PLUS: Capital One's AI bug hunter, Nvidia's new AI math, and an AI that found a crypto flaw
Happy reading
The global AI race just got a major new competitor from China. Moonshot AI has unveiled Kimi K3, a massive 2.8 trillion-parameter model that rivals the performance of top American systems at a fraction of the cost.
With its upcoming release as the world's largest open-weight model, Kimi K3 isn't just a competitor—it's a new foundation for developers everywhere. Does this signal an end to America's commanding lead in AI and a major shift in the global innovation landscape?
In today’s Next in AI:
China's massive Kimi K3 model
Capital One's open-source AI bug hunter
Nvidia's 'intelligence per dollar' metric
An AI auditor that found a major crypto flaw
China's AI Moonshot
Next in AI: Chinese startup Moonshot AI just dropped Kimi K3, a massive 2.8 trillion-parameter model that rivals top US systems. The model is shaking up the global AI landscape and will be released as an open-weight model later this month.
Explained:
Kimi K3 is immediately competitive, outperforming models like Anthropic's Fable 5 in front-end coding tests while costing 40% less than comparable premium US models.
Set for a July 27 release, it will be the world's largest open-weight model, allowing any developer or company to download, customize, and run it on their own systems.
This release effectively erases America's once-commanding lead in AI, suggesting China has closed a technology gap previously estimated to be 6-12 months in a fraction of the time.
Why It Matters:
Kimi K3's arrival puts immense pressure on the pricing and perceived technological edge of top American AI labs. This signals an acceleration in the global AI race, where powerful, open-weight models from China are now a major competitive force for developers and businesses.
Capital One's AI Bug Hunter
Next in AI: Capital One has open-sourced VulnHunter, an agentic AI security tool that autonomously finds, validates, and suggests fixes for code vulnerabilities by mimicking an attacker's mindset.
Explained:
The tool flips traditional security scanning by using an attacker-first approach, starting at potential entry points like APIs and reasoning forward to find viable exploit paths.
A built-in falsification engine challenges the AI's own conclusions, a unique step designed to dramatically reduce the false positives that often overwhelm development teams.
Releasing this as a public defensive resource democratizes advanced security, allowing the wider community to help find and fix vulnerabilities before they are exploited.
Why It Matters:
This release signals a major shift toward using AI for proactive defense, not just reactive scanning. By open-sourcing VulnHunter, Capital One provides a powerful resource for collective security in an era of AI-accelerated threats.
Nvidia's New AI Math
Next in AI: Nvidia is reframing the economics of AI, shifting the focus from just the cost per token for inference to a new metric: 'intelligence per dollar.' This new approach highlights the growing importance of continuously improving a model through post-training.
Explained:
Post-training is where models learn to perform complex tasks like writing code or planning multi-step actions, primarily through a process called reinforcement learning.
This new metric doesn't replace cost per token but builds upon it; a more capable model makes every token it serves more valuable, increasing the return on the initial investment.
Nvidia’s new Nemotron 3 Ultra model, which can fix real software bugs with over 71% accuracy, showcases this focus and is built to run on the upcoming Vera Rubin platform designed to maximize this new metric.
Why It Matters: This signals that the AI industry is maturing beyond simply scaling models to building and continuously refining specialized agents. The 'intelligence per dollar' framework will guide future investments in AI infrastructure, prioritizing systems that support ongoing learning and improvement.
AI Finds Crypto's Flaw
Next in AI: An AI-powered auditor zkao discovered a major security flaw in OpenVM's zero-knowledge virtual machine, a foundational technology for many web3 protocols.
Explained:
The tool found a critical soundness bug in a core pairing library that could allow a malicious actor to forge cryptographic proofs. This vulnerability effectively breaks the security model for protocols relying on KZG opening proofs or Groth16 SNARKs.
The discovery highlights the value of specialized AI, as zkao succeeded where more general-purpose LLMs failed to find the exploitable issue. The AI found the bug after more than nine hours of scanning the complex codebase.
In response, the OpenVM team quickly acknowledged the finding and has already patched the vulnerability in version 1.6.0, demonstrating a successful cycle of AI-assisted discovery and human-led resolution.
Why It Matters: This event showcases AI's expanding role from a productivity tool to a specialized auditor capable of securing incredibly complex cryptographic systems. As digital infrastructure grows more intricate, AI-powered security analysis is becoming an essential layer of defense.
AI Pulse
Anthropic is in preliminary talks to lease a significant amount of AI computing power from Meta, with a potential deal reportedly valued at around $10 billion.
Wired revealed Meta’s AI app for its smart glasses quietly embeds face-recognition technology capable of creating biometric data, a feature internally called “NameTag.”
Nurses report that AI-driven workplace surveillance at Kaiser Permanente is threatening patient care, with a new survey finding a majority of nurses believe AI undermines safety.
Viasat won a $229 million patent infringement case against Japanese chipmaker Kioxia after a federal jury found Kioxia's flash memory technology violated its patents.