Base Labs recruits Hugging Face and Goodfire for open-weight AI safety
Baseten's research arm plans to publish training and monitoring methods, though its proposed safety standard still lacks a technical specification.
By RuntimeWire Staff · Published
Primary source: TechCrunch
Why it matters
Baseten is trying to make inference providers part of open-model safety. A reproducible standard could give developers controls that remain visible across training, modification and deployment.
Baseten, founded by Tuhin Srivastava, Amir Haghighat, Phil Howes and Pankaj Gupta, launched an open-weight AI safety partnership through its Base Labs research arm on September 17th, bringing Hugging Face and Goodfire into an effort to develop public methods for training, evaluating and monitoring open models.
The move extends the founders' original infrastructure thesis into a messier part of the model lifecycle. Srivastava and his co-founders started Baseten in 2019 after repeatedly seeing capable machine learning models stall before production, trapped by slow deployments, fragmented tooling and systems that failed under real traffic. Base Labs is applying the same operator's instinct to safety: controls have to survive contact with the infrastructure that distributes and serves a model.
That gives the partnership a practical opening. Open-weight developers can inspect and modify a model's parameters, which expands the pool of researchers able to study failures. The same access lets users strip out refusal behavior and redistribute the altered model. TechCrunch reported that Hugging Face listed over 6,000 models associated with "abliteration," a family of techniques used to weaken or remove model safeguards.
Safety at the serving layer
Base Labs says it will develop and publish methods that put safety into model training and deployment. Base Labs wrote, "We believe openness to be an advantage for AI safety," arguing that access to model behavior can make controls more visible and actionable.
The partners cover three distinct positions in the open-model supply chain. Hugging Face operates the distribution platform where developers publish and download models. Goodfire develops interpretability systems intended to expose, monitor and shape behavior inside neural networks. Baseten supplies the runtimes and infrastructure used to serve models in production.
Their likely division of labor can be inferred from those businesses. Hugging Face can provide access to model publishers and users, Goodfire can contribute methods for detecting internal behaviors, and Baseten can test whether controls remain useful under production serving conditions. The partners have not committed to that architecture, however, and the announcement does not assign formal responsibilities.
Goodfire's role gives the project its strongest technical premise. The San Francisco research lab was founded by interpretability researchers and operators including Eric Ho, Tom McGrath and Dan Balsam. Goodfire says its systems are designed to help researchers understand, monitor and align models by inspecting their internal representations. Ho previously described the company's goal as making neural networks easier to understand and repair "from the inside out".
That approach could move safety work closer to the weights and activations that produce model behavior. Output filters and application-level policies can still matter, but they become easier to bypass once a modified model is downloaded and deployed elsewhere. A monitor tied to internal features may offer another layer of evidence about what changed and how the modified system is behaving.
The standard still needs a specification
Base Labs is calling the planned output a safety standard, a label that currently runs ahead of the published details. The partners have not identified a model architecture, dataset, evaluation suite, certification process, software package or licensing framework. They also have not said how a compliant model would be distinguished from one that merely performs well on a narrow safety benchmark.
Those choices will determine whether the project produces infrastructure developers can adopt or another set of research recommendations. A useful standard needs measurable requirements, repeatable tests and a way to account for modifications made after release. Open weights can be copied, fine-tuned and repackaged without the original developer or host remaining in the loop.
The proposed framework also has to separate safety progress from general capability gains. Some evaluations reward models that recognize dangerous scenarios or follow complicated instructions, skills that can improve alongside the very capabilities being measured for risk. Publishing methods, datasets and negative results would let outside researchers inspect those tradeoffs instead of accepting a single aggregate score.
Base Labs has already set expectations that could help. Its public research manifesto promises full publication of experiments, data and negative findings, along with concrete problems and falsifiable results. Earlier work from the lab examined post-training supervision, continual learning and model memory. Safety therefore fits the lab's existing research agenda, even as the Hugging Face and Goodfire partnership pushes Base Labs toward standards work with consequences outside Baseten's own platform.
Baseten expands beyond inference
The timing follows a rapid expansion at both infrastructure partners. Baseten raised a $1.5 billion Series F on June 22nd at a $13 billion valuation, led by Altimeter Capital, Conviction Partners and Spark Capital, with Sands Capital and Wellington Management as co-leads. Goodfire says it has raised over $200 million from investors including Menlo Ventures, Lightspeed Venture Partners and B Capital.
That capital gives the project access to compute, interpretability research and production infrastructure. It also gives Baseten a commercial incentive to help define how open-model safety is implemented. If model developers adopt safety checks at training and serving time, inference providers become part of the control plane instead of neutral pipes carrying whatever weights a customer uploads.
Baseten's founders built the company around the belief that specialized models would spread through products and workflows. That future creates a larger operational problem than serving models quickly. Thousands of customized systems will be harder to audit through centralized policies, especially when their weights and training histories differ.
Base Labs is inviting developers and model builders to contribute to the framework. The first meaningful test will be whether the partners publish a concrete implementation that outside teams can reproduce, modify and challenge. Openness gives the initiative credibility only if the standard itself arrives with inspectable code, evaluations and failure cases.