Google Updates Gemma 4 with Performance Gains and Bug Fixes
Google released a stealth update for Gemma 4, improving tool calling and response completion, with performance boosts up to 70% on Nvidia Hopper GPUs.
Archive
Browse all published articles.
ARCHIVE
-
-
Thinking Machines Launches Inkling, 975B-Parameter Model Leading U.S. Labs
Thinking Machines, founded by former OpenAI CTO Mira Murati, released Inkling, a 975B-parameter model that leads U.S. open-weights benchmarks but trails top Chinese models.
-
Applied Computing Unveils Orbital AI Model for Oil and Gas Plants
Applied Computing, a London-based startup, raised $20 million in Series A funding to develop Orbital, an AI model designed to help oil and gas operators analyze data from entire plants.
-
ROCm 7.14 Launches TheRock as Production-Ready AI Platform
AMD releases ROCm 7.14, featuring TheRock, its first production-ready automated build system, expanding AI support across MI350 and Ryzen AI hardware.
-
Microsoft Training Sales Team to Downplay OpenAI and Anthropic
Microsoft is reportedly preparing its salespeople to highlight the advantages of its AI models over those of OpenAI and Anthropic, with a focus on efficiency and cost-effectiveness.
-
OpenAI Uses AI to Test Its Own AI, Outperforming Humans
OpenAI's internal AI model GPT-Red found security flaws in 84% of test scenarios, surpassing human red teamers' 13% success rate.
-
SpaceX's Stock Falls to $135 Below IPO Price
SpaceX’s shares dropped to $135 on Wednesday, below the IPO price set for its June 12 public offering, which raised nearly $86 billion.
-
OpenAI Launches $230 Keyboard for Codex
OpenAI introduced a $230 keyboard designed to work with its Codex coding assistant, offering features like light-up agent keys and customizable command shortcuts.
-
SpaceXAI Open-Sources Grok Build Coding Agent
SpaceXAI has open-sourced Grok Build, its coding agent and terminal UI, making the code available on GitHub as of July 15, 2026.
-
GPT-5.6 Sol Pro Disproves 30-Year-Old Statistics Conjecture
A University of Pennsylvania researcher used GPT-5.6 Sol Pro to disprove a long-standing statistics conjecture in 90 minutes, outperforming its predecessor which failed after 20 hours.
-
OpenAI Unveils GPT-Red, LLM Super-Hacker for Cybersecurity
OpenAI released GPT-Red, an LLM designed to simulate cyberattacks, to strengthen its models' defenses. The tool outperformed human red-teamers in testing GPT-5.6.
-
Meta and Stanford Test AI with Baby-Like Learning
Researchers developed EgoBabyVLM, a test that challenges AI models to learn like infants, using footage from babies' cameras. The test reveals current models struggle to interpret messy, real-world data.