AI News: Anthropic Leak is Bigger Than You Think

This week brought significant developments in the AI landscape, starting with the unexpected leak of Anthropic's Claude code. The leaked source code revealed a sophisticated three-layer memory architecture that moves beyond traditional "store everything" retrieval. This system, at its core, uses "memory MD," a lightweight index of pointers that is perpetually loaded into context. This index stores locations rather than raw data, and transcripts are "grepped" for specific identifiers, meaning the system searches text for patterns to retrieve relevant information. Even more intriguing was the discovery of "Chyros" within the leaked code, described as an "autonomous demon mode." This feature would allow Claude to operate as an always-on background agent, performing memory consolidation while the user is idle. Chyros is designed to be a proactive AI that acts without explicit prompts. It runs 24/7, checking every few seconds if there's anything worth doing. If it acts, it can fix code errors, respond to messages, update files, and run tasks—essentially anything Claude can already do, but autonomously. It also possesses exclusive tools: push notifications to reach users outside the terminal, file delivery for sending created content without being asked, and pull request subscriptions to monitor GitHub for code changes. Examples of its potential include automatically restarting a website that goes down, replying to customer complaint emails at 2 AM, or fixing typos on a website, all while the user is away. This hints at a "post-prompting era" where AI becomes a proactive background agent, learning user needs and acting on their behalf. Similar proactive AI developments are also seen with OpenClaw's cron jobs and autonomous heartbeats. The leak also showed more evidence of "Capiara" (or "Mythos"), Anthropic's next-level model, and a "hidden buddy system" resembling a Tamagotchi-style terminal pet, which some speculate was a planned April Fool's joke that was sidelined due to the leak. Anthropic's head of Claude, Boris Churnney, attributed the leak to a developer error, emphasizing a focus on process improvement rather than individual blame. While no private customer information was leaked, the incident may lead to spin-offs of Claude code leveraging the open-source material. Recraft launched its V4 family of models, designed for professional aesthetics and agency-quality design assets. The V4 and V4 Pro models offer elevated art-directed outputs for cohesive brand visuals, website prototyping, product packaging, and detailed illustrations, understanding elements like lighting and negative space. The V4 vector models generate native vector graphics, such as logos and icons, as scalable and editable SVGs, unlike typical AI generations that only fake the look of vector art. These models demonstrate strict prompt adherence and handle text on images surprisingly well, a crucial feature for professional brand assets. They can also create photorealistic product mock-ups, complete with realistic lighting and shadows, at a fraction of the cost of traditional product shoots. OpenAI raised $122 billion at an $852 billion valuation, marking the largest raise by any company ever and making it the fastest-growing company in history. Microsoft was among the investors, dispelling rumors of a falling out between the two companies. OpenAI is now generating $2 billion in revenue per month. In their fundraising announcement, OpenAI also revealed plans to build a "unified AI super app." This super app aims to integrate ChatGPT, Codex, browsing, and broader agentic capabilities into a single, agent-first experience. The goal is to move away from disconnected tools, offering a system that understands intent, takes action, and operates across applications and workflows. This move mirrors Anthropic's approach with its unified Claude app and suggests a similar future for proactive, background AI agents, especially given OpenAI's recruitment of Peter Steinberger, creator of OpenClaw. OpenAI's decision to phase out Sora and focus on core competencies like intelligent models for chatting, research, coding, and agentic use cases was also noted. A Wall Street Journal article revealed that Sora was losing approximately a million dollars a day, making its shutdown financially understandable. Microsoft released MAI Transcribe 1, a new speech recognition model offering best-in-class accuracy across 25 languages, specifically built for transcription. Benchmarks show it significantly outperforms models like GPT Transcribe, Scribe V2, Gemini 3.1 Flash, and OpenAI's Whisper, even in noisy environments. The demonstration highlighted its ability to differentiate homonyms based on context. MAI Transcribe 1, along with MAI Voice One (text-to-speech) and MAI Image 2, is now available within Microsoft Foundry for developers. Google rolled out VO3.1 Light, a more cost-effective version of its video generation model. While likely having a trade-off in quality, it’s priced at 5 cents for 720p video, compared to 40 cents for standard VO3.1 with audio and 15 cents for VO3.1 Fast. VO3.1 Light does not support 4K generation. Google also announced price reductions for VO3.1 Fast, effective April 7th, hinting at potential further news around that date. Google also introduced "AI Inbox" with smart prioritization and daily personalized briefings, currently in beta for Google AI Ultra subscribers ($250/month). This feature provides suggested to-dos based on inbox content and topics to catch up on. Several new large language models were released. Google launched Gemma 4, an open-weight, open-source model designed to run on Android devices and laptop GPUs, suitable for local agentic systems like OpenClaw. Alibaba released two models: Quinn 3.5 Omni, an omnimodal model working with text, image, audio, and video, claiming to outperform Gemini 3.1 Pro in audio and match its audiovisual understanding. A standout feature is "audiovisual vibe coding," where users describe a vision to the camera, and Quinn 3.5 Omni Plus instantly builds a functional website or game. Days later, Alibaba released Quinn 3.6 Plus, designed for real-world agents with enhanced multimodal vision and a 1 million token context window. Benchmarks show it competing with models like Opus 4.5 and Gemini 3 Pro, excelling in software engineering, real-world Q&A, video reasoning, and document recognition. Finally, RC, a new company, released Trinity Large Thinking, an open-source model under Apache 2.0, showing comparable performance to models like Opus 4.6 and GLM5, and demonstrating capabilities in code generation and agentic workflows. These open-source models continue to gain ground on state-of-the-art closed-source models. In rapid-fire news, Claude's computer use feature, allowing it to control mouse and keyboard, is now available in Claude Code for Pro or Max subscribers. A CodeX plugin for Claude Code allows users to integrate OpenAI's CodeX for code review and task delegation within Claude Code's UI. ChatGPT is now available in Apple CarPlay. OpenAI acquired TBPN (Tech Business Production Network), a daily live show focused on tech, business, and production. This acquisition was unexpected, given OpenAI's stated focus on core competencies, and raises questions about editorial independence. Perplexity released "Computer for Taxes," a new use case for its computer model that loads tax modules for US federal tax questions, assisting with drafting returns, reviewing prepared returns, and building tools for complex tax code. Salesforce is making Slack more "agentic" with 30 new capabilities, transforming Slackbot into a "teammate." Features include meeting transcription, reusable AI skills, native customer management, deep research, and voice input, allowing Slack to handle chatbot-like tasks and even record external meetings. These features are rolling out to business plus and enterprise plus plans, with free and pro plans gaining access soon. GM is using AI to visualize cars before they exist, transforming hand-drawn sketches into concept videos, suggesting design changes, and running early aerodynamic testing, significantly compressing the design timeline. Lastly, Instacart introduced physical AI "smart carts" with touchscreens, cameras, location sensors, and scales, powered by Nvidia Jetson. These carts process real-time data on aisle activity and basket contents to influence grocery shopping with timely recommendations, aiming to improve effectiveness compared to post-checkout suggestions.