AI Digest — Jul 20, 2026 (Morning)

Researchers introduced ActiveVision, a benchmark to measure active observation in multimodal large language models (MLLMs). The benchmark consists of 17 tasks that require repeated visual perception. Top MLLMs, including GPT-5.5 and Claude Fable 5, performed poorly on these tasks, with the highest-scoring model solving only 10.6% of items. In contrast, human participants averaged 96.1% accuracy. The results suggest that current MLLMs lack robust active visual observation, highlighting the need for new architectures and training objectives. The benchmark's findings have implications for the development of more advanced MLLMs that can effectively integrate perception and reasoning.

A recent study explores the relationship between pretraining and reinforcement learning (RL) in large language models (LLMs), using chess as a controlled testbed. The researchers pretrained language models of varying sizes on human chess games, fine-tuned them on synthetic reasoning traces, and applied RL to chess puzzles. They found that post-RL performance is predictable from pretraining loss and that RL improves the model's policy in a way that depends on puzzle difficulty. The study also demonstrated that these findings transfer to other domains, such as math. The results provide insights into the pretraining-to-RL interface and the science of reasoning in LLMs.

A developer replaced a $120,000 bowling center system with a network of ESP32 microcontrollers at a cost of $1,600. The ESP32s were used to control various aspects of the bowling center, including scoring, lighting, and machinery. This project demonstrates the potential for affordable, open-source hardware to disrupt traditional industrial control systems. The use of ESP32s in this context highlights their versatility and capabilities in IoT applications.

A study found that people who received advice from AI were less accurate in their answers but more confident in their responses. The research suggests that AI advice can suppress critical thinking, leading to incorrect conclusions. This study involved participants who were given AI-generated answers to questions and then asked to provide their own responses. The results indicate that over-reliance on AI advice can be detrimental to decision-making. The study's findings have implications for the development of AI systems that provide advice or guidance to humans.

Sam Altman, in an email to OpenAI's board, discussed plans to create and release a language model similar to GPT-3 that can run on consumer hardware. The goal is to discourage others from releasing similar models and make it harder for new efforts to get funded. This move is part of OpenAI's open-source strategy discussions. The email was exposed in the Musk v. Altman case in 2026. The plan aims to preempt Stability or others from releasing comparable models.

Alibaba has introduced Triton, a programming language designed for its SAIL (Self-Adaptive Intelligent Learning) framework. Triton aims to simplify the development of AI models by providing a high-level abstraction. The language is now available on GitHub, allowing developers to explore and contribute to the project. This development is significant for the AI community as it indicates Alibaba's continued investment in AI research and development. The release of Triton may also influence the development of future AI frameworks and languages.

Researchers introduced Loopie, a looped Transformer model that addresses the challenge of increasing parameter count versus looping a model. Loopie consists of two Mixture-of-Experts models with 20B and 6B parameters. Extensive studies show Loopie outperforms vanilla Transformer baselines with the same compute budget. Loopie achieves strong reasoning abilities through a novel post-training pipeline and gold-medal performance at the 2025 IMO and IPhO. This development is significant in the field of natural language processing.