Anthropic Poaches an OpenAI Chip Engineer, New Insights into How LLMs Learn "Rare Skills", and Gemma 4 Runs Smoothly Without a GPU — AI News for June 7

Anthropic Poaches an OpenAI Chip Engineer — The Stakes Behind the IPO Race and In-House Chip Development

Anthropic has hired Clive Chan, the second hardware engineer to depart OpenAI's custom chip program. Chan brings experience from Tesla's Autopilot ASIC as well as the OpenAI–Broadcom partnership. With both companies heading toward IPOs, this is part of an escalating talent war, and Anthropic is reportedly also exploring the development of its own AI chips.

The move suggests that major AI companies are getting serious about bringing chip design in-house. The apparent aim is to reduce dependence on NVIDIA and optimize inference costs.

New Research Unravels How Large LLMs Acquire "Rare Skills"

A new study using models ranging from 4 million to 4 billion parameters has revealed why smaller models fail to learn rare tasks. The culprit: learning frequent tasks constantly overwrites the knowledge of rare tasks the model has already acquired. Notably, rather than scaling up the model, the problem can be mitigated simply by increasing the frequency of the target task in the training data — an instructive example of a practical, data-design-driven approach instead of an all-in bet on ever-larger models.

Gemma 4 26B-A4B Runs Without a GPU on a $150 Used PC

A report on Reddit's LocalLLaMA community has been drawing attention: Gemma 4 26B-A4B running at roughly 7 tokens per second on an aging GPU-less desktop (Core i5-8500 / 32GB RAM). The poster says they "were considering buying a GPU, but might not need one after all" — a real-world demonstration that the efficiency of MoE architectures holds up even in low-end environments.

At the same time, debate over how dense and MoE models tolerate quantization is heating up in the same community. One report says Gemma 4 26B A4B starts looping around the 45K context mark at Q5 quantization, with the issue resolved at Q6.

GitHub's CPO on the Future of AI Coding Agents

In an interview, GitHub CPO Mario Rodriguez discussed AI coding agents and "macro delegation," offering the outlook that workflows where developers direct the overall trajectory rather than the fine details of code will become mainstream.

Meanwhile in Japan, a 2026-edition comparison of Claude Code, Gemini CLI, and Kiro has been published on Qiita, drawing attention to the practical differences between the tools.