Meta Is Building "Hatch", Its First Paid AI Product, at Up to $200 Per Month

Meta is reportedly developing "Hatch", an AI agent product that automatically executes tasks like building tools, making reservations, and sending emails from natural-language instructions alone. A price point of up to $200 per month is reportedly under consideration.

CEO Mark Zuckerberg positions the product as a new revenue stream beyond advertising and a way to recoup the company's massive AI investments. For a company that has so far offered the Meta AI chatbot for free, this will be its first paid AI product.

OpenAI and Google are also accelerating their push into the AI agent market, and the growth potential of paid consumer-facing services is drawing attention.

xAI Trained Its Coding Model on Claude Outputs — and Reportedly Kept Going After Anthropic Cut Off Access

xAI, led by Elon Musk, reportedly trained its in-house coding model on outputs from Anthropic's Claude for several months.

According to reports, even after Anthropic blocked access, xAI continued accessing Claude through personal accounts and the Blackbox AI service. On top of that, xAI's pretraining team has shrunk to fewer than five people, with multiple leaders having departed, and the compute resources Musk stockpiled are now being rented out to Anthropic and Google rather than used to run the company's own models.

The episode has once again thrown the ethical and legal questions around data use between AI companies into sharp relief.

Sebastian Raschka Publishes a Systematic Roundup of LLM Research Papers from Early 2026

Sebastian Raschka, a prominent machine learning researcher, has published a large reference list organizing notable LLM research papers released between January and May 2026 by category.

The list spans ten categories, and he highlights the following trends as particularly noteworthy:

Hybrid architectures**: Nemotron 3 Super adopts a design that alternates attention layers with Mamba-2 layers. Qwen3.6 also uses a similar design with Gated DeltaNet layers. The main focus is making long-context processing more efficient

Mamba-3 and Gated DeltaNet-2**: New versions of these state space models have arrived, and adoption in next-generation LLMs is expected

Nemotron 3 Ultra**: A scaled-up version of the Super variant (550B-A55B) has also been released

Inference efficiency and KV caching**: With the growth of agent usage, efficient long-context processing has become the most critical challenge

Reinforcement learning, agent systems, and coding agents**: Research aimed at practical deployment is surging

Overall, the first half of 2026 shows a clear transition from the "scale things up" stage to the "get clever with architecture" stage.

Domino: Qwen3 Gets Up to 5.8x Faster with Speculative Decoding

The paper "Domino: Decoupling Causal Modeling from Autoregressive Drafting in Speculative Decoding" has been making the rounds in Reddit's LocalLLaMA community.

The method decouples causal modeling from autoregressive drafting, substantially improving the efficiency of speculative decoding. It is reported to achieve up to 5.8x throughput gains on Qwen3 models.

The paper and the code are available on arXiv and GitHub respectively, making this a practical result for practitioners interested in speeding up inference.

Meta Repeatedly Delays Developer Releases of New AI Models

According to a WSJ report, Meta has repeatedly delayed the release of new AI models it had planned for developers.

The reasons remain unclear, but the story has also gained traction on Hacker News, and concerns are growing about the impact on the open-source AI community. While the company focuses on monetizing Hatch, its strategy for releasing developer-facing models is looking increasingly opaque.

---