ML Times

Jul 18, 2026

Daily

Weekly

• GPT-5.6 Sol Pro successfully closed a 30-year gap in convex optimization by proving a lower bound that had eluded researchers since 1996, verified in Lean after a focused 148-minute session.

• Open-source AI is now a dominant force, with open-weight models accounting for over 50% of production tokens by mid-2026, reflecting a significant shift in developer preference towards cost-effective and accessible solutions.

• Kimi K3, with 2.8 trillion parameters, is hailed as the first "open 3T-class model," outperforming many competitors in benchmarks while promising an open weight release by July 27, 2026.

• Vāgdhenu is a vṛtta-aware text-to-speech system that accurately renders Sanskrit verses into chant, utilizing a retrained neural vocoder and a meticulously designed chant corpus of approximately 5 hours.

• Kimi K3 matches Claude in performance while offering significantly lower costs, with API rates at $3 per million input tokens compared to Claude's $10, highlighting a shift in the competitive landscape of AI models.

• AI auditor zkao identified a critical soundness bug in OpenVM's openvm-pairing library, allowing malicious actors to forge pairing equality, which undermines the integrity of cryptographic proofs built on this library.

• Google DeepMind's initiative, AlphaGenome, leverages AI to enhance our understanding of the genome, aiming to unlock new avenues in bioresilience and health innovation. Learn more.

• Stereo2Spatial is a model that transforms stereo music into spatialized binaural mixes, utilizing a flow-matching diffusion model in the latent space of a VAE, which enhances audio quality by modeling raw waveforms instead of latent representations.

• NVIDIA's Vera Rubin platform enhances post-training efficiency, achieving the lowest cost per token through extreme codesign, which is crucial for maximizing intelligence per dollar in agentic AI applications.

• NVIDIA NeMo Automodel and 🤗 Diffusers enable efficient fine-tuning of video and image models at scale, enhancing performance and adaptability across various applications.

• The Isomorphic Labs Drug Design Engine (IsoDDE) significantly enhances predictive accuracy in drug design, outperforming AlphaFold 3 by more than doubling its accuracy on challenging protein-ligand benchmarks and predicting binding affinities faster and more accurately than traditional methods.

• GPT-2 Small’s embedding geometry around "Trump": discretized vs. continuous nearest neighbours [P]