About This Week’s Covers

This week’s cover celebrates the opening of the Minecraft movie. I used the Minecraft theme to test all of the various image generation tools to see if they could create a complete cover, including the text. I’m not doing any post-production editing to really try to show how the image engines are doing without any help. I’ll share examples of each engine below this description. Ideogram was the consistent winner for this week’s challenge (above).

A Minecraft-style screenshot with a 16:9 aspect ratio for an AI newsletter cover. The scene shows a vibrant spring meadow with cherry blossom trees in bloom and flowers dotting the landscape. A friendly blocky robot character is sitting under a tree, reading a digital newspaper with holographic AI symbols floating above it. There’s a small farm with AI helper robots tending to crops. In the sky, cubic drones are flying in formation. The Minecraft-style sun is rising/setting with beautiful pixel art lighting. Text overlay at the top reads “AI News 80: 2025/04/11” in the classic Minecraft font. Include some spring animals like rabbits or bees interacting with small robot companions.

Below: ChatGPT did a great job, but it was a bit dark and curvy for MineCraft. MidJourney was horrible. Flux Pro did great but wrote “Minecraft” across the top. Claude wrote the image as SVG code, which is cool, but not the winner.

The category images are also references to Minecraft. I used Claude 3.7 and Flux Pro Ultra to automate the creation of 40 category covers in one bulk command (aka one press of a button generates all 40 images). Here are my favorite six from the batch. The goal is to create these quickly, not perfectly, and test the capabilities of the models. Shout out to my nephew Calvin (resident Minecraft expert) for selecting the best of four tries for each category.

This Week By The Numbers

Total Organized Headlines: 567

This Week’s Executive Summaries

It was another jam-packed week of artificial intelligence developments. The pieces are coming together quickly to enable AI systems to talk to each other, and the memory capabilities continue to grow. Those are the two biggest themes, and I believe the pace of change will start to quicken now that the pieces are in place.

ChatGPT can now remember and reference your entire conversation history going back in perpetuity.

Google launched their own open agent protocol that allows agents from different vendors to communicate. At the same time, Google announced they will adopt and support Anthropic’s model context protocol (MCP) agent structure. One cool thing about Google’s new protocol is that it allows agents to discover and read each other’s capabilities through ‘business cards’ that explain what each agent is capable of completing.

OpenAI has gone on the offensive and is suing Elon Musk, claiming he has no goals other than to be a self-serving heckler.

The CEO of Shopify sent a memo out to all employees saying that everyone needs to be proficient in AI in order to be employed there.

Google released a Deep Research competitor to OpenAI and Perplexity. Thanks to Google’s deep integration with search, Google’s Deep Research is already outperforming PhD-level researchers when given access to the web.

Amazon launched a web operations AI last week, and this week they released an incredibly powerful audio model, which can both listen and respond and process tasks within one model. Clearly, at least for now, Amazon is committed to the next iteration of Alexa.

Most people think of NVIDIA when it comes to AI chips; however, Google has been making their own proprietary chips for a long time. Their newest chip is insanely powerful with 4,614 trillion operations per second. Take a second to think about that.

The Washington Post predicts a massive job disruption among white-collar workers similar to that of outsourcing jobs overseas.

NVIDIA has put its stake in the ground regarding embodied robots. While people debate the ethics of large language models and image creation, the world of artificial intelligence leaders has already moved on to creating millions of humanoid robots in the near future that will start to appear all around us. This hour-and-a-half video is worth watching.

OpenAI released a benchmark to test how well AI systems can reproduce scientific research. Right now, artificial intelligence is only at a 21% level compared to humans.

University of California San Diego researchers demonstrated that GPT-4.5 successfully passes the Turing test. The AI system was able to fool 75% of humans into thinking it was human.

Google has integrated vision into its AI tools, finally catching up to OpenAI’s multimodal product.

Audio leader ElevenLabs has adopted Anthropic’s model context protocol (MCP) to enable locally running servers to create voice agents to make outbound calls and clone voices.

Adobe plans to integrate AI agents throughout its Creative Cloud product lineup to automate the mundane parts of creative operations as well as improve context-aware search within videos and photos.

The new MidJourney image creation model came out and is a bit of a dud, sadly.

All this and more in this week’s artificial intelligence newsletter!!! All gas, no brakes!

ChatGPT Expands Memory to Include All Past Conversations
ChatGPT now references your entire conversation history to provide more personalized responses. The improved memory system automatically draws on your preferences and interests from previous chats, making interactions more relevant for writing, advice, and learning. Users maintain full control—you can opt out of memory features or use temporary chats when privacy is preferred. This update represents OpenAI’s vision for AI systems that become increasingly personalized over time.

“we have greatly improved memory in chatgpt–it can now reference all your past conversations! this is a surprisingly great feature imo, and it points at something we are excited about: ai systems that get to know you over your life, and become extremely useful and personalized.” / X https://x.com/sama/status/1910380643772665873

“Starting today, memory in ChatGPT can now reference all of your past chats to provide more personalized responses, drawing on your preferences and interests to make it even more helpful for writing, getting advice, learning, and beyond. https://x.com/OpenAI/status/1910378768172212636

“As always, you’re in control of ChatGPT’s memory. You can opt out of referencing past chats, or memory altogether, at any time in settings. If you’re already opted out of memory, you’ll be opted out of referencing past chats by default. If you want to change what ChatGPT knows” / X https://x.com/OpenAI/status/1910378772789854698

OpenAI Countersues Elon Musk After Ongoing Dispute
OpenAI has filed a countersuit against Elon Musk, claiming his actions aim to impede the organization’s progress and gain control of AI innovations for personal gain. The company states they are preparing to establish a well-funded nonprofit structure, contradicting Musk’s allegations about their organizational direction. OpenAI claims Musk previously attempted to merge the company with Tesla as a for-profit venture before departing when his proposal was rejected. While acknowledging Musk’s entrepreneurial accomplishments, OpenAI characterizes his current actions as self-serving and inconsistent with their mission.

“Elon’s nonstop actions against us are just bad-faith tactics to slow down OpenAI and seize control of the leading AI innovations for his personal benefit. Today, we counter-sued to stop him.” / X https://x.com/OpenAINewsroom/status/1910105151492575611

Big News: Google Unveils Agent2Agent Protocol for AI Collaboration
Google has launched Agent2Agent (A2A), an open protocol enabling AI agents from different vendors to communicate and collaborate across platforms. The protocol allows agents to discover each other’s capabilities through JSON-based “Agent Cards,” negotiate tasks, and work together on complex enterprise workflows regardless of their underlying frameworks. A2A supports long-running tasks, multimodal interactions (text, audio, video), and enterprise-grade security. Built on standard technologies (HTTP, JSON-RPC, SSE), the protocol complements Anthropic’s Model Context Protocol and has garnered support from over 50 technology partners including Salesforce, Workday, and major consulting firms. In parallel, Google released their Agent Development Kit, allowing developers to build multi-agent applications in under 100 lines of Python code.

“Huge news. Google just released their Agent Development Kit (ADK). It’s an open-source framework to build, manage, evaluate and deploy multi agents. You can get a multi-agent application running in <100 lines of Python. https://x.com/LiorOnAI/status/1910041530183893221

GitHub – google/A2A: An open protocol enabling communication and interoperability between opaque agentic applications. https://github.com/google/A2A

“NEW: Google announces Agent2Agent Agent2Agent (A2A) is a new open protocol that lets AI agents securely collaborate across ecosystems regardless of framework or vendor. Here is all you need to know: https://x.com/omarsar0/status/1909977142311690320

Announcing the Agent2Agent Protocol (A2A) – Google Developers Blog https://developers.googleblog.com/en/a2a-a-new-era-of-agent-interoperability/

“Regarding the new Google agent-to-agent protocol: Love the idea of agents having “cards” Humans get business cards, why not AI?” / X https://x.com/mathemagic1an/status/1910198673512017947

“NEW: Google presents Agent Development Kit (ADK) Features: – code-first – multi-agents – rich tool ecosystem – flexible orchestration – integrated dev xp – development-ready – streaming – state, memory, artifacts – extensibility > pip install google-adk https://x.com/omarsar0/status/1910004370864742757

Shopify CEO Demands AI Adoption as “Baseline Expectation” for All Staff
Shopify CEO Tobi Lütke declared AI usage a mandatory skill for all employees in a company-wide memo that later circulated online. The directive states that effective AI use is now “a fundamental expectation of everyone at Shopify” with Lütke bluntly adding he “cannot see this working out” for those who opt out. The company will add AI proficiency questions to performance reviews and require teams to demonstrate why tasks cannot be accomplished with AI before requesting additional headcount. Lütke emphasized this applies to everyone including himself and the executive team, positioning the mandate as essential to Shopify’s mission of reimagining entrepreneurship in an AI-powered world.

“Reflexive AI usage is now a baseline expectation at Shopify” / X https://x.com/tobi/status/1909251946235437514

Google Enters The Deep Research Game In A Big Way
Competing with and potential beating OpenAI and Perplexity, Google has expanded its Deep Research capability to Gemini Advanced subscribers using the 2.5 Pro Experimental model. This enhancement improves the AI’s ability to analyze information and create comprehensive research reports on virtually any topic. According to Google’s testing, users preferred reports from the 2.5 Pro-powered Deep Research over competing services by a 2-to-1 margin. The feature works across web, Android, and iOS platforms, and includes an Audio Overviews option that converts reports into podcast-style conversations for on-the-go consumption. Users can access Deep Research by selecting “Gemini 2.5 Pro (experimental)” from the dropdown menu and tapping “Deep Research” in the prompt bar.

The Deep Research feature is now powered by the Gemini 2.5 Pro model, making it even more effective! Give it a try to generate detailed, easy-to-read reports on just about any research topic at https://t.co/oSWSqcZCZw

“Your favorite personal AI research assistant just got smarter. 🧠 Deep Research on @GeminiApp is now available to Advanced users on Gemini 2.5 Pro, our most intelligent model – making it even better at analyzing information to create insightful reports on almost any topic → https://x.com/GoogleDeepMind/status/1909943627218129004

“Google Gemini 2.5 is the first public AI model to definitively beat the level of performance of human PhDs with access to Google on hard multiple choice problems inside their field of expertise (around 81%). All AI tests are flawed, but GPQA Diamond has been a pretty good one.” / X https://x.com/emollick/status/1907737487176286418

Amazon Has Entered the Chat with Nova Sonic (who names these things): All-in-One Voice AI for Natural Conversations
The new Alexa might be pretty awesome. Amazon has launched Nova Sonic, a unified voice model that handles both speech understanding and generation in a single system, eliminating the need to stitch together separate models while preserving important speech elements like tone and speaking style. The platform supports function calling and integration with external services through RAG technology, currently available for American and British English with more languages planned. Accessible through a new bidirectional streaming API that enables natural conversational flow with smooth handling of interruptions, Nova Sonic includes built-in content moderation and watermarking protections and is now available in the US East region with a 300K context window and support for 8-minute sessions.

Introducing Amazon Nova Sonic: Human-like voice conversations for generative AI applications | AWS News Blog https://aws.amazon.com/blogs/aws/introducing-amazon-nova-sonic-human-like-voice-conversations-for-generative-ai-applications/

Amazon’s Nova Sonic foundation model understands voice in a whole new way https://www.aboutamazon.com/news/innovation-at-amazon/nova-sonic-voice-speech-foundation-model

“Amazon launched Nova Sonic speech-to-speech AI for human-like interactions —Outperforms OpenAI’s voice models with ~ 80% less cost —4.2% word error rate across languages — 46.7% better accuracy than GPT-4o for noisy environments —On Amazon Bedrock https://x.com/rowancheung/status/1909845011551633891

Google Unveils Ironwood, An Unfathomably Powerful AI Chip
Google has introduced Ironwood, its seventh-generation Tensor Processing Unit (TPU), designed specifically for AI inference tasks. Each chip delivers 4,614 trillion calculations per second with 192GB of high-speed memory. At maximum configuration, a 9,216-chip pod achieves 42.5 exaflops – 24 times more powerful than the world’s largest supercomputer. The system offers twice the power efficiency of Google’s previous TPU generation, with specialized capabilities for handling complex AI models that perform reasoning tasks. Ironwood will be available to Google Cloud customers later this year, supporting both smaller 256-chip configurations and the full-scale version for the most demanding AI workloads.

Ironwood: The first Google TPU for the age of inference https://blog.google/products/google-cloud/ironwood-tpu-age-of-inference/

“Google’s TPUv7 is out! ML accelerator marketing material is usually pretty inscrutable (what numbers are even comparable?), so here I’ll explain concretely how this compares with Nvidia. 🧵 https://x.com/itsclivetime/status/1910026066129014868

“Google just announced Ironwood, their 7th-gen TPU competitor to Nvidia’s Blackwell B200 GPUs – 4,614 TFLOPs per chip (FP8) – 192 GB HBM, 7.2 Tbps HBM bandwidth – 1.2 Tbps bidirectional ICI – 42.5 exaflops per 9,216-chip pod (24x El Capitan) https://x.com/scaling01/status/1909949372965564896

“Google’s new TPUv7 rumored to have 2000x the performance of the latest iPhone” / X https://x.com/scaling01/status/1909958867066175802

Washington Post: AI Impact on Knowledge Workers Mirrors Manufacturing’s “China Shock”
The Washington Post identifies a parallel between today’s “AI shock” and manufacturing’s earlier “China shock,” with automation now targeting knowledge workers, particularly in entry-level white-collar positions.

“⚡ The Washington Post draws a crucial parallel: We’re facing an “AI shock” similar to manufacturing’s “China shock” – but hitting knowledge workers. Especially entry-level, white-collar work could get automated. The key difference? “Winning the AI tech competition with other” / X https://x.com/fdaudens/status/1908226743636246875

“🔄 Apparently, the Great Recalibration has begun. We’re now heading into an era where AI is fundamentally redefining the nature of work itself, by forcing a complete reassessment of human value in the workplace, according to a piece in Fast Company. But it might be driven more by” / X https://x.com/fdaudens/status/1908226742310867065

Microsoft Copilot Adds Web Browsing and Action Features
Microsoft Copilot can now use websites, including booking tickets and restaurant reservations through partnerships with travel and shopping sites. The AI assistant can also monitor online deals, generate conversational podcasts about content, and view what’s on your camera or screen to answer questions. Copilot introduces Pages for organizing research and Deep Research for complex queries. The system remembers user preferences but allows control over what information is stored. These updates arrive as Microsoft reportedly considers incorporating more of its in-house technology into Copilot, which has traditionally relied on OpenAI models.

Microsoft’s Copilot can now browse the web and perform actions for you | TechCrunch https://techcrunch.com/2025/04/04/microsofts-copilot-can-now-browse-the-web-and-perform-actions-for-you/

PayPal Launches MCP Servers for AI-Powered Commerce
PayPal rolled out Model Context Protocol (MCP) servers to help developers implement AI-driven capabilities for merchants. The Anthropic-created standard allows AI agents to access third-party services like PayPal through a unified protocol. The initial feature focuses on automated invoice generation, letting merchants create invoices through conversational AI prompts without visiting PayPal’s website or using traditional APIs. Developers can access PayPal’s MCP through either a downloadable local server compatible with Claude Desktop and Cursor AI, or a remote server option that works across multiple devices with PayPal authentication.

“Today, we launched @PayPal’s #ModelContextProtocol server, enabling #developers to create next-gen payment experiences powered by agentic #AI. Our journey starts with PayPal Invoicing feature, available now. Read more here. https://x.com/paypaldev/status/1907477953501605955

Great Vdeo on Humanoid Robots from NVIDIA’s GTC Event
NVIDIA released a video featuring leaders from top robotics companies discussing the emerging era of humanoid robots. Executives from 1X, Boston Dynamics, Skild AI, NVIDIA, Agility Robotics, and Tiffintech explore how these human-shaped machines could become the most widely deployed form of general-purpose robotics in history. The panel examines the critical role of foundation models, simulation technologies, and advanced computing in making versatile humanoid robots a reality, comparing this development to ChatGPT’s impact on AI conversation.

A New Era for Generalist Robotics: The Rise of Humanoids | NVIDIA GTC 2025 – YouTube https://www.youtube.com/watch?v=BmD22FNOAY4

OpenAI’s PaperBench Will Track Improvement of AI Models Replicating Research
OpenAI has introduced PaperBench, a framework testing AI systems’ ability to reproduce machine learning research results from scientific papers. In tests, AI agents were given research PDFs, web access, and coding tools to replicate findings within 12 hours—tasks that typically take human experts days. Claude 3.5 Sonnet achieved the highest replication score at approximately 21%, demonstrating both progress and the significant gap remaining between AI and human PhD-level expertise. Researchers noted that models performed better when encouraged to “think longer” rather than stopping prematurely. While each replication attempt costs hundreds of dollars in computing resources, this research represents a step toward AI systems that could potentially advance scientific discovery independently.

“.@OpenAI dropped a new research paper showing AI agents are now capable of replicating cutting-edge AI research papers from scratch. This is one step closer to the Intelligence Explosion: AI that can discover new science and improve itself. Here’s what they learned: 🧵 https://x.com/MatthewBerman/status/1907890944223097145

GPT-4.5 Passes Turing Test in UC San Diego Study
UC San Diego researchers have demonstrated that GPT-4.5 successfully passes Alan Turing’s famous test of machine intelligence. In a three-party experiment where participants spoke simultaneously with both a human and AI, GPT-4.5 (when given a human-like persona) was mistakenly identified as the human 73% of the time. This represents significant improvement over previous studies where GPT-4 was judged as human only about 50% of the time in traditional two-party tests. The findings, published in a new preprint, suggest current AI systems can consistently fool humans into believing they’re communicating with another person.

“New preprint: we evaluated LLMs in a 3-party Turing test (participants speak to a human & AI simultaneously and decide which is which). GPT-4.5 (when prompted to adopt a humanlike persona) was judged to be the human 73% of the time, suggesting it passes the Turing test (🧵) https://x.com/camrobjones/status/1907086860322480233

“Researchers at UC San Diego demonstrated that AI systems can consistently pass Alan Turing’s famous test of machine intelligence OpenAI’s GPT-4.5 was mistaken for human nearly three-quarters of the time in controlled trials https://x.com/adcock_brett/status/1908913665706721489

Google Expands AI Seach Mode With Multimodal Features
Google is extending AI Mode to millions more U.S. users while adding multimodal capabilities from Lens. The feature allows users to upload images or take photos and ask complex questions about what they see. AI Mode combines Gemini’s visual understanding with Google’s search technology to comprehend entire scenes, identify objects, and provide contextually relevant responses with helpful links. Early feedback shows users are submitting queries twice as long as traditional searches, particularly for exploratory questions and complex tasks like product comparisons and trip planning.

AI Mode in Google Search adds multimodal search https://blog.google/products/search/ai-mode-multimodal-search/

Google DeepMind and Gemini to Support Anthropic’s MCP Protocol
Google DeepMind CEO Demis Hassabis announced that the company will implement Anthropic’s Machine Control Protocol (MCP) for its Gemini AI models and software development kit. The protocol, which helps manage AI systems that can take actions, is gaining traction as an industry standard. DeepMind plans to collaborate with Anthropic and other companies to further develop the protocol for the emerging era of AI agents.

“MCP is a good protocol and it’s rapidly becoming an open standard for the AI agentic era. We’re excited to announce that we’ll be supporting it for our Gemini models and SDK. Look forward to developing it further with the MCP team and others in the industry” / X https://x.com/demishassabis/status/1910107859041271977

ElevenLabs Launches MCP Server for AI Voice Integration
ElevenLabs has introduced its Model Context Protocol (MCP) server, allowing users to access the company’s AI audio capabilities through text prompts in tools like Claude and Cursor. The locally-running server connects to ElevenLabs’ cloud APIs, enabling users to create voice agents that can make outbound calls, transcribe audio, generate speech, and clone voices. The system supports practical applications like meeting transcription and creative uses such as developing character voices for entertainment projects, all while giving users control over their data and workflows.

“The WhatsApp MCP server can now send and receive images, videos, and voice notes Combine it with the new ElevenLabs MCP server to give it superpowers — using AI to transcribe the voice notes and send audio messages with 3,000+ voices https://x.com/LukeHarries_/status/1909303780941640041

“Introducing the official ElevenLabs MCP server. Give Claude and Cursor access to the entire ElevenLabs AI audio platform via simple text prompts. You can even spin up voice agents to perform outbound calls for you — like ordering pizza. https://x.com/elevenlabsio/status/1909300782673101265

Anthropic Study Reveals How College Students Use Claude
Anthropic analyzed one million education-related conversations with Claude, finding students primarily use the AI for creating educational content (39.3%) and technical explanations (33.5%). Computer Science students are the most active Claude users, accounting for 38.6% of conversations despite representing only 5.4% of US degrees. The study identified different interaction patterns across disciplines—Science and Math students favored collaborative problem-solving while Education students tended toward direct output creation. Researchers also flagged concerning behaviors suggesting academic dishonesty, including requests for answers and help avoiding plagiarism detection. The findings raise questions about students potentially offloading important cognitive tasks to AI assistants.

“New Anthropic research: How university students use Claude. We ran a privacy-preserving analysis of a million education-related conversations with Claude to produce our first Education Report. https://x.com/AnthropicAI/status/1909626720476365171

Adobe Is Creating AI Agents to Assist Creative Production
Adobe plans to integrate AI agents throughout its product lineup to help users work more efficiently. The company is developing agent technology for Acrobat to analyze documents and serve as research assistants, Express to guide non-designers through creation processes, and Creative Cloud apps to handle repetitive tasks. In Photoshop, Adobe will introduce a new Actions panel that suggests context-aware edits with one-click implementation, while in Premiere Pro, Media Intelligence now helps editors quickly find specific visual content in video clips. Adobe emphasizes these tools are designed to complement human creativity rather than replace it, positioning AI agents as assistants that free professionals to focus on higher-level creative work.

Our vision for accelerating creativity and productivity with agentic AI | Adobe Blog https://blog.adobe.com/en/publish/2025/04/09/our-vision-for-accelerating-creativity-productivity-with-agentic-ai

WordPress Launches AI Website Builder
WordPress has introduced an AI website builder that creates complete websites based on simple text prompts. Users describe their vision, and the system automatically generates a fully-designed site with text, layouts, and images. The tool targets entrepreneurs, freelancers, and bloggers who need professional sites without technical skills. Each account receives 30 free prompts, with unlimited access available through hosting plans. While not yet supporting e-commerce or complex integrations, the builder maintains full WordPress compatibility for manual customization. The tool is currently available only for new WordPress.com websites.

Try Our New AI Website Builder for Free – WordPress.com https://wordpress.com/blog/2025/04/09/ai-website-builder/

The New MidJourney Is Out and It’s Not Impressing People
I kicked the tires on it, and it failed almost every test I gave it. I still love the OG MidJourney for certain styes, but at this point it should be a lot further ahead, sadly. TLDR: Most people agree.

“So, let’s hear those Midjourney v7 reviews. Did they cook? Ngl, doesn’t seem promising…” / X https://x.com/bilawalsidhu/status/1908180816963739774

Runway Speeds Up AI Video Generation with Gen-4 Turbo
Runway has launched Gen-4 Turbo, a faster version of its AI video model that can produce 10-second videos in just 30 seconds. This accelerated generation time makes the tool particularly useful for quick creative iterations and experimentation. The company is rolling out Gen-4 Turbo across all subscription plans, including the free Basic tier which comes with 125 one-time credits.

“Runway released Gen4, its next-gen video generation model It produces 5-10 second videos at 1080p with strong consistency in characters, objects, and locations, and world understanding Also delivers better prompt adherence than Gen3 https://x.com/adcock_brett/status/1908913485800431653

“Runway released Gen-4 Turbo, a faster version of its new AI video model It can produce 10-second videos in just 30 seconds. Now rolling out across all plans, including the free ‘Basic’ tier with 125 one-time credits. https://x.com/rowancheung/status/1909496768473014408

“Today we’re introducing Gen-4 Turbo. The fastest way to generate with our most powerful video model yet. With Gen-4 Turbo it now takes just 30 seconds to generate a 10 second video, making it ideal for rapid iteration and creative exploration. Now rolling out across all plans. https://x.com/runwayml/status/1909302613192876102

Meta Introduces Llama 4 Models with Massive Context Window
Meta has unveiled Llama 4 Scout and Llama 4 Maverick, its most advanced AI models to date. The 17B-parameter Scout features a 10-million token context window—large enough to process multiple classic book series simultaneously with room to spare. “You could fit the entire Harry Potter series, A Song of Ice and Fire (Books 1–5), The Lord of the Rings, The Hobbit, The Bible, The Quran, and Dune. Still with millions of tokens remaining.” Maverick, with 128 experts, delivers superior image understanding capabilities and outperforms GPT-4o and Gemini 2.0 Flash on key benchmarks. Both models stem from the still-in-training Llama 4 Behemoth, which is already showing results that surpass models from OpenAI, Anthropic, and Google on STEM tasks.

“Llama 4 is here! Meta has released two smaller versions of its new Llama 4 family of models: Llama 4 Scout and Maverick, and announced a larger version called Behemoth that is still in training. In this thread, we dig into their training details and benchmark performance 🧵” / X https://x.com/EpochAIResearch/status/1909699970594394173

“Today is the start of a new era of natively multimodal AI innovation. Today, we’re introducing the first Llama 4 models: Llama 4 Scout and Llama 4 Maverick — our most advanced models yet and the best in their class for multimodality. Llama 4 Scout • 17B-active-parameter model https://x.com/AIatMeta/status/1908598456144531660

“Llama 4 just released with a 10M context window. You could fit the entire Harry Potter series, A Song of Ice and Fire (Books 1–5), The Lord of the Rings, The Hobbit, The Bible, The Quran, and Dune. Still with millions of tokens to spare. This is library-scale reasoning. https://x.com/skirano/status/1908613559069635032

Together AI and Agentica Team Launch DeepCoder
Together AI and Agentica have released DeepCoder-14B-Preview, a code reasoning model that matches the performance of much larger systems while using only 14 billion parameters. The model achieves 60.6% accuracy on LiveCodeBench and performs well on competition-level coding tasks. Despite its focus on coding, DeepCoder shows strong capabilities in mathematics, scoring 73.8% on AIME problems. The team has made the dataset, code, training logs, and system optimizations publicly available to support broader research in reinforcement learning for AI development.

“DeepCoder-14B-Preview achieves 60.6% on LiveCodeBench and a 1936 score on CodeForces, performing on par with o3-mini (low) and o1 on competition-level coding tasks. Despite being trained solely on coding data, DeepCoder generalizes remarkably well to the math domain, reaching https://x.com/togethercompute/status/1909697131645903065

“NEW code model in town, DeepCoder 14B – beats O3 Mini, MIT licensed, works w/ vLLM, TGI, Transformers and more! 🔥 https://x.com/reach_vb/status/1909706239577329915

“Announcing DeepCoder-14B – an o1 & o3-mini level coding reasoning model fully open-sourced! We’re releasing everything: dataset, code, and training recipe.🔥 Built in collaboration with the @Agentica_ team. See how we created it. 🧵 https://x.com/togethercompute/status/1909697122372378908

DeepCoder: A Fully Open-Source 14B Coder at O3-mini Level https://www.together.ai/blog/deepcoder

6 AI Visuals and Charts: Week Ending April 11, 2025

Pew Research Center on X: “AI experts and public largely see jobs for cashiers, journalists and factory workers at risk due to AI; views differ widely on truck drivers. https://t.co/VfoVGg1fnN https://t.co/BmnqIwgvKA” / X https://x.com/pewresearch/status/1907856958826139684

The 2025 AI Index Report | Stanford HAI https://hai.stanford.edu/ai-index/2025-ai-index-report

“@tkasasagi Edo period cat meme https://x.com/hardmaru/status/1908022570789773516

“Now with GPT-4o a year and a half later. (The prompts are just the punchlines: “Interrupting cow,” etc. Interestingly the better prompt adherence makes them kind of boring compared to less faithful image models) https://x.com/emollick/status/1908738416994295963

“Runway Gen 4 (below) is a big step up for the Runway models, but I think Veo 2 is still the winner for best video now. This is using the same prompt: “golden retriever running through a brutalist art gallery with modern art on the walls as a single red balloon floats overhead” https://x.com/emollick/status/1908008991357079813

AI Index 2025: State of AI in 10 Charts | Stanford HAI https://hai.stanford.edu/news/ai-index-2025-state-of-ai-in-10-charts