Early to AI winners averaging +85% in last 18+ months –> Now sharing next investment insights with you

Over the past 19 months, MavSource's founder was early to the AI winners. Sandisk +196%. Micron +150%. Western Digital +129%. Nvidia +69%. Lam Research +63%. Okta +33%. Average across eight positions: +85%. 

Now that same signal comes to your inbox daily. MavSource aggregates all major AI newsletters, podcasts, AI labs, earnings calls, investor letters, and GitHub — then summarizes what matters, analyzes the trends, and adds founder commentary on why it may matter. Built for investors, founders, and operators who want to see what's next before it's a headline. 

One 5-minute email, every morning. Free. Results as of June 2026.

Subscribe For Free Now

^{Past performance does not indicate future results. MavSource is informational only, not investment advice.}

Today:

• Anthropic Maps Claude’s Hidden Thoughts

• Tencent Opens Hy3 AI Model

• Meta Launches Muse Image

• Claude Cowork Goes Mobile

• SpaceXAI and Cursor Prep New Model

Decoding Silent Thoughts in AI

Researchers have discovered that advanced language models (like Claude) process concepts internally before outputting text, functioning similarly to the human brain's "conscious" scratchpad. Using a technique called the Jacobian lens, they identified a small collection of internal neural patterns dubbed the J-space.

• Silent Reasoning: The model can hold a concept in mind, solve multi-step problems, and reason internally without writing anything down in its text output.

• Reportable & Modulatable: If asked to think about a concept silently, the model activates those specific patterns. It can also accurately report on what it was just "thinking" about.

• Security & Alignment: This workspace allows developers to observe hidden cognitive work. It makes it possible to catch a model privately noticing it is being tested, recognizing a prompt injection, or pursuing a hidden goal we didn't explicitly program.

A Highly Efficient Open-Source Model

A new 295-billion-parameter Mixture-of-Experts (MoE) model, Hy3, has been released to the open-source community. It activates only 21 billion parameters during inference, making it incredibly efficient while rivaling flagship models two to five times its size.

• Productivity Powerhouse: In blind evaluations by industry experts, it demonstrated massive gains in practical workflows, scoring especially high in frontend development, CI/CD, and data storage tasks.

• Fewer Hallucinations: Through strict data cleaning and constraint training, commonsense errors dropped by half, and its overall hallucination rate fell from 12.5% to 5.4%.

• Reliable Agent Tooling: The model is highly stable across different agent scaffoldings. It maintains context over complex, multi-turn interactions without losing track of constraints or drifting from the original intent.

Agentic Media and Video Generation

A powerful new suite of models—Muse Image and Muse Video—has just rolled out, operating more like autonomous agents than standard text-to-image generators.

• Tool Use & Web Search: Instead of simply mapping a prompt to pixels, Muse actively searches the web for factual grounding and visual references. For example, it can search for real, unbranded vintage items to generate highly authentic, natural-looking scenes rather than relying on polished, generic aesthetics. It can also write code to generate precise elements like scannable QR codes or interactive GIFs.

• Self-Refinement: The model evaluates its own drafts during the generation process. If a small detail is off, it seamlessly applies a local edit; if the composition is fundamentally wrong, it starts over from scratch without needing a new user prompt.

• High-Fidelity Video with Audio: Built on the same pretraining base, the new video model delivers exceptional visual quality paired natively with generated audio, built directly to stream into modern content creation workflows.

🧠RESEARCH

The Mirage of Optimizing Training Policies: Monotonic Inference Policies as the Real Objective for LLM Reinforcement Learning

This paper says AI models can improve during training but still get worse when used in real apps. The cause is a gap between the training system and the serving system. The authors propose MIPU, a two-step check that keeps only updates that improve real-use performance and stability in tests.

Securing the AI Agent: A Unified Framework for Multi-Layer Agent Red Teaming

AI-Infra-Guard is an open-source safety toolkit for testing AI agents. It checks four weak spots: the software setup, connected tools, agent behavior, and the model itself. It uses rule checks, AI-assisted reviews, and red teaming, meaning stress tests that try to break the system before attackers do in real use.

OmniOpt: Taxonomy, Geometry, and Benchmarking of Modern Optimizers

OmniOpt maps more than 100 optimizer methods, which are training rules that help AI models learn. The paper groups them by how they work and what they improve, such as speed, memory use, and stability. It also benchmarks them across tasks so researchers can choose methods more clearly with evidence.

📲SOCIAL MEDIA

🗞️MORE NEWS

Claude Cowork Goes Mobile and Web Anthropic is expanding its Claude Cowork feature to work seamlessly across mobile devices and web browsers. This means you can start a complex task like analyzing documents or drafting a report on your laptop and check its progress later on your phone. Cowork runs in the background even if your device is off, allowing the AI to handle long-running, multi-step work while you are away.

SpaceX/xAI + Cursor SpaceXAI and Cursor are preparing to launch their first shared AI model as soon as Wednesday, according to a memo reported by The Information. The model is built to answer quickly and could compete with top systems from Anthropic and OpenAI, but Reuters said it has not confirmed the report on its own.

TeraWulf + Anthropic TeraWulf signed a 20-year lease with Anthropic worth about $19 billion to build a large AI data center site in Hawesville, Kentucky. The site is expected to support 401 megawatts of IT power, with full capacity planned by early 2028.

China Considers Curbing Foreign Access to Its Top AI Beijing is discussing ways to restrict overseas access to China's most advanced AI models due to national security concerns. These talks were prompted in part by the rising global popularity of Chinese AI systems, which are becoming highly capable and cheaper to use than many Western alternatives. Officials are also considering tougher penalties for anyone who steals proprietary AI technology or leaks unreleased models.

Microsoft Swaps Out OpenAI for In-House Models To lower operating costs, Microsoft is quietly starting to replace models from OpenAI and Anthropic with its own internally developed AI models (called MAI) in several software applications like Excel and Outlook. While Microsoft will still use top-tier partner models for complex reasoning tasks, it is shifting toward cheaper, specialized in-house models to handle the massive volume of routine requests in its Office suite.

Midjourney Countersues Hollywood Over AI Use The AI image generator Midjourney is fighting back in a copyright lawsuit brought against it by major film studios like Disney and Warner Bros. Midjourney is demanding that the studios reveal how they are secretly using AI internally for things like storyboarding or idea generation. The company hopes to prove that downloading copyrighted material to train AI models is actually standard industry practice, even among the studios suing them.