AI Daily — May 13, 2026
2026-05-13
openai.com
How NVIDIA engineers and researchers build with Codex
OpenAI details how NVIDIA teams are using Codex paired with GPT-5.5 to ship production systems and accelerate research-to-experiment pipelines. The case study is notable as a signal of Codex's positioning as a serious engineering tool at one of the most infrastructure-heavy AI companies, not just a developer toy. It implies GPT-5.5 is now in active production use at NVIDIA beyond API experimentation.
openai.com
What Parameter Golf taught us about AI-assisted research
OpenAI's Parameter Golf competition drew 1,000+ participants and 2,000+ submissions, tasking entrants with ML research, coding agents, quantization, and novel model design under strict parameter-count constraints. The writeup shares findings about where AI-assisted research workflows succeeded and failed in a structured competitive setting. This is one of the more rigorous public evals of AI-in-the-loop ML research conducted at scale.
rss.arxiv.org
The Bicameral Model: Bidirectional Hidden-State Coupling Between Parallel Language Models
Researchers propose coupling two frozen pretrained LLMs through a trainable neural interface on their intermediate hidden states, enabling continuous bidirectional communication rather than serialized text exchange. A learned suppression gate (~1% of combined parameters) selectively controls what each model shares with the other at every generation step, trained end-to-end from task loss alone. This is a structurally novel alternative to tool-use and multi-agent frameworks that avoids the information bottleneck of token-level communication.
rss.arxiv.org
Steering Without Breaking: Mechanistically Informed Interventions for Discrete Diffusion Language Models
This paper applies sparse autoencoders to four discrete diffusion LMs (124M–8B parameters) to analyze when different attributes commit during the denoising process, finding that topic locks in within the first 2% of steps while sentiment unfolds over ~20%. Uniform intervention schedules borrowed from autoregressive controlled generation cause compounding quality degradation in DLMs, especially under multi-attribute steering. The work provides a mechanistic foundation for designing attribute-specific intervention schedules tailored to diffusion-based text generation.