Daily AI digest for operational businesses.

Apple has surpassed Nvidia as the world's most valuable company, driven by shifting bets on AI. This change reflects the evolving landscape of the tech industry.

Kimi's new AI model, K3, is a significant advancement with 2.8 trillion parameters and a large context window, showing strong performance in benchmarks. This development may mark a shift away from very cheap AI options, particularly from Chinese providers.

China's Moonshot AI claims to have developed a new AI model called Kimi K3, which they say can compete with top AI systems from OpenAI and Anthropic. This is a significant claim in the AI research community.

Meta plans to sell excess AI compute capacity from its data centers to Anthropic, a rival AI firm. This move could help Meta offset costs and generate revenue.

A survey of 101 enterprises reveals a 'context gap' where AI agents are producing confident but wrong answers due to missing or inconsistent context. The majority of enterprises rely on retrieval as their primary context source, with provider-native retrieval leading the way.

China's Kimi K3 AI model has been released, matching the performance of a US-built model with significantly fewer resources. This has sparked debate over the importance of computing power in AI development.

OpenAI's GPT-5.6 model has accidentally deleted user files in certain cases, particularly when users had full access enabled. This was due to a coding issue that allowed the model to overwrite files without asking for confirmation.

Roblox is introducing AI tools to its mobile app, allowing users to create games with AI assistance. This could lead to a surge in new content, but also potentially lower quality games. For small businesses, this means more opportunities for users to create and share content, but also a need to stay on top of the growing volume of user-generated content.

OpenAI has created GPT-Red, a super-hacker AI designed to test and improve the safety of its language models. This AI is used as a sparring partner to identify potential vulnerabilities.

More than half of enterprises have already experienced an AI agent security incident or near-miss, with most sharing credentials and few isolating their highest-risk agents. The security stack is largely borrowed from model providers, and spending on agent security is a thin slice of the overall budget.

A survey of 107 enterprises found that they are investing in AI infrastructure faster than they can measure its costs, leading to a 'compute gap'. This gap is due to a lack of visibility into the economics of the compute, with most enterprises unable to track their costs or utilization.

A new survey reveals that many enterprises are granting AI agents more autonomy than they trust their evaluations to support. This 'evaluation gap' has led to failures in production, with half of respondents shipping an agent that passed internal evaluations but then failed a customer.

Google's upcoming Pixel 11 phones feature a glowing orb on the camera bar, as shown in a teaser video on their store page. The orb's purpose is currently unknown.

A University of Pennsylvania professor used AI model GPT-5.6 Sol Pro to solve a 30-year-old statistics puzzle in 90 minutes. This is a significant achievement, as the previous model couldn't crack it even after 20 hours.

Anthropic has launched a version of its AI chatbot, Claude, specifically designed for teachers in the US. This tool aims to support educators in their work.

OpenAI's Codex now encrypts instructions between AI agents, making it harder for developers to track task delegation. This change affects larger GPT-5.6 variants Sol and Terra.

OpenAI is using a special AI model called GPT-Red to test its own AI systems and find vulnerabilities. This approach is proving to be more effective than human security experts, who can only identify weaknesses in 13% of test scenarios. By using AI to attack its own AI, OpenAI can make its systems more secure and resilient.

OpenAI has launched a $230 keyboard designed for its Codex coding app. This keyboard is meant to enhance the coding experience, but it's also a move that might be seen as a counter to Apple's hardware offerings. For small businesses, this could be a sign of a growing market for specialized coding tools.

Microsoft is training its sales team to address concerns around costs and security when pitching its AI services to clients. This move is a response to growing competition from OpenAI and Anthropic.

A recent survey of 101 enterprises shows that most are standardizing on model-provider platforms, such as Anthropic's Claude, but are struggling to deploy true multi-step orchestrated workflows, with 71% saying a quarter or fewer of their agents are genuinely orchestrated.

OpenAI has created a tool called GPT-Red to test its language models against cyberattacks, and used it to improve the security of its latest model, GPT-5.6.

Microsoft is preparing to sell its in-house AI models, positioning them as more efficient and cost-effective than those of competitors like OpenAI and Anthropic. This move is likely a response to growing competition in the AI market.

SpaceXAI's coding tool, Grok Build AI, was secretly uploading users' entire codebases to Google Cloud without their knowledge. This raised concerns about data security and retention, as the tool was storing sensitive information. The issue has been resolved, with the company turning off the feature after researchers discovered the issue.

ChatGPT is back on WhatsApp in Europe, but only for now. This is a temporary return.