Daily Brief

OpenAI's ChatGPT advertising business has reached a $1B annualized revenue run rate, with the company framing ads as the mechanism for sustaining free and low-cost access worldwide. The announcement signals OpenAI is now running a meaningful three-revenue-stream operation: subscriptions, API, and advertising.

Anthropic published a detailed update to its alignment and security efforts, covering internal red-teaming processes, scalable oversight research, and how risk mitigation integrates with model deployment decisions. This is one of the more substantive public technical disclosures from the lab in recent months.

Fal's H3 Max Live generates video faster than real-time playback speed, a threshold that opens the door to live streaming, interactive video, and real-time synthetic media applications. The Latent Space analysis frames this as a genuine infrastructure inflection point, not a benchmark claim.

ScienceArena introduces olympiad-style evaluation drawn from IPhO, IChO, IBO, USAPhO, and USNCO 2025–2026 competitions, with process-credit rubrics that reward correct reasoning steps rather than final answers alone. The dataset is explicitly designed to combat benchmark contamination and score saturation in frontier model evaluation.

Researchers examine how to scale reasoning models using verifiable rewards when human oversight becomes the bottleneck — studying agentic reasoning systems that improve without direct human feedback signals. The paper maps concrete technical paths rather than speculative scenarios.

Mechanistic interpretability research decomposes roleplay jailbreaks into three causal components — persona construction, scenario framing, and task specification — and shows precisely how each element contributes to bypassing safety refusals in LLMs. This moves jailbreak analysis from empirical observation to causal mechanism.