Anthropic Leak Reveals ‘Demon Mode’ AI Agents
The world of artificial intelligence saw a major development this week with the unexpected leak of Anthropic’s Claude code. This leak has given the public an unprecedented look into the company’s cutting-edge research and future plans. While Anthropic is working to remove the leaked material, the code is already being used by developers to build new AI versions.
A Sophisticated Memory System
One of the most significant discoveries from the leaked code is a new three-layer memory architecture. Instead of storing everything, this system uses a self-healing memory. At its heart is ‘memory MD,’ a lightweight index that is always loaded into the AI’s context. This index doesn’t store raw data but rather the locations of information. This means the AI doesn’t need to read through long transcripts; it can quickly find specific details when needed.
Introducing ‘Chyros’: The Autonomous Agent
Perhaps the most striking revelation is a feature called ‘Chyros.’ This represents a new direction in user experience, often described as an ‘autonomous demon mode.’ Essentially, Chyros allows Claude to act as an always-on background agent. It works quietly while users are busy or even asleep, performing tasks like consolidating memory. It receives regular prompts asking if there’s anything important to do. If action is needed, it can fix code errors, respond to messages, update files, or run other tasks that Claude can already perform, but without direct commands.
Proactive AI: The ‘Post-Prompting’ Era
Chyros has unique tools: push notifications to alert users, file delivery for sending created content, and pull request subscriptions to monitor code changes on platforms like GitHub. This moves AI beyond simply responding to prompts. It suggests a future where AI acts proactively, learning user needs and performing tasks in the background. This shift towards AI becoming more integrated and less directly controlled is seen as a major step forward. Other companies like OpenClaw are also exploring similar autonomous functions.
Other Anthropic Discoveries
The leak also provided more details about ‘Capiara,’ believed to be Anthropic’s next-generation AI model. Evidence of a ‘hidden buddy system,’ a Tamagotchi-like digital pet with stats like chaos and snark, was also found. Many speculate this was intended as an April Fool’s joke that was overshadowed by the code leak.
Recraft V4: Elevating Design Assets
Beyond Anthropic, Recraft has launched its V4 family of AI models for professional design. These new models focus on creating cohesive brand visuals, website prototypes, packaging, and scalable vector illustrations. The V4 and V4 Pro models are designed for photorealistic scenes and complex compositions, understanding lighting and materials. The V4 vector models can generate actual, editable SVG files, a significant improvement over AI tools that only mimic vector art. Recraft V4 aims to provide agency-quality design assets efficiently.
OpenAI’s Massive Funding and Super App Vision
OpenAI has secured a staggering $122 billion in funding, valuing the company at $852 billion. This makes it the fastest-growing company in history, with revenues reportedly growing four times faster than tech giants like Alphabet and Meta. Microsoft remains a key investor in this new round. OpenAI revealed plans to build a unified AI ‘super app.’ This app would combine chat, GPT models, coding tools, browsing, and agentic capabilities into a single, intuitive experience. The goal is to move beyond disconnected tools and create a system that understands user intent and acts across various applications and data sources.
Microsoft’s AI Transcription Breakthrough
Microsoft has released MAI Transcribe 1, a new speech recognition model. It boasts best-in-class accuracy in 25 languages and is specifically built for transcription tasks. Benchmarks show it significantly outperforms models like GPT Transcribe and OpenAI’s Whisper, even in noisy environments. This technology is now accessible through Microsoft Foundry, allowing developers to build on top of Microsoft’s AI models.
Google’s Video Model Updates and AI Inbox
Google has introduced VO3.1 Light, a more affordable version of its video generation model. While potentially lower in quality than previous versions, it offers a significantly reduced cost for generating video content. Google is also lowering prices for its VO3.1 Fast model. Additionally, Google is testing ‘AI Inbox,’ a feature for Google AI Ultra subscribers that offers smart email prioritization and personalized daily briefings. This tool aims to help users manage their inboxes more effectively.
New Open-Source Models Emerge
Several new large language models have been released. Google launched Gemma 4, an open-weight model designed for local use on devices like Android phones and laptops. Alibaba released two new models: Quinn 3.5 Omni, an omnimodal model capable of understanding text, image, audio, and video, and Quinn 3.6 Plus, optimized for real-world agents with a massive 1 million token context window. RC released Trinity Large Thinking, another open-source model showing strong performance on benchmarks and capable of complex tasks like code generation and app building.
The Future of AI: Proactive and Integrated
The trend towards more proactive and integrated AI is clear. From Anthropic’s ‘Chyros’ to OpenAI’s ‘super app’ vision and Google’s AI Inbox, the focus is shifting from users commanding AI to AI anticipating and fulfilling needs. While closed-source models often lead, the rapid advancements in open-source alternatives suggest a future where powerful AI tools are increasingly accessible and adaptable.
Source: AI News: Anthropic Leak is Bigger Than You Think (YouTube)
