This week saw a mix of breakthroughs and disputes across AI and computing. IBM demonstrated a quantum computer solving a problem beyond classical reach, while music publishers sued Anthropic over training data. Google released new models for speech and geospatial work, and researchers found a prompt injection flaw in Claude Code. The week also brought advances in model efficiency and evaluation.

  1. IBM quantum computer solves classically intractable problem in 15 minutes - IBM and University of Chicago researchers used 70 error-corrected logical qubits to complete a quantum computation that leading classical methods could not practically reproduce. The problem involved simulating a quantum system, and the result suggests error-corrected quantum computers are approaching useful tasks. This matters because it marks a step toward quantum advantage in real-world applications.
  2. Sony and Warner sue Anthropic over music copyright - Major music publishers accuse Anthropic of torrenting and scraping copyrighted works to train Claude, seeking damages and an injunction. The lawsuit claims Anthropic used pirated lyrics and recordings without permission. This case could set a precedent for how AI companies source training data and may influence future copyright law.
  3. Claude Code auto mode bypassed by prompt injection - Researcher Johann Rehberger found a prompt injection attack that bypasses Claude Code's auto mode safety classifier 80% of the time. The attack tricks the model into executing harmful commands by embedding instructions in files or web content. This highlights ongoing security challenges in AI coding assistants and the need for stronger safeguards.
  4. Google launches Gemini 3.5 Transcribe for real-time speech-to-text - Google DeepMind introduced Gemini 3.5 Transcribe, a speech-to-text model with smart formatting, custom vocabulary, and low word error rates for developers. It supports multiple languages and can handle noisy audio. This release improves accessibility and transcription workflows for businesses and developers.
  5. 4-bit model beats its full-precision source after quantization-aware healing - A new healing method lets a compressed, 4-bit model outperform its bfloat16 original on most benchmarks. The technique repairs quantization errors during fine-tuning, recovering lost accuracy. This could make large models cheaper to deploy without sacrificing performance, benefiting edge devices and cloud services.
  6. Double-blind AI evaluations keep benchmarks secret - Google DeepMind pilots a cryptographic method to evaluate AI models without exposing test questions or model weights. The approach uses secure multi-party computation to score models privately. This addresses benchmark contamination and could lead to more trustworthy model comparisons.

The strongest shared signal is the push toward more capable and efficient AI systems, alongside growing legal and security scrutiny. Quantum computing reached a milestone, while copyright lawsuits and prompt injection attacks show the field's unresolved tensions. Advances in quantization and private evaluation point to a focus on practical deployment and trust.