# Jul 19, 2026ML Times

### 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.

- **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.

- **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](https://deepmind.google/blog/alphagenome-ai-for-better-understanding-the-genome/).

- **Researching tokenomics** revealed that using multiple AI models in a shared memory setup can significantly extend research duration without additional costs, achieving roughly 10 times longer operational time.

- **Harness engineering** enhances agent output by optimizing the **context** and **tools** surrounding it, allowing for a **100x improvement** in performance as noted by Ryan Lopopolo.

- **UnifiedIR** aims to consolidate various Intermediate Representation (IR) data structures in Julia into a single, extensible framework, enhancing usability and integration across the ecosystem.

- **Kurt Gödel's** groundbreaking work revealed that **mathematics cannot fully explain itself**, demonstrating that any sufficiently complex system will contain truths it cannot prove, a concept that remains crucial in today's AI landscape.

- **GFX1250**, AMD's new accelerator chip for machine learning, features significant enhancements like **1024 addressable VGPRs** per wave and **320kB of local memory**, marking a substantial leap from previous architectures.

- The **Muon optimizer**, designed for deep learning, operates in function space rather than parameter space, which can lead to instability when applied to Transformer Q and K weight matrices, causing training to collapse.

- **Post-training techniques** like supervised fine-tuning (SFT) and reinforcement learning with verifiable rewards (RLVR) significantly improved the accuracy of open-weight LLMs on the Swedish Medical Licensing Exam, with MedGemma-1.5-4B achieving **60.8%** accuracy after SFT, sufficient to pass the exam.

- **Deep learning** significantly enhances **single-cell RNA sequencing (scRNA-seq)** analysis, with a survey detailing **25 methods** across **6 subcategories** that improve data interpretation and biological insights.

- **GPT-2 Small’s** embedding geometry around “Trump”: discretized vs. continuous nearest neighbours.