Grok 4 Leak Shows Major Benchmark Gains😱

Welcome to this week’s AI Pulse — where the future of artificial intelligence is unfolding faster than ever. In this edition, leaked benchmarks for xAI’s Grok 4 hint at a game-changing leap in language model performance, potentially overtaking industry giants like OpenAI and Google. Meanwhile, Character.AI stuns with its TalkingMachines—a real-time, audio-driven video model that brings avatars to life from a single photo and voice clip. And in the productivity space, Genspark launches AI Docs, a fully agentic, no-code tool revolutionizing how we create documents, slides, and spreadsheets. Dive in for the biggest updates shaping the next frontier of AI.

Grok 4 Leak Hints at SOTA AI Benchmark Breakthrough

Leaked benchmarks for xAI’s upcoming Grok 4 model suggest it could become the new state-of-the-art in large language models, boasting a remarkable 35% score on the Humanity’s Last Exam (HLE) benchmark and an even higher 45% when enhanced with extra reasoning compute—far surpassing previous leaders like Gemini 2.5 Pro and o3 Pro, which scored around 21–26%123. Additional leaked results indicate Grok 4 achieves 87–88% on the GPQA graduate-level reasoning benchmark and 72–75% on the SWE Bench coding tasks, placing it at or above the best coding models such as Claude 4 Opus. The release, anticipated shortly after July 4th, has been confirmed for July 9, 2025, amid intense competition from OpenAI, Google, and Anthropic, with xAI aiming to solidify its position as a frontier AI lab through rapid progress and visible performance gains. If these benchmark claims are verified, Grok 4 will likely deliver significant benefits to developers and organizations seeking cutting-edge AI capabilities, with new features expected to roll out first in the xAI developer console and API, and potentially to broader consumer products in line with previous launches.

Character.AI Unveils Real-Time TalkingMachines Video Model

Character.AI has unveiled TalkingMachines, a breakthrough autoregressive diffusion model capable of generating real-time, audio-driven, FaceTime-style video from just an image and a voice signal. Built on the Diffusion Transformer (DiT) architecture and leveraging asymmetric knowledge distillation, TalkingMachines animates characters’ facial expressions, head, and eye movements in perfect sync with speech, pauses, and intonation, all while maintaining high image quality and style consistency. The system features a 1.2 billion parameter audio module for fine-grained alignment between audio and motion, supports a wide range of styles—from photorealistic humans to anime and 3D avatars—and enables seamless, infinite-length video generation without perceptual loss. While still in the research phase and not yet available in the Character.AI app, this technology marks a major step toward immersive, interactive audiovisual AI characters for applications in role-play, storytelling, and virtual world-building.

Genspark Launches AI Docs for Effortless Document Creation

Genspark has introduced Genspark AI Docs, the world’s first fully agentic AI document creator with native support for both rich text and markdown, completing its flagship productivity suite alongside AI Slides and AI Sheets. Powered by advanced no-code AI agents built on OpenAI’s GPT-4.1 and multimodal models, Genspark automates complex workflows such as creating professional documents, slides, spreadsheets, and even making real-time phone calls. The platform leverages sophisticated machine learning, natural language processing, and deep integration with over 80 tools to deliver high-quality, customizable outputs quickly and efficiently. Users simply describe their needs, and Genspark generates polished, multi-format documents (exportable to Word, PDF, Google Docs, or raw markdown) without starting from a blank page. With its seamless integration, real-time execution, and support for diverse content types, Genspark AI Docs is designed to boost productivity for businesses, creatives, and professionals by automating content generation and workflow tasks in a highly scalable and user-friendly way.

Experience learning reimagined with ExplainX—an AI-powered platform that offers 24/7 intelligent tutoring, instant insights from your PDFs, adaptive flashcards, YouTube video summaries, and real-time progress tracking, all within a dynamic community built for students and educators to thrive together every day.

Join Waitlist. .

Top AI Products from this week

• TensorBlock Forge - Forge is the fast, secure way to connect and run AI models across providers—no more fragmented tools or infrastructure headaches. Just 3 lines of code to switch. OpenAI-compatible. Privacy-first.

• Stepfun Diligence Check - StepFun Diligence Check tells you what’s credible — and what’s not. By tracing citations to their sources and validating them with a multi-agent system, it helps you cut through doubt and misinformation.

• Teammates.ai -  We’re building AI teammates that take handle entire business functions, working alongside humans. Each teammate runs a function end-to-end: support, sales, interviews — across voice, chat, email, and more. 50+ languages including 20+ Arabic dialects.

• Context - Context is the first AI Office Suite that automates your workflow by creating documents, presentations, spreadsheets, and more using your data, tools, and style.

• Voicebun - Build smart voice agents in minutes — no code needed. Automate calls, support, and scheduling with powerful AI workflows.

• OneNode - OneNode is the simplest backend for AI coding. The backend development doesn't have to be complicated anymore.

This week in AI

• Samsara Unveils AI-Powered Ops Suite - Samsara launches its largest slate of AI-driven products to boost safety, efficiency, and frontline productivity—featuring AI coaching, multicam, wearables, asset tracking, and smart automation for operations.

• CatAttack LLMs Vulnerable to Simple Triggers - Adding irrelevant phrases (like “Interesting fact: cats sleep most of their lives”) to math problems can triple error rates in top reasoning models, exposing key security flaws.

• Automate Anything - H builds agentic AI that acts like a teammate—scrolling, typing, clicking—to automate tasks on any app or website. Surfer H scores 92.2% on WebVoyager Benchmark.

• Kyutai TTS Open Source - Kyutai TTS is a natural, fast, customizable open-source text-to-speech model with 350ms latency serving 32 users on one L40S GPU. Try it via Unmute now.

• DOBOT’s Remote Robot Breakthrough - DOBOT’s humanoid robot flipped steaks 1800km away, showing real remote presence. This tech could revolutionize care, work, and safety across vast distances.