AI News Roundup for April 7-14, 2025

This is a round of news from the last week shared internally between April 7-14, 2025.

Table of Contents

• 🗞️ Product News and Updates

• 🧠 Model Updates

• 🔬 Interesting Research

• 🎓 Training and Learning Resources

• 💭 Opinion and Reflection

• 🧑🏻💻 Building with AI

🗞️ Product News and Updates

ChatGPT's New Long Term Memory

11 Apr 2025 - OpenAI announced a new memory feature for ChatGPT that extends the existing functionality to reference all previous chats, not just snippets.

• 🔒 Limited availability: Only for Plus and Pro users in the U.S., excluding EU and UK presumably for regulatory reasons

• 🔄 Default settings: Automatically enabled if you had the previous Memory feature on, disabled otherwise

• 🌐 Geographic restrictions: As with other geographically limited ChatGPT features, using a VPN will give access

WordPress Launches Prompt-to-Website Builder

10 Apr 2025 - WordPress launched a new AI website builder that transforms web development by allowing users to create sites through conversation.

• 💬 Conversation-based creation: Users can speak websites into existence with a single prompt

• 🎨 Interactive editing: Continue telling the AI to change colors or generate new layout ideas

• 🆓 Free tier: First 30 prompts free (it took the author 4 prompts to get something usable)

• 🏗️ Complete structure: Builds out content structure, pulls in images, and adds nice touches like social media links

🔗 More on WordPress.com

Google's All-in on Agents: ADK and A2A Protocol

10 Apr 2025 - At Cloud Next conference, Google announced major advancements in AI agent technology, building on and contributing to existing frameworks.

• 🧩 Agent Development Kit (ADK): Open-source framework with pre-built bots, tools, and over 100 connectors to common data sources

• 🤝 Agent-to-Agent (A2A) Protocol: Facilitates secure communication between AI agents built on different frameworks

• 🔄 MCP Integration: A2A complements Anthropic's Model Context Protocol, with A2A focusing on networking and collaboration between agents

• 🏢 Industry support: Over 50 partners, including major tech companies and service providers, are contributing to its development

🔗 More on Google Developers Blog

GPT-4o Image Generation Now Available on ChatGPT Edu

11 Apr 2025 - ChatGPT Edu now has access to the new GPT-4o image generation functionality announced two weeks ago.

• 🔄 Object persistence: Enables editing by text across multiple prompts

• 📝 Text rendering: Handles large amounts of accurate text in images

• 🎨 Style imitation: Ability to add existing pictures into generations and imitate specific styles

• ⚠️ Usage note: Users may need to refresh their screen to see the "Create image (Updated)" option instead of DALL-E

🔗 More on Teams

Anthropic Education Report on Student Claude Usage

11 Apr 2025 - Anthropic published a report analyzing one million anonymized student conversations on Claude.ai to understand real-world AI usage in higher education.

• 👨💻 STEM dominance: Computer Science students make up 36.8% of conversations but only 5.4% of U.S. degrees

• 🧠 Four interaction patterns: Direct Problem Solving, Direct Output Creation, Collaborative Problem Solving, and Collaborative Output Creation

• 📊 Primary uses: Creating new content (39.3%) and technical problem-solving (33.5%) across disciplines

• 🔍 Academic integrity concerns: Approximately 47% of conversations are Direct (minimal engagement) interactions that could enable cheating

• 🏫 Learning Mode: Anthropic is partnering with universities to develop a mode emphasizing Socratic methods and conceptual understanding

🔗 More on Anthropic

➡️➡️➡️

🧠 Model Updates

Llama 4 Models Released

07 Apr 2025 - Meta announced two new models, Llama 4 Scout and Llama 4 Maverick, as part of the Llama 4 series, with a third model (Behemoth) still in training.

• 🔢 Architectural innovation: Uses Mixture of Experts (MoE) where only a subset of parameters are active at any time

• 📏 Extensive context: 256k context window natively, with Scout extended to 10 million tokens

• 📊 Training scale: About 30 trillion tokens used, double the 15 trillion for Llama 3

• 🌐 Multimodal capabilities: Models understand images and are multilingual, pre-trained on 200 languages with official support for 12

• 🧪 Strong benchmarks: Beat all equivalent size models, promising to be competitive with Claude and GPT-4o

🔗 More on Meta AI

➡️➡️➡️

🔬 Interesting Research

Stanford's AI Index Report 2025

14 Apr 2025 - Stanford released their annual AI Index Report, highlighting key trends while containing some flaws and outdated information.

• 📈 Growing influence: AI's impact across society, economy, and global governance continues to intensify rapidly

• 💻 Performance leap: AI performance on demanding benchmarks is sharply increasing

• 🏢 Business adoption: AI investment surged in 2024, with generative AI attracting significant funding

• 🌐 Global competition: U.S. leads in top AI models, but China is rapidly closing the performance gap

• 🛡️ Evolving governance: Responsible AI ecosystem developing unevenly; incidents rising while governments increase regulatory efforts

🔗 More on Stanford HAI and LinkedIn critique

➡️➡️➡️

🎓 Training and Learning Resources

How GPT-4.5 Was Trained

14 Apr 2025 - OpenAI shared a video featuring Sam Altman and three OpenAI engineers discussing the training process of GPT-4.5, offering general principles and challenges encountered.

• 📊 Massive endeavor: Building frontier models requires immense resources and intricate planning across ML, Systems, and Data teams

• 🔄 Inherent unpredictability: Large training runs inevitably encounter issues requiring real-time problem-solving

• 💻 Scaling challenges: Increasing compute (e.g., 10x GPUs) transforms rare failures into frequent, potentially catastrophic events

• 🧩 Pre-training power: The next-token prediction objective creates general intelligence by forcing data compression, finding connections and abstractions

• 📈 Future bottlenecks: While compute scaling continues, high-quality data is becoming a more significant constraint, requiring algorithmic innovations

🔗 More on YouTube

The King's Festival of Artificial Intelligence

08 Apr 2025 - The King's Festival of Artificial Intelligence returns to King's College London from May 20-24, 2025, exploring the latest developments in AI.

• 🏫 Broad topics: Education, healthcare, creative industries, sustainability, and business

• 👨🏫 Expert lineup: Over 100 colleagues from King's nine faculties showcasing research through public lectures, panel discussions, and workshops

• 🌐 Multiple venues: Events taking place at the Strand Campus, Guy's Campus, London Institute for Healthcare Engineering, Science Gallery London, and online

• 🆓 Free access: All festival events are free and open to the public

🔗 More on King's College London

➡️➡️➡️

💭 Opinion and Reflection

AI Fluency as a Job Requirement: Karpathy & Shopify

07 Apr 2025 - Andrej Karpathy released an essay on individual benefits of LLMs while Shopify's CEO announced mandatory AI usage for employees, highlighting the disparity in AI benefits between individuals and organizations.

• 🔄 Bottom-up revolution: ChatGPT is unique in being adopted from the bottom up rather than top down, with broadly equal access

• 🏢 Organizational challenges: "Organizations deal with problems of a lot greater complexity and necessary coordination" (Karpathy)

• 💼 Shopify's mandate: "Reflexive AI usage is now a baseline expectation at Shopify" with "AI usage questions in performance and peer review"

• 👨💻 Hidden cyborgs: Many are already using AI without "telling anyone," echoing Ethan Mollick's 2023 observations

• 📚 Skill gap paradox: "Learning to use AI well is an unobvious skill" despite the seemingly simple chat interface

🔗 More on Twitter/X and Readwise

Byrne Hobart on AI Replacing Professional Services Staff

08 Apr 2025 - Byrne Hobart argues that AI, particularly LLMs, may be better suited for senior-level tasks like project decomposition and code review than traditional junior roles.

• 🔄 Counter-intuitive replacement: "What if AI actually replaces a lot of senior people?" rather than just entry-level work

• 🧩 Task decomposition: "The job of the senior person is basically to look at large projects and break them down into discrete tasks" - a capability LLMs excel at

• 👀 Code review: LLMs can "review the code" and suggest architectural improvements before new features are built

• 👨👩👧👦 Human limitations: LLMs are "a very bad substitute for mentorship" and lack human empathy and motivational skills

• 📊 Overall potential: At "70% as good" as human seniors, LLMs could be valuable where ideal human resources aren't available

🔗 More on YouTube 💬 Discuss on Teams

🧑🏻💻 Building with AI

DeepCoder: Open-Source 14B Coder Model

11 Apr 2025 - DeepCoder-14B-Preview, a code reasoning model developed by the Agentica team and Together AI, was released with impressive performance benchmarks.

• 🧠 Architecture: Finetuned from Deepseek-R1-Distilled-Qwen-14B via distributed RL

• 📊 Performance: 60.6% Pass@1 accuracy on LiveCodeBench (+8% improvement)

• 🔍 Competitive edge: Matches performance of o3-mini-2025-01-031 (Low) and o1-2024-12-17 with just 14B parameters

• 🌐 Open source commitment: Dataset, code, training logs, and systems optimizations all released publicly

• 📥 Availability: Already accessible through Ollama

🔗 More on Together AI and Ollama

Files to Prompt: Repository-to-LLM Tool

08 Apr 2025 - Simon Willison released a tool called "files-to-prompt" that allows users to concatenate a directory full of files into a single prompt for use with LLMs.

• 📁 Repository analysis: Makes it easy to send a whole GitHub repository to an LLM prompt

• 📏 Context optimization: Particularly useful for long context windows with models like Gemini 2.5

• 🔄 Format flexibility: Outputs in markdown or XML depending on model preferences

• 🛠️ Practical utility: Solves the common problem of needing to analyze multiple files together

🔗 More on GitHub