Research News & Updates

Generative AI introduces unique forms of technical debt, including tool sprawl, prompt stuffing, opaque pipelines, inadequate human feedback systems, and insufficient stakeholder engagement, requiring new development practices to effectively manage and pay down these debts.

Logical Intelligence, led by Eve Bodnia, has developed an energy-based reasoning model (EBM) called Kona 1.0, which can solve sudoku puzzles faster than leading large language models (LLMs).

Shayne Longpre and Sayna Ebrahimi introduce ATLAS, adaptive transfer scaling laws designed to optimize the scaling of multilingual language models.

Researchers have demonstrated that attackers can manipulate vision language models (VLMs) by generating adversarial images, exploiting vulnerabilities in these deep learning architectures.

Researchers unlock GPT-OSS for agentic reinforcement learning training, addressing challenges and improving performance through a series of practical engineering fixes.

Our study shows that relying on AI assistance can hinder the acquisition of new coding skills, as it reduces comprehension and understanding of the code being written.

Researchers at Google's DeepMind developed AlphaGenome, an AI model that quickly unravels the 'dark genome', a crucial part of DNA with a significant role in disease discovery.

A recent study found that a small fraction of AI conversations exhibit disempowerment potential, where users' autonomy is compromised, although severe disempowerment occurs rarely.

Researchers at Google Research identified that multi-agent coordination can dramatically improve performance on parallelizable tasks but degraded performance on sequential ones. A predictive model correctly identifies the optimal coordination strategy for 87% of unseen task configurations.

A recent study, featured in Wired, mathematically proves that large language models have limitations in complexity and accuracy.