May 16, 2026

Daily

Key Updates

• arXiv implements 1-year ban for papers containing incontrovertible evidence of unchecked LLM-generated errors, such as hallucinated references or results.
  arXiv has instituted a 1-year ban for authors whose papers contain incontrovertible evidence of unchecked LLM-generated errors, emphasizing the authors' responsibility for all content, regardless of its origin.

• Orthrus-Qwen3: up to 7.8×tokens/forward on Qwen3, identical output distribution
  Orthrus combines the exact generation fidelity of autoregressive LLMs with the high-speed parallel token generation of diffusion models, achieving up to a 7.8× speedup in inference tasks while ensuring strictly lossless generation.

• Europe built sovereign clouds to escape US control. Forgot about the processors
  Europe's €2 billion investment in sovereign cloud initiatives aims to reduce reliance on US technology, yet most cloud operators still depend on American-made Intel and AMD processors, which harbor unmonitored management engines that could compromise digital sovereignty.

• Frontier AI has broken the open CTF format
  Frontier AI has fundamentally altered the CTF landscape, rendering traditional skill measurement obsolete as AI tools can now solve many challenges with minimal human input, leading to a shift in competition dynamics.

• Δ-Mem: Efficient Online Memory for Large Language Models
  $δ$-mem introduces a lightweight memory mechanism that enhances large language models by integrating a compact online state, allowing for efficient historical information reuse without the need for costly context window expansions.

• DeepSeek-V4-Flash means LLM steering is interesting again
  DeepSeek-V4-Flash revitalizes interest in LLM steering, enabling engineers to manipulate model outputs by adjusting internal activations, making it accessible for local models to compete with frontier agents.

• Kioxia and Dell cram 10 PB into slim 2RU server
  Kioxia's LC9 SSDs enable Dell to create a 10 PB all-flash storage server in a compact 2 RU form factor, enhancing storage density for enterprise applications.

• After 8 years, I rewrote my open-source PyTorch curvature library
  The hessian-eigenthings module enables efficient eigendecomposition of the Hessian and other curvature matrices for PyTorch models, utilizing methods like Lanczos and stochastic power iteration to overcome memory limitations associated with full Hessian computation.

• Do you agree with Judea that learning from data is not everything?
  Judea Pearl asserts that there are mathematical limits to learning from data alone, emphasizing that correlation does not equate to causation, which is often misunderstood in machine learning.

• I broke AppLovin's mediation cipher protocol
  I decrypted AppLovin's mediation cipher, revealing that it transmits enough device data to uniquely identify an iPhone across apps, even when users deny ATT, undermining the assumption that ATT is the sole method for user identification.

• Orthrus: Memory-Efficient Parallel Token Generation via Dual-View Diffusion
  Orthrus introduces a trainable diffusion attention module in a frozen AR Transformer, achieving up to 7.8× token processing speed and maintaining accuracy comparable to Qwen3-8B.

• PINN is predicting trivial solution for stiff ODE
  PINN struggles with stiff ODEs, particularly when the stiffness parameter (k) exceeds 50, leading to a trivial solution prediction despite various adjustments in training parameters.

• Further Notes on Our Recent Research on AI Delegation and Long-Horizon Reliability
  AI delegation in workflows can lead to 19–34% degradation in artifact fidelity over 20 iterations, highlighting the need for robust evaluation methods in long-horizon tasks.