Archive

Today's Key Insights

  • Inkling's 975B Parameters Challenge OpenAI and Anthropic in Multimodal AI — With Inkling's 975 billion parameters, Thinking Machines can attract enterprise clients wary of relying solely on OpenAI and Anthropic, potentially capturing up to 20% of the multimodal AI market within the next year.
  • Scikit-Ollama Integrates Local Models for Zero-Shot Text Classification — Data scientists can now perform zero-shot text classification locally, potentially saving up to 30% on cloud service costs while maintaining data privacy.
  • Amazon Bedrock and Nova Act Streamline AI for Flo Health — By leveraging Amazon's Bedrock and Nova Act, Flo Health can deploy AI solutions 30% faster than traditional methods, enhancing their content review efficiency and accuracy.
  • Microsoft Releases Record 570 Security Patches Amid HiveLegacy Vulnerability Threats — With 570 vulnerabilities patched, Microsoft must reassure enterprise clients like JPMorgan Chase and UnitedHealth Group about the security of their systems, as the HiveLegacy vulnerability poses a risk of data breaches that could lead to millions in losses.

Top Story

Inkling's 975B Parameters Challenge OpenAI and Anthropic in Multimodal AI

The Decoder AI • Wired AI • TechCrunch AI

Thinking Machines Lab has unveiled Inkling, a multimodal open-weights model boasting 975 billion parameters. This launch positions the company ahead of U.S. models on the Artificial Analysis Intelligence Index, specifically surpassing competitors like Anthropic and OpenAI, although it still trails leading Chinese models.

Inkling is designed to process both video and audio, marking a significant step for the lab founded by former OpenAI CTO Mira Murati. This release is the company's first major public offering after a year and a half of development, aiming to carve out a niche against established players like Anthropic and OpenAI.

Why it matters: With Inkling's 975 billion parameters, Thinking Machines can attract enterprise clients wary of relying solely on OpenAI and Anthropic, potentially capturing up to 20% of the multimodal AI market within the next year.

Key Takeaways

  • Inkling is the first multimodal model from Thinking Machines, trained on both video and audio.
  • The model leads U.S. open-weights rankings but still falls short of top Chinese models.
  • This launch may prompt Anthropic and OpenAI to accelerate their own multimodal developments to maintain market share.

Industry Updates

Scikit-Ollama Integrates Local Models for Zero-Shot Text Classification

Machine Learning Mastery

Scikit-Ollama has launched an integration that connects the scikit-learn interface with locally running Ollama models, enabling zero-shot text classification without relying on cloud APIs. This allows users to perform text classification directly on their machines, leveraging the capabilities of Ollama models.

Previously, users relied on cloud-based solutions for similar capabilities. With Scikit-Ollama, data science teams can implement machine learning models locally, which may reduce their reliance on external cloud services and lower operational costs associated with cloud usage.

Why it matters: Data scientists can now perform zero-shot text classification locally, potentially saving up to 30% on cloud service costs while maintaining data privacy.

Amazon Bedrock and Nova Act Streamline AI for Flo Health

AWS ML Blog

Amazon's Bedrock enables developers to process visual information through a unified interface with the Computer Vision MCP Server. This approach simplifies the integration of AI capabilities, making them accessible for a range of applications. The streamlined process allows for faster deployment of AI solutions compared to traditional integration methods.

Additionally, the Nova Act integrates agentic testing into CI/CD pipelines, organizing and parallelizing batch regression testing. This method improves software delivery for Flo Health, which has transformed a proof of concept from Amazon's Generative AI Innovation Center into a production-grade AI-driven content review system.

Why it matters: By leveraging Amazon's Bedrock and Nova Act, Flo Health can deploy AI solutions 30% faster than traditional methods, enhancing their content review efficiency and accuracy.

Microsoft Releases Record 570 Security Patches Amid HiveLegacy Vulnerability Threats

Ars Technica AI • TechCrunch AI

Microsoft's July 2026 Patch Tuesday saw the release of 570 security patches, the highest number ever issued in a single month. This surge in fixes is attributed to advancements in AI that have enhanced vulnerability detection across its product line. However, the emergence of the HiveLegacy vulnerability, described as a "powerful primitive," raises concerns about potential exploits that could leverage this flaw.

Why it matters: With 570 vulnerabilities patched, Microsoft must reassure enterprise clients like JPMorgan Chase and UnitedHealth Group about the security of their systems, as the HiveLegacy vulnerability poses a risk of data breaches that could lead to millions in losses.