Andrej Karpathy Joins Anthropic's Pre-training Team
**Andrej Karpathy**, one of the co-founders of OpenAI and former architect of Tesla's Autopilot, announced on his X account that he has joined **Anthropic's pre-training team**.
Pre-training is the stage where massive compute is poured into a model to equip it with core knowledge and capabilities, making it one of the most expensive steps in building frontier models. Karpathy has said that "the next few years will be a particularly formative period for the LLM frontier," and his choice of Anthropic over a return to OpenAI is a clear loss for his former employer.
Karpathy has long argued that reinforcement learning from human feedback (RLHF) has only limited effectiveness in training language models, so his focus on foundational model training is consistent with his past stance.
Microsoft Releases Computer-Use-Focused Model "Fara-7B"
Microsoft Research has released **"Fara-7B"**, an agentic small language model specialized for Computer Use. Offered under the MIT license, the model has just **7 billion parameters** yet reportedly achieves performance on par with agent systems that require far larger resources.
Architecturally, it is a multimodal decoder-only language model that takes screenshot images and text context as input, then outputs actions as tool calls after a chain-of-thought reasoning process. The baseline is Qwen 2.5-VL (7B), with a context length of 128K tokens.
Training used 64 H100 GPUs and took only 2.5 days. It is a static model whose training was completed in late October 2025, and it has recorded more than 14,000 downloads on Hugging Face.
Former OpenAI Employees Warn of xAI Safety Issues, Potentially Affecting SpaceX IPO
Former OpenAI employees have founded a new AI watchdog organization and are warning that **xAI has an insufficient track record on safety**. They argue that investors deserve more information about the safety practices of the Elon Musk-led xAI, which could affect SpaceX's planned initial public offering (IPO).
The former employees say xAI falls short of industry standards on safety, highlighting a new question: how the safety governance of frontier AI companies intersects with investor accountability at a publicly listed company.
Hugging Face Releases DNA Foundation Model "Carbon"
Loubna at Hugging Face has announced the release of **"Carbon"**, an open DNA foundation model. Carbon is a family of models aimed at decoding the language of life, with applications in understanding and analyzing genomic sequences.
Its open-source release could significantly contribute to democratizing AI in bioinformatics. Highly practical tools continue to emerge at the intersection of life science and AI.
Slopinator: A Tool for Poisoning AI Training Data Emerges
A tool called **"Slopinator"** has been published, designed to sneak fake code into GitHub repositories and poison the training data of AI models.
The attack exploits the fact that AI companies scrape public code from GitHub for training data, with poisoned data degrading model behavior. The provenance and trustworthiness of AI training data sources are once again emerging as a security concern.
AI Graduation Announcer Mispronounces and Skips Names
An AI announcer introduced at a graduation ceremony in the United States mispronounced graduates' names and even skipped some of them. The Verge reported on the incident at Glendale Community College.
The case illustrates that AI deployment in public settings still faces reliability challenges, and some argue that automation—especially in moments tied to personal dignity, such as the reading of names—calls for greater care.
Anthropic Tops CNBC's "Disruptor 50"
**Anthropic** has taken the top spot in CNBC's 2026 "Disruptor 50" startup ranking. The ranking evaluates companies that are transforming their industries, and Anthropic was reportedly recognized for its safety-focused approach and the market penetration of the Claude series.