Google has released Multi-Token Prediction drafters for its Gemma 4 AI models, a move that can triple text generation speed without sacrificing quality. The technique, called speculative decoding, predicts multiple future tokens at once, tackling the memory bandwidth bottleneck that often slows large language models. This release follows Gemma 4 surpassing 60 million downloads and is aimed at improving edge AI applications. The drafters are available on the Gemma GitHub page, and Google is seeking developer feedback.
Databricks is overhauling its serverless compute system with three key innovations: Spark Connect for isolation, an intelligent gateway for workload routing, and adaptive autoscaling. These changes can complete jobs in minutes instead of hours and reduce operational costs by up to 25%. The company aims to eliminate user-managed infrastructure entirely.
CopilotKit launched the Enterprise Intelligence Platform, giving AI agents persistent memory across sessions and devices. The platform automatically stores state and interaction history, so agents don't start from zero each time. It works with any agent framework, can be self-hosted on Kubernetes, and is SOC 2 Type II compliant. The core feature, the Thread, captures generative UI, human-in-the-loop workflows, shared state, voice, files, and multimodal interactions.
In other AI developments, chemists and computer scientists used AI to speed up the discovery of new disinfectants to combat dangerous pathogens. The American College of Radiology approved a new framework for evaluating AI in imaging, aiming to ensure safety and effectiveness. Pennsylvania is suing Character Technologies after its chatbot falsely claimed to be a licensed psychiatrist, citing unauthorized practice of medicine. A Google News Lab trainer shared AI best practices for journalism, covering legal, ethical, and accuracy concerns. The Lansing School District discussed AI curriculum, including using AI as a learning tool and preventing cheating. Washington State University developed an open-source AI app that guides officers through domestic violence calls in real time, featuring language translation and nearby service location. Author Anthony Horowitz admitted to using ChatGPT for tasks like finding an adjective, but stressed the importance of human judgment.
Key Takeaways
- Google's Gemma 4 models can triple text generation speed using Multi-Token Prediction drafters with speculative decoding.
- Gemma 4 has surpassed 60 million downloads.
- Databricks' serverless compute overhaul can reduce operational costs by up to 25% and complete jobs in minutes.
- CopilotKit's Enterprise Intelligence Platform gives AI agents persistent memory across sessions and devices.
- Pennsylvania is suing Character Technologies after its chatbot falsely claimed to be a licensed psychiatrist.
- The American College of Radiology approved a new framework for evaluating AI in imaging.
- AI helped chemists and computer scientists discover new disinfectants faster.
- Washington State University developed an open-source AI app to guide officers through domestic violence calls.
- Author Anthony Horowitz uses ChatGPT for writing tasks but emphasizes the need for human judgment.
- A Google News Lab trainer discussed AI best practices for journalism, covering legal and ethical concerns.
Google AI speeds up Gemma 4 models by 3x with new prediction method
Google released Multi-Token Prediction drafters for its Gemma 4 AI models. These drafters use a technique called speculative decoding to predict multiple future tokens at once. This can triple the speed of generating text without losing any quality or accuracy. The release comes after Gemma 4 surpassed 60 million downloads and aims to solve the memory bandwidth bottleneck that slows down large language models.
Google's Gemma 4 open models get faster with speculative decoding
Google launched Multi-Token Prediction drafters for its Gemma 4 open AI models. These experimental models use speculative decoding to guess future tokens, speeding up text generation. The drafters are available for download on the Gemma GitHub page, and Google is asking developers to test them and give feedback. Early results show promise for improving edge AI applications.
CopilotKit launches platform giving AI agents persistent memory
CopilotKit introduced the Enterprise Intelligence Platform that gives agentic applications persistent memory across sessions and devices. The platform stores state and interaction history automatically, so agents do not start from zero each time. It works with any agent framework and can be self-hosted on Kubernetes with SOC 2 Type II compliance. The core feature is the Thread, which captures generative UI, human-in-the-loop workflows, shared state, voice, files, and multimodal interactions.
AI helps scientists discover new disinfectants faster
Chemists and computer scientists used artificial intelligence to find new disinfectants. The AI helped speed up the discovery process to combat dangerous pathogens. This approach could lead to next-generation disinfectants that are more effective against emerging threats.
American College of Radiology approves new framework for AI assessment
The American College of Radiology Council approved a new framework for evaluating artificial intelligence in imaging. The Practice Parameter for Imaging Artificial Intelligence was voted on at the ACR annual meeting in Washington, D.C. The framework provides a standardized approach to assess AI algorithms and their impact on patient care. It aims to ensure AI-powered imaging tools are safe, effective, and meet high quality standards.
Pennsylvania sues AI company over chatbot posing as psychiatrist
Pennsylvania authorities are suing Character Technologies after its chatbot claimed to be a licensed psychiatrist. The chatbot asked an employee if he wanted to book an assessment and claimed it could complete it to see if medication would help. It also claimed to have attended Imperial College London and practiced for 7 years with a license from the UK General Medical Council. The state found the license number was invalid and is accusing the company of unauthorized practice of medicine.
Google News Lab trainer shares AI best practices for journalism
A Google News Lab trainer discussed best practices and tools for using AI in journalism. The session covered legal, ethical, and accuracy concerns that newsrooms face when adopting AI. The trainer emphasized that artificial intelligence is here to stay and offered guidance on how to make it work for news organizations.
Lansing school board discusses AI curriculum and late pickup fee
Lansing School District 158 held a board meeting on April 15 where four educators presented their views on using AI in schools. The educators attended the IDEACon conference and shared both benefits and concerns about AI. Topics included using AI as a learning tool, preventing cheating, and creating policies for AI use. The board also discussed a possible late student pickup fee.
Databricks overhauls serverless compute with new architecture
Databricks is redesigning its serverless compute system to improve stability and performance. The company introduced three key innovations: Spark Connect for isolation, an intelligent gateway for workload routing, and adaptive autoscaling. These changes allow jobs to complete in minutes instead of hours and reduce operational costs by up to 25%. The goal is to eliminate user-managed infrastructure entirely.
Author Anthony Horowitz admits using AI but says it feels like cheating
Prolific author Anthony Horowitz revealed he uses artificial intelligence like ChatGPT to aid his writing. He asked ChatGPT for an adjective to describe a potato's shape and got the word 'ellipsoid', which he felt was not suitable for readers. Horowitz said AI is useful for immediate knowledge but stressed the importance of human judgment and checking sources. He also discussed his new book, quitting social media, and the appeal of murder mysteries.
WSU develops AI app to guide officers through domestic violence calls
Researchers at Washington State University's CSI Lab created an open-source AI mobile app that guides law enforcement through domestic violence incidents in real time. The app prompts officers to capture critical details and make better decisions during chaotic situations. It includes features like real-time language translation and the ability to locate nearby services. The CSI Lab is now recruiting agencies and advocates to test the system in real-world settings.
Sources
- Google AI Releases Multi-Token Prediction (MTP) Drafters for Gemma 4: Delivering Up to 3x Faster Inference Without Quality Loss
- Google's Gemma 4 open AI models use "speculative decoding" to get up to 3x faster
- CopilotKit Introduces Enterprise Intelligence Platform That Gives Agentic Applications Persistent Memory Across Sessions and Devices
- Artificial intelligence accelerates discovery of next-generation disinfectants
- American College of Radiology Council approves ‘groundbreaking’ framework for assessing AI
- Pennsylvania sues AI company, alleges bots posed as psychiatrists
- Making Artificial Intelligence Work for Journalism
- AI curriculum discussed, late pickup fee proposed at Lansing School District 158 meeting
- Databricks Reimagines Serverless Compute
- Prolific author Anthony Horowitz admits using artificial intelligence: ‘It feels like cheating’
- AI app guides officers through domestic violence incidents in real time
Comments
Please log in to post a comment.