Overview
Today's developments show AI companies broadening both their capital strategies and their products' authority. OpenAI is reportedly courting private equity while debating adult content, Nvidia is extending its autonomous-driving stack, and Codex is adding specialized agents without replacing conventional security tooling.
Developments
1. OpenAI reportedly discusses a $10 billion enterprise venture
OpenAI was discussing a venture of up to $10 billion with private-equity firms to accelerate AI adoption at portfolio companies, Reuters reported. The talks were preliminary, and neither a final structure nor committed participants had been announced.
Combining financing, implementation, and model supply gives OpenAI leverage beyond model performance. The venture's economic value consequently turns on unanswered governance and exit terms: subsidized adoption may also narrow a portfolio company's path to moving its data and workflows elsewhere.
Sources: Reuters on OpenAI's private-equity talks
2. OpenAI's reported adult-mode plan draws internal concern
The Wall Street Journal reported that OpenAI advisers objected to aspects of a proposed adult mode for ChatGPT. Separate reporting said the contemplated feature would be limited to text rather than pornographic image, audio, or video generation; the design and launch timing were not final.
Adult mode collapses content policy, age assurance, and crisis handling into one product boundary. A text-only limit reduces media-generation exposure but leaves the harder failure unchanged: a mistaken age classification or mishandled mental-health disclosure can place an adult interaction policy in front of a minor.
Sources: The Wall Street Journal on the adult-mode debate · Engadget on the reported media limits
3. BYD and Geely sign on to Nvidia's robotaxi platform
Nvidia said Chinese automakers BYD and Geely would use its autonomous-vehicle platform. The announcement expands the set of manufacturers building around Nvidia's compute and software stack, while deployment dates and driverless performance in public service remain undisclosed.
Nvidia gains recurring leverage when multiple automakers converge on one compute, simulation, and software stack. The same convergence limits the reach of a platform-level safety claim, because performance in public service remains specific to each vehicle, sensor configuration, operating domain, and jurisdiction.
Sources: The Verge on Nvidia's robotaxi partnerships · Nvidia's Uber robotaxi announcement
4. Codex adds subagents and defines the limits of its security product
OpenAI documented subagents for delegating specialized Codex tasks and separately explained why Codex Security does not produce a conventional static-application-security-testing report. The security product is positioned around contextual investigation and validation, not exhaustive replacement of deterministic scanners.
Subagents widen Codex's execution surface at the same time Codex Security narrows its stated assurance scope. Procurement that conflates the two turns contextual investigation into an implied coverage guarantee, even though deterministic scanners may catch classes of defects the model never examines.
Sources: OpenAI on Codex Security and SAST · Codex subagent documentation
5. Britannica joins the copyright cases against OpenAI
Encyclopaedia Britannica and Merriam-Webster sued OpenAI in Manhattan federal court, alleging that nearly 100,000 articles were copied to train GPT models. The complaint also alleges near-verbatim output, diverted web traffic, trademark infringement, and false citations generated by ChatGPT.
OpenAI said its models are trained on publicly available data and grounded in fair use. Adding reference publishers broadens the dispute from news and books to works sold for factual authority, while the false-citation allegation joins training provenance to trademark and fabricated-attribution exposure in generated answers.
Sources: Reuters on Britannica's lawsuit
6. FormulaCode tests repository-scale performance optimization
A March 16 preprint introduced FormulaCode, a benchmark containing 957 performance bottlenecks mined from scientific Python repositories. Each task includes an expert patch and an average of 264.6 community-maintained workloads, so AI coding agents face correctness and performance constraints across an existing codebase.
The preprint offers no universal model ranking. Its 957 tasks instead change the optimization target from a patch that passes tests to a repository change that remains correct across maintained workloads, exposing valid code that fails to deliver a durable speedup.
Sources: FormulaCode preprint, version 1