In partnership with

Good morning. It’s Monday, July 14th.

On this day in tech history: In 1983, Nintendo released the arcade game Mario Bros. in Japan. While a video game, its underlying logic and pathfinding algorithms for character movement influenced early AI research in game theory and agent-based systems, precursors to reinforcement learning techniques.

In today’s email:

• China’s Free Trillion-Parameter Open Model

• OpenAI Delays Open-Source Release

• xAI Fixing Grok 4’s Bias

• 5 New AI Tools

• Latest AI Research Papers

You read. We listen. Let us know what you think by replying to this email.

China’s Kimi K2 outperforms GPT-4.1 on code and math, and it’s free

Chinese startup Moonshot AI just dropped Kimi K2, a trillion‑parameter open‑source model built to handle coding, tool use, and full-on autonomous workflows, and it’s already beating GPT‑4.1 on big benchmarks. Kimi K2 hits 65.8% on SWE‑bench Verified, 53.7% on LiveCodeBench, and a huge 97.4% on MATH‑500. It runs a mixture‑of‑experts setup with 32 billion active parameters, plus a custom MuonClip optimizer that keeps training stable and cost‑effective at scale.

It’s not just about raw power either. Moonshot’s pairing the free open release with cheap API access to undercut big players and get enterprises on board fast. The focus on real agentic behavior, executing multi-step tasks autonomously, cues the shift from conversational chatbots to practical AI assistants. Read more.

OpenAI’s cautious pause meets Google’s strategic play for AI coding talent

OpenAI has indefinitely delayed the launch of its highly anticipated open AI model, originally slated for this summer. The release would have given developers free access to model weights with reasoning capabilities comparable to OpenAI’s o-series models. CEO Sam Altman cited the irreversible nature of publishing weights and the need for deeper safety reviews before making them public.

At the same time, OpenAI’s separate effort to acquire AI coding startup Windsurf for around $3 billion fell through, reportedly due to complications tied to its partnership with Microsoft. Seizing the opening, Google DeepMind secured Windsurf CEO Varun Mohan, cofounder Douglas Chen, and key R&D engineers in a talent-driven deal valued at roughly $2.4 billion over several years.

While Windsurf remains independent under interim CEO Jeff Wang, the departing team will lead “agentic coding” initiatives to prop up Google’s Gemini models. Read more.

xAI says it wants to fix Grok 4 because referencing Musk's views is not right for a truth-seeking AI

The Wall Street Journal reports SpaceX may invest $2 billion in Elon Musk’s AI startup xAI, as part of a planned $10 billion equity and debt raise arranged by Morgan Stanley. This would mark SpaceX’s first direct investment into xAI, which already supplies its Grok chatbot to support Starlink customer service. Earlier this year, xAI merged operationally with Musk’s social platform 𝕏. However, Musk responded saying, these rumors are false.

xAI also acknowledged technical and design flaws in Grok 4 after independent testing revealed the chatbot routinely referenced Musk’s personal posts on hot-button issues like abortion and the Israel-Palestine conflict, despite system prompts instructing it to seek a diverse set of sources.

xAI admitted this contradicts its goal of developing a truth-seeking AI and promised a fix to the underlying model. Read more.

• Meta acquires voice startup Play AI

• Deepmind’s GenAI processors Promise Real‑Time, Plug‑and‑Play AI Pipelines

• New open source coding models from Mistral target practical AI agents

• Anthropic tests API Playground and Claude Code analytics in Workbench

• Goldman Sachs is piloting its first autonomous coder in major AI milestone for Wall Street

• AI startups believe Google’s Chrome is vulnerable to a new wave of intelligent browsers

• Turns out, aligning LLMs to be "helpful" via human feedback actually teaches them to bullshit

• Hugging Face's new robot is the Seinfeld of AI devices

• EU's Model Documentation Form makes AI providers explain their models like it's tax season

• Energy-Based Transformers aim to teach AI to really think, trading brute force for iterative, System 2‑style reasoning

• Reka.ai Launches Flash 3.1 and Reka Quant for Multimodal AI Solutions

• Exists opens public access to AI-powered 3D world creator

  

  5 new AI-powered tools from around the web

📄 Traceable Evidence Enhanced Visual Grounded Reasoning: Evaluation and Methodology

📄 Test-Time Scaling with Reflective Generative Model ​​

📄 CLiFT: Compressive Light-Field Tokens for Compute-Efficient and Adaptive Neural Rendering

📄 NeuralOS: Towards Simulating Operating Systems via Neural Generative Models

📄 T-LoRA: Single Image Diffusion Model Customization Without Overfitting

Thank you for reading today’s edition.

Your feedback is valuable. Respond to this email and tell us how you think we could add more value to this newsletter.

Interested in reaching smart readers like you? To become an AI Breakfast sponsor, reply to this email or DM us on 𝕏!