engblogs

Scaling Categorical Flow Maps demonstrates training a 1.7B base flow on 2.1T tokens to self-distill into CFMs for few-step, high-quality text generation, while introducing a semi-discrete likelihood bound and LM benchmarking.

Performance characterization comparing Diffusion Language Models (DLMs) and Autoregressive Language Models (ARMs), revealing that DLMs gain arithmetic intensity through parallelism but struggle with long-context scaling, with block-wise decoding and reduced sampling steps as key routes to competitive latency and throughput versus ARMs.

Scaling Categorical Flow Maps demonstrates training a 1.7B-parameter base flow on 2.1T tokens to enable diverse, high-quality text generation in as few as four inference steps using CFMs, while introducing a semi-discrete likelihood bound and practical guidance on loss weighting and time scheduling for large-scale diffusion-flow models.

An in-depth look at LM's object model for representing JavaScript objects, arrays, and closures inside an Automerge-backed persistent heap, using a transpiler and proxies to serialize, deserialize, and share state for multi-user collaboration while preserving invariants and performance.

Data-driven analysis of how ChatGPT moves from asking to doing across the world, revealing work-focused adoption, rising multimedia use, and demographic shifts via OpenAI Signals.

Collaboration between OpenAI and the American Psychological Association to align AI development and use with psychology-grounded, evidence-informed safeguards for youth well-being, emphasizing developmentally appropriate design and practical resources for families, clinicians, and schools.

A technical overview of GPT-5.6 Sol improvements in ChatGPT, the expansion of GPT-5.6 Luna to free users with unlimited chats, and the new Think button and reasoning slider for higher-quality, safer outputs.

From asking to doing: a data-driven look at how ChatGPT is moving from simple queries to task-oriented work across the globe, fueled by country-by-country adoption and rising multimedia use.

OpenAI collaborates with the American Psychological Association to bring psychological science into responsible AI for youth, outlining evidence-based safeguards, developmentally appropriate design, and tools for families, clinicians, and schools.

A data-driven look at how ChatGPT is moving from answering to doing tasks worldwide, with country-by-country adoption, rising multimedia use, and demographic shifts shaping its deployment.

Details the rollout of GPT-5.6 Sol improvements in ChatGPT for Plus/Pro users and the expansion of GPT-5.6 Luna to Free users, introducing a Think button and a response-depth slider to balance speed, accuracy, and deeper reasoning.

OpenAI partners with the American Psychological Association to align youth AI use with evidence-based psychology, building safeguards, practical tools for families and clinicians, and developmentally appropriate design.