Alice’s safety training effort pairs Inflection AI and AI21 with NVIDIA Nemotron open models

Alice, the research and product initiative backing a suite of AI safety demonstrations, says it has collaborated with Inflection AI and AI21 to show that NVIDIA’s Nemotron open models can be trained for frontier safety use cases. The work is positioned as an approach to improve safety behavior through training, rather than relying only on guardrails or after-the-fact filtering.

The effort centers on post-training, a common machine-learning technique where an already-trained model is further adjusted using additional data. In this case, Inflection AI and AI21 are described as post-training NVIDIA Nemotron open models using Alice’s safety data, which the collaboration says is designed to address safety and security topics that become more difficult as models get more capable.

The announcement describes the goal as demonstrating that “powerful open models” can be adapted across a range of safety and security challenges. It also suggests the training is meant to preserve capability, framing the collaboration as evidence that open models can be made safer without “compromising” performance in ways that would reduce usefulness.

While the communication emphasizes training-based safety, it does not, in the material available for this review, provide granular results such as benchmark names, scoring methodology, target threat models, or the size of the datasets used. It also does not spell out whether the post-training produced specific model variants (for example, distinct checkpoint versions) or how broadly the safety data generalizes across domains.

NVIDIA, whose Nemotron open models sit at the center of the demonstration, is a key supplier of accelerated computing and software used across the AI industry. In sector terms, the collaboration reflects a broader shift toward open model ecosystems, where multiple labs can iterate on shared model families and contribute tooling or training recipes.

The collaboration’s framing is also consistent with a market dynamic that has intensified over the past year, where major AI vendors and downstream labs are pressed to show not only performance, but also controllability, robustness, and security. Training on safety data is one of the most direct ways to attempt those improvements, and the involvement of multiple companies indicates that the method is intended to be reproducible across environments.

Still, several practical details remain undisclosed in the available information. The communication does not list the specific safety categories covered, the evaluation framework used to compare pre- and post-training models, or any assurances about how the changes affect behavior outside the tested scenarios.

For the next phase, the most consequential indicates to watch would be whether the collaboration publishes more complete evaluation data, clarifies the safety-and-security threat coverage, and indicates whether the approach can be scaled or transferred to other open-weight model families beyond Nemotron. Separate from the demonstration, investors and customers will likely look for evidence that safety training can be repeated with predictable outcomes as open model adoption grows.