByteDance's SeedFold outperforms AlphaFold 3 in protein structure prediction, while Google releases AlphaGenome weights to academic researchers.
newsWorld Models Are Taking Over
Google's Project Genie, World Labs' World API, and DeepMind's D4RT shift AI from generating videos to building persistent simulated worlds where agents can reason and act.