Jan 23, 2026

ML Times

Daily

Weekly

• Qwen3-TTS family is now open sourced: Voice design, clone, and generation

  • Qwen operates across multiple domains, including qwen.ai and chat.qwenlm.ai, indicating a robust infrastructure for AI-driven applications and services. Link to article

• GPTZero finds 100 new hallucinations in NeurIPS 2025 accepted papers

  • GPTZero's analysis of 4841 NeurIPS 2025 papers revealed at least 100 confirmed hallucinations, highlighting a significant oversight in the peer review process that failed to catch these errors despite rigorous evaluations by multiple reviewers.

• 100 Hallucinated Citations Found in 51 Accepted Papers at NeurIPS 2025

  • 100 hallucinated citations were identified in 51 accepted papers at NeurIPS 2025, raising concerns about the integrity of peer-reviewed research.

• Show HN: Text-to-video model from scratch (2 brothers, 2 years, 2B params)

  • Linum v2 is a 2B text-to-video model that enhances video generation capabilities, allowing for more dynamic and engaging content creation.

• Keeping 20k GPUs healthy

  • Modal operates a robust GPU worker pool exceeding 20,000 GPUs, employing a comprehensive reliability system that includes instance type testing, machine image preparation, and both passive and active health checks to ensure optimal performance across various cloud providers.

• Floating-Point Printing and Parsing Can Be Simple and Fast

  • Fast unrounded scaling enables efficient floating-point printing and parsing, achieving speeds surpassing existing algorithms like Dragon4 and Grisu3, with a Go implementation expected in Go 1.27 by August 2026.

• I solved CartPole-v1 using only bitwise ops with Differentiable Logic Synthesis

  • The author achieved a perfect score on CartPole-v1 by distilling the control policy into four bitwise rules using Differentiable Logic Synthesis, treating the state as a boolean input vector rather than numerical values.

• LLM-in-Sandbox Elicits General Agentic Intelligence

  • LLM-in-Sandbox enables LLMs to explore a code sandbox, demonstrating generalization to non-code tasks without additional training, such as accessing external resources and executing scripts for formatting.

• Launch HN: Constellation Space (YC W26) – AI for satellite mission assurance

  • Constellation Space has developed an AI system that predicts satellite link failures 3-5 minutes in advance with over 90% accuracy, enabling proactive traffic rerouting to prevent data loss.

• Waypoint-1: Real-Time Interactive Video Diffusion from Overworld

  • Waypoint-1 is a real-time interactive video diffusion model that allows users to control and prompt video generation through text and inputs, creating immersive experiences with zero latency.

• Flying with Photons: Rendering Novel Views of Propagating Light

  • Novel imaging technique captures light propagation with picosecond-level resolution, enabling the synthesis of videos from multiple viewpoints, showcasing effects like scattering and reflection.

• Teacher-Free Self-Distillation: Fixing the Softmax "Infinite Gap" with Euclidean alignment

  • Teacher-Free Self-Distillation (TFSD) addresses the "Infinite Gap" problem in Cross-Entropy loss by employing negative squared Euclidean distance, which bounds logits and prevents infinite values during optimization.

• Agentic Uncertainty Quantification

  • The Dual-Process Agentic UQ (AUQ) framework innovatively transforms verbalized uncertainty into active control signals, addressing the limitations of traditional uncertainty quantification methods that merely diagnose risks without resolution.

• Unrolling the Codex agent loop

  • From Pilot to Profit: Survey Reveals the Financial Services Industry Is Doubling Down on AI Investment and Open Source

  • Financial institutions are significantly increasing AI investments, with nearly 100% planning to maintain or boost budgets, driven by a clear return on investment from AI applications like fraud detection and risk management.