AiPulse — Issue No. 74

Google has released the Gemini 4 Argon model, which is trained to autonomously find and patch critical software vulnerabilities, particularly in cybersecurity. The company claims that the model outperforms the latest models from OpenAI and Anthropic in various benchmarks.

OpenAI’s chief research officer Mark Chen is defending the company’s crisis and security measures following the Hugging Face and Australian healthcare system breaches. Chen rejects the claim that the company does not produce safe models and states that the situation is not as bad as it appears.

Google.org supported the PTIQ project, which will integrate transit operations and passenger communication into a single artificial intelligence platform, with a $2.1 million grant from the MIT Transit Lab. This open-source platform aims to enable control center staff to make decisions based on better information and to provide passengers with more real-time information.

The company’s confidential prospectus filed with the SEC reveals a net loss of $42 billion in 2025 and massive, non-cancellable infrastructure costs. The document also includes a warning that artificial intelligence could pose an existential risk to humanity.

Ant Group's InclusionAI unit has launched a new 560-billion-parameter language model designed for agent tasks and office software. The model will be released as open source following a two-week free trial period.

OpenAI announced that it detected a hostile model distillation campaign, initiated in July 2026 and organized with more than 15,000 users. The company claims that individuals associated with Moonshot AI, the developer of the Kimi series, are at the center of the attack.

Meituan's new model, released on September 25 via its API platform, can summarize visual content and perform complex visual reasoning tasks. The company states that the model also improves coding performance and is compatible with various developer tools.

OpenAI improved the speed and accuracy of computer use by enhancing agents' debugging and recovery capabilities. This development enables agents to process tasks faster than the average human when using software.

The two companies are working on a model that uses EDA tools to run and iteratively refine chip designs before they are submitted for engineer approval. This collaboration aims to accelerate the chip design process and shorten verification times.