# Feb 20, 2026

## ML Times    Feb 20, 2026

- **Gemini 3.1 Pro** enhances AI capabilities for complex tasks, achieving a verified score of **77.1%** on the ARC-AGI-2 benchmark, which evaluates novel logic pattern solutions.

- **AI's potential is stifled by high latency and costs**, with current models requiring extensive infrastructure and resources, yet Taalas aims to revolutionize this by creating custom silicon that drastically improves performance and efficiency.

- **Nvidia** is set to finalize a **$30bn investment** in OpenAI, replacing a previously planned **$100bn deal**, as part of a broader funding round expected to exceed **$100bn** and value OpenAI at **$730bn**.

- **AI summarization tools can subtly manipulate outputs** through customized policies, as demonstrated by the Bilingual Shadow Reasoning technique, which can bypass safety guardrails while appearing neutral.

- **Consistency diffusion language models (CDLM) achieve up to 14.5x faster inference** by integrating multi-token finalization with block-wise KV caching, enhancing efficiency in math and coding tasks.

- **AI agents, like Claude Code, are increasingly operating autonomously, with session durations nearly doubling from under 25 minutes to over 45 minutes in just three months, indicating a growing trust and capability in real-world applications.**

- **Tabular data competitions** are evolving, with **AutoML** and **tabular foundation models** like **TabPFN** emerging alongside traditional **GBDTs** such as **XGBoost** and **LightGBM** in winning solutions.

- **Claude Code Security** introduces a novel approach to cybersecurity by scanning codebases for vulnerabilities and suggesting targeted patches, enhancing the ability of teams to address complex security issues that traditional tools often overlook.

- **Minions** are Stripe's innovative **end-to-end coding agents**, generating over **1,000 pull requests** weekly, with human oversight ensuring quality control.

- **Fast context compaction** via Attention Matching enables the construction of compact key-value caches that maintain performance while achieving up to **50x compaction** in seconds with minimal quality loss.

- The **SoftDTW-CUDA for PyTorch** package offers a **GPU-accelerated** and **memory-efficient** implementation of Soft Dynamic Time Warping, achieving **~67× faster** performance and **~98% lower GPU memory** usage compared to existing methods.

- **Vision-Language Models (VLMs) achieve ~84% F1 on text-rendered binary grids but drop to 29-39% F1 on filled squares**, revealing a significant performance gap that highlights their reliance on textual cues for spatial reasoning.

- **Wizwand v2** enhances dataset consistency and task granularity by utilizing **LLM** for accurate dataset descriptions, significantly reducing nonsensical comparisons.

- **Key finding:** The **Cheap Anchor** score predicts edge importance in GPT-2's induction circuit with a **Spearman correlation of ρ=0.623**, achieving a **125x speedup** over traditional methods by relying solely on weight structure without requiring model runs or training data.

- **Gemini 3.1 Pro** enhances AI capabilities for complex tasks, achieving a verified score of **77.1%** on the ARC-AGI-2 benchmark, more than double that of its predecessor.