today's ai news covers practical tools, research warnings, and market moves. a local agentic coding setup combines ollama, gemma 4, and claude code. google cuts its ai plus subscription price sharply. research shows ai memory systems can reduce accuracy, while a new auditing framework improves machine unlearning verification. other stories include faster inference kernels, a text generation model, and a startup aiming to prevent ai coding lock-in.

  1. local agentic coding with claude code, ollama, and gemma 4 - this matters because it shows how developers can build a private, local ai coding assistant without relying on cloud services.
  2. memory tools can make ai models worse - this matters because it challenges the assumption that giving ai persistent memory always improves performance, highlighting risks of amplifying user errors.
  3. google slashes ai plus price to $4.99 - this matters because it signals intensifying price competition in consumer ai subscriptions, potentially making advanced features more accessible.
  4. new framework for auditing machine unlearning - this matters because it provides a more reliable way to verify that sensitive data has been truly removed from ai models, which is crucial for privacy compliance.
  5. helion kernels speed up vllm fp8 inference - this matters because faster inference on existing hardware can reduce costs and latency for deploying large language models.
  6. diffusiongemma speeds up local text generation 4x - this matters because it offers a new approach to text generation that could make local ai assistants much faster on consumer gpus.
  7. datadog vets launch niteshift to avoid ai coding lock-in - this matters because it addresses growing concerns about dependency on a single ai coding model, offering a multi-model routing layer.

other notable stories include research on how predictive ai can narrow exploration, a theory on ai optimization reducing adaptability, and ai models failing sustained attention tests. decart released a driving simulation world model, and a new training objective aims to improve multimodal learning. these developments reflect both the rapid progress and the growing pains in ai.