Who is leading xAI now?

xAI’s Original Founding Team Is Gone

All remaining co-founders of xAI have reportedly departed, marking a complete exit from the company’s original founding team. It is a notable leadership reset for one of the most closely watched AI labs.

Brett Adcock Launches Hark With $100M Bet

Brett Adcock has come out of stealth with Hark, a new AI lab backed by $100 million of his own money. The company is taking a full-stack approach, building foundation models, software, and purpose-built hardware from scratch. Its team includes talent from Apple, Meta, Google, Tesla, and Amazon, with models slated for this summer and dedicated devices to follow shortly after.

• Hark is starting with $100 million from Adcock’s own money.
• The company is building models, software, and hardware as one stack from day one.
• It plans to release first models this summer and devices shortly after.

Google’s Agent Smith Gets So Popular Access Tightens

Google’s internal AI tool Agent Smith has become popular enough that access is being restricted. The company is also tying adoption to performance reviews and standardizing AI use across teams. At the same time, NotebookLM, Gemini for Business, and a new Skills framework suggest Google is pushing a broader AI operating layer across products and workflows.

• Demand for Agent Smith has led to restricted access as employees rush to use it.
• Google is reportedly tying AI use to performance reviews to drive adoption across teams.
• The broader rollout includes NotebookLM and Gemini updates plus a new Skills framework.

Meta’s Brain Model Outperforms Real fMRI

Meta has open-sourced TRIBE v2, a model trained on brain scans from more than 700 people that simulates neural activity across vision, hearing, and language. The synthetic predictions reportedly outperform real fMRI recordings in some settings, and the team says the model can reproduce decades of neuroscience findings in software. Meta also released the code, weights, and a live demo for researchers.

• TRIBE v2 was trained on 700+ people and more than 1,000 hours of brain data.
• Its synthetic predictions reportedly outperform real fMRI in some population-level comparisons.
• Meta also released code, weights, and a live demo for researchers to use.

MetaClaw Tries To Teach Agents On The Fly

MetaClaw is designed to upgrade AI agents while they are in use. When an agent fails, a separate model extracts a compact rule from the error and injects it into the system prompt, while reinforcement learning with LoRA fine-tuning updates weights during idle windows. The results are based on simulations, but the reported gains are substantial.

• When an agent fails, MetaClaw extracts a compact rule from the error and adds it to the prompt.
• It also uses idle windows to update model weights with LoRA fine-tuning.
• The reported gains come from simulation tests rather than real-world deployment.

Chroma’s Context-1 Targets Faster Agentic Search

Context-1 is Chroma’s 20B parameter agentic search model, trained on more than 8,000 synthetically generated tasks. It aims to match frontier retrieval performance at much lower cost and with up to 10 times faster inference. The system separates search from generation and can iteratively refine sub-queries across multiple turns.

• The model has 20B parameters and was trained on more than 8,000 synthetic tasks.
• Chroma says it can reach frontier-level retrieval at a fraction of the cost.
• Its design separates search from generation and iterates across multiple turns.