AI Daily — May 12, 2026
2026-05-12
openai.com
OpenAI launches DeployCo to help businesses build around intelligence
OpenAI has spun up DeployCo, a dedicated enterprise deployment company aimed at helping organizations move frontier AI models from experimentation into production. The move signals OpenAI expanding beyond model development into professional services and implementation, competing more directly with system integrators and consulting firms in the enterprise AI space.
openai.com
How ChatGPT adoption broadened in early 2026
OpenAI's Q1 2026 usage data shows the fastest ChatGPT growth is now among users over 35, with gender usage becoming more balanced — indicators that AI assistant adoption is moving past early-adopter demographics into broader mainstream use. This demographic shift has practical implications for product design and safety considerations as the user base diversifies.
rss.arxiv.org
On Distinguishing Capability Elicitation from Capability Creation in Post-Training: A Free-Energy Perspective
This paper argues that the SFT-as-imitation vs. RL-as-discovery framing in post-training is too coarse, and proposes a more precise distinction: capability elicitation (reweighting behaviors already in the model's accessible support) versus capability creation (expanding what the model can practically reach). The authors operationalize this via the concept of 'accessible support' under finite compute budgets, offering a framework with direct implications for alignment and safety reasoning about what post-training actually changes in a model.
rss.arxiv.org
Where Reliability Lives in Vision-Language Models: A Mechanistic Study of Attention, Hidden States, and Causal Circuits
Across three open-weight VLM families (LLaVA-1.5, PaliGemma, Qwen2-VL), the paper finds that attention map sharpness is essentially uncorrelated with answer correctness (R_pb ≈ 0.001), debunking a common assumption used in VLM interpretability and uncertainty estimation. Hidden-state geometry, not attention structure, turns out to be a more predictive signal of model reliability, with implications for how practitioners should instrument VLMs for production deployment.