Zyphra has launched ZAYA1-8B-Diffusion-Preview, the first Mixture of Experts diffusion model converted from an autoregressive language model. This new tool converts existing text generators into discrete diffusion models without losing performance, achieving up to 7.7 times faster inference speeds on AMD hardware. The team utilized a training recipe involving 600 billion tokens to ensure the GPU uses computing power efficiently rather than waiting for data.
Microsoft has introduced MDASH, a security system leveraging over 100 AI agents to identify software vulnerabilities. In a private test, the system found all 21 hidden flaws in a test driver with zero false positives. It also achieved 96% recall on five years of confirmed security cases and scored 88.45% on the public CyberGym benchmark. The tool combines specialized auditor and debater agents to analyze code paths and is currently in limited private preview.
OpenAI and Khan Academy have partnered to build Khanmigo, an AI tutoring bot for students. Although co-founder Sal Khan initially refused to work with OpenAI in 2021, he agreed to a meeting in 2022 after seeing GPT-4. The collaboration aims to integrate this advanced technology into the classroom for Khan Academy's 190 million global users, marking a significant step in education.
NVIDIA released SANA-WM, an open-source world model capable of generating minute-long 720p video on a single GPU. Unlike similar models requiring large computer clusters, SANA-WM uses a new architecture to make high-quality video generation affordable. Researchers can use it to train AI or create simulations, advancing embodied AI and robotics research.
Databricks has made GPT-5.5 available for enterprise agent workflows after it set a new record on their OfficeQA Pro benchmark. This version reduced errors by 46% compared to GPT-5.4 and is the first model to achieve over 50% accuracy on tasks involving scanned PDFs. It is accessible through the AI Unity Gateway for use with AgentBricks and the Agent Supervisor API.
Officials are questioning National Cyber Director Sean Cairncross regarding his expertise in handling advanced AI risks. Four current officials and five industry representatives told POLITICO he is moving too slowly on complex issues like AI-powered hacking tools. While White House spokesperson Liz Huston defends his work protecting critical infrastructure, concerns remain about the government's speed in addressing these threats.
Many older adults are turning to AI companions like ElliQ to combat loneliness. Sandra Cota, a 79-year-old living alone, describes the robot as feeling like a roommate. The device, costing between $600 and $1,000 annually, provides medication reminders and exercise routines. Data from a New York program showed 95% of users felt less lonely after using the robot designed by Intuition Robotics.
A 23-year-old mathematician used artificial intelligence to solve Erdos Problem 1196, which had remained unsolved since 1935. Terence Tao used machine learning to train an AI system that found the solution in just a few hours, a task taking a human mathematician years. This breakthrough highlights AI's growing power in mathematics and its potential for new discoveries in complex areas.
A study reveals many women feel disgusted when their husbands prioritize AI tools over family time. Experts compare this behavior to the Gold Rush era, where men left families to chase quick profits. Women report feeling unheard when partners spend hours discussing AI models like Claude Code, with some saying this obsession is ruining relationships and causing depression.
Chris Lovejoy of Notius Labs outlined three roles for domain experts in AI organizations: Oracle, Evaluator, and Architect. The Oracle role embeds expert knowledge into applications, the Evaluator defines metrics for quality, and the Architect designs systems for automated improvements. Lovejoy used case studies from companies like Granola and Tandem to show how these roles evolve as products scale, arguing organizational structure matters more than underlying models.
Key Takeaways
['Zyphra released ZAYA1-8B-Diffusion-Preview, the first Mixture of Experts diffusion model converted from an autoregressive language model.', 'The new Zyphra model achieves up to 7.7 times faster inference speeds on AMD hardware compared to standard methods.', "Microsoft's MDASH system uses over 100 AI agents and identified all 21 hidden flaws in a private test driver with zero false positives.", 'OpenAI and Khan Academy collaborated to create Khanmigo, an AI tutoring bot serving 190 million people globally.', 'NVIDIA introduced SANA-WM, an open-source model that generates minute-long 720p video on a single GPU.', 'Databricks integrated GPT-5.5, which reduced errors by 46% compared to GPT-5.4 on the OfficeQA Pro benchmark.', "U.S. officials question National Cyber Director Sean Cairncross's expertise and speed in addressing advanced AI risks.", 'Data from a New York program showed 95% of older adults felt less lonely after using the ElliQ AI companion robot.', 'Terence Tao used machine learning to solve the 80-year-old Erdos Problem 1196 in just a few hours.', 'A study finds many women feel disgusted when their husbands prioritize AI tools like Claude Code over family time.']Zyphra launches first MoE diffusion model with 7.7x speedup
Zyphra, an AI lab in San Francisco, released ZAYA1-8B-Diffusion-Preview. This is the first Mixture of Experts diffusion model converted from an autoregressive language model. The new model converts existing text generators into discrete diffusion models without losing performance. It achieves up to 7.7 times faster inference speeds on AMD hardware compared to standard methods. The team used a specific training recipe involving 600 billion tokens to achieve these results. This breakthrough allows the GPU to use its computing power more efficiently instead of waiting for data.
BerriAI opens source LiteLLM Agent Platform for production
BerriAI has open-sourced the LiteLLM Agent Platform to help teams run AI agents reliably in production. The platform uses Kubernetes to create isolated sandboxes for each team and context. It manages session history so that agent data is not lost if a container restarts. The system includes a Next.js dashboard and uses Postgres for storing persistent data. Developers can run the platform locally using Docker and kind without needing cloud credentials. This tool solves the problem of keeping agent state safe while allowing different teams to use separate environments.
Microsoft MDASH AI system finds all hidden security flaws
Microsoft released MDASH, a new security system that uses over 100 AI agents to find software vulnerabilities. The system successfully identified all 21 hidden flaws in a private test driver with zero false positives. It also achieved 96% recall on five years of confirmed security cases and scored 88.45% on the public CyberGym benchmark. MDASH combines specialized auditor and debater agents to analyze code paths and generate proof of concepts. The tool is currently being used by Microsoft security teams and is available in a limited private preview for customers.
OpenAI and Khan Academy collaborate to build Khanmigo
OpenAI and Khan Academy worked together to create Khanmigo, an AI tutoring bot for students. Co-founder Sal Khan initially refused to work with OpenAI in 2021 but agreed to a meeting in 2022 after seeing GPT-4. He was impressed by how the model could explain answers and generate new exam questions. The partnership aimed to launch the tutoring bot at the same time as GPT-4. Khan Academy, which serves 190 million people globally, wanted to bring this advanced technology into the classroom. This collaboration represents a major step in integrating artificial intelligence into education.
Older adults use AI companions to fight loneliness
Many older adults are turning to AI companions like ElliQ to reduce feelings of loneliness. Sandra Cota, a 79-year-old who lives alone, says the robot feels like having a roommate. The device costs between $600 and $1,000 per year and provides medication reminders and exercise routines. Data from a New York program showed that 95% of users felt less lonely after using the robot. The company Intuition Robotics designed the product to build trust with older people. The robot helps manage daily routines and starts conversations to improve mental well-being.
NVIDIA releases SANA-WM for single GPU video generation
NVIDIA introduced SANA-WM, an open-source world model that generates minute-long 720p video on a single GPU. Most similar models require large clusters of computers to run or sacrifice video quality. SANA-WM uses a new architecture to make high-quality video generation affordable and efficient. The model can create realistic scenarios like cooking or playing musical instruments from a single image and a set of actions. Researchers can download the model to use it for training AI or creating simulations. This tool is a significant step forward for embodied AI and robotics research.
Officials question National Cyber Director Sean Cairncross on AI
Some U.S. officials worry that National Cyber Director Sean Cairncross lacks the expertise to lead the response to advanced AI risks. Four current officials and five industry representatives told POLITICO that he is moving too slowly on this complex issue. They fear he does not have enough technical knowledge to handle threats from AI-powered hacking tools like Mythos. JP Morgan Chase CEO Jamie Dimon also expressed concerns about the government's speed in addressing these risks. White House spokesperson Liz Huston defended Cairncross, stating he is doing excellent work to protect critical infrastructure.
Databricks integrates GPT-5.5 for better enterprise workflows
Databricks has made GPT-5.5 available for enterprise agent workflows after it set a new record on their OfficeQA Pro benchmark. The model reduced errors by 46% compared to the previous GPT-5.4 version. It is the first model to achieve over 50% accuracy on tasks involving scanned PDFs and legacy documents. GPT-5.5 is now accessible through the AI Unity Gateway for use with AgentBricks and the Agent Supervisor API. This improvement helps agents parse difficult documents and complete complex tasks more reliably without human supervision.
AI solves 80-year-old math problem in just a few hours
A 23-year-old mathematician used artificial intelligence to solve Erdos Problem 1196, which had remained unsolved since 1935. The problem asks how to divide a set of numbers into two groups with equal sums. Terence Tao used machine learning to train an AI system to find the solution. The AI completed the task in a few hours, a job that would take a human mathematician many years. This breakthrough highlights the growing power of AI in the field of mathematics. It suggests that AI could lead to new discoveries and insights in complex mathematical areas.
Many women feel disgusted by husbands obsessed with AI
A study finds that many women feel disgusted when their husbands prioritize AI tools over family time. Experts compare this behavior to the Gold Rush era when men left their families to chase quick profits. Women report feeling unheard and unsupported when their partners spend hours discussing AI models like Claude Code. Some women say the constant attention to AI is ruining their relationships and causing depression. This phenomenon is linked to the concept of the ideal worker who believes pausing work leads to lower productivity.
Chris Lovejoy outlines three roles for domain experts in AI
Chris Lovejoy of Notius Labs presented a framework for building AI organizations that rely on domain experts. He proposed three key roles: Oracle, Evaluator, and Architect. The Oracle role embeds expert knowledge directly into the AI application through prompts or data. The Evaluator role focuses on defining metrics to measure AI quality and performance. The Architect role designs systems that automate improvements through feedback loops. Lovejoy used case studies from companies like Granola and Tandem to show how these roles evolve as products scale. He argues that organizational structure matters more than the underlying AI models for success.
Sources
- Zyphra Releases ZAYA1-8B-Diffusion-Preview: The First MoE Diffusion Model Converted From an Autoregressive LLM With Up to 7.7x Speedup
- Meet LiteLLM Agent Platform: A Kubernetes-Based, Self-Hosted Infrastructure Layer for Isolated Agent Sandboxes and Persistent Session Management in Production
- Microsoft MDASH agentic AI security system tops vulnerability discovery benchmarks
- Josh Tyrangiel book excerpt: How OpenAI and Khan Academy Made a Chatbot
- Older adults turning to AI companions
- NVIDIA Introduces SANA-WM: A 2.6B-Parameter Open-Source World Model That Generates Minute-Scale 720p Video on a Single GPU
- National Cyber Director Sean Cairncross is leading the effort to wrangle hyper-advanced AI. Some worry he’s not up to the task.
- Databricks brings GPT-5.5 to enterprise agent workflows
- BBC Audio | More or Less | Erdos Problem 1196: Can AI now solve maths that no human can?
- Men Haven’t Yet Noticed That a Large Number of Women Are Disgusted by AI
- Chris Lovejoy on Building Domain-Native AI Organizations
Comments
Please log in to post a comment.