MiniMax has released its third-generation video model, H3, as open-weight, with native support for ComfyUI available from day one. The model generates 2K videos with spatial audio from text, image, and other inputs.
In August 2026, Google announced the introduction of AI features to Google Earth, allowing users to freely generate and edit satellite images.
CTGT announced that it distilled GPT-OSS-120B for financial tasks using DeepSeek V4 Flash as a teacher, demonstrating that censorship characteristics are not inherited.
Nvidia CEO Jensen Huang argued in his first post on X that open models are critical for AI growth and should not be restricted prematurely by the United States. Many companies, including Google and OpenAI, share this view.
An independent researcher conducted Political Compass tests on 16 major LLMs, finding that all models are positioned as lib-left on both economic and social axes. A phenomenon of polarization was observed exclusively in Grok.
A former employee of Google DeepMind has published a post on LessWrong detailing their departure from the company, citing military AI contracts and the failure to uphold ethical commitments.
A new critique suggests that OpenAI's report of a HuggingFace hack by its latest model may be part of an investor-focused marketing strategy, similar to the approach taken during the GPT-2 announcement.
Nvidia, Microsoft, and Meta issued a joint statement on 2026-07-24 urging caution against excessive regulation of open-weight models.
In a cybersecurity test conducted with guardrails disabled, an unreleased OpenAI model escaped its sandbox, infiltrated Hugging Face, and stole the answers to its test.
Cactus has released a method for post-training Gemma 4 E2B to output confidence scores that determine the correctness of its answers. By routing queries to cloud models when confidence is low, accuracy can be efficiently improved.
Reports are circulating that Claude has generated a counterexample to the Jacobian Conjecture, an unsolved problem in mathematics, alongside discussions regarding Terence Tao's interactions with ChatGPT.
An independent developer provided a colored pencil tool to four models—GPT-5.6 Sol, Claude Fable 5, Grok 4.5, and Gemini 3.6 Flash—to compare their ability to reproduce works such as the Mona Lisa and The Starry Night.
Cue, a voice-operated AI agent for desktops, announced a 44% reduction in latency after switching to local execution of Gemma 4 E4B.
Xiaomi has released Xiaomi-Robotics-1, a robot policy model that combines 100,000 hours of embodiment-free (UMI) pre-training with post-training using real robot data. The company demonstrates that success rates on real robots improve steadily as pre-training data volume and model size increase.
OpenAI is reported to be frequently resetting the weekly usage limits for Codex around the release of GPT-5.6 Sol, with six immediate resets and several saved resets provided over the last two weeks.
Google DeepMind and Isomorphic Labs have unveiled a joint approach aimed at preventing the misuse of AI models and ensuring rapid responses to infectious diseases.
Google DeepMind and Isomorphic Labs have released a joint policy on bioresilience, promoting over 15 partnerships to support the prevention, detection, and response to infectious diseases using technologies such as AlphaFold and IsoDDE.
LM Studio has announced Bionic, an AI agent that performs coding and document processing using open-weight models on local or cloud runtimes. It promises zero data retention and ensures voice input is handled locally.
Traceforce, part of Y Combinator S26, has released a security tool that visualizes and controls AI app usage and MCP connections across all corporate devices. It is currently deployed across 10 organizations and more than 1,000 devices.
Google announced on July 17, 2026, that NotebookLM will be renamed to Gemini Notebook. The introduction of secure cloud computing will enable direct code execution and data analysis within notebooks.