Anthropic's open-weights holdout turns a safety argument into a supply question

A buyer weighing an API against a weight file is choosing which permanence to live with. Anthropic's case against opening sufficiently capable models is that safeguards can be removed once weights are out and access cannot simply be withdrawn ^[3]. The letter's case for opening them is wider access, better competition and control over how and where an organisation deploys AI ^[2].

Who applies the test is where the Forbes Tech Council column parts company with Anthropic. Harini Gopalakrishnan, listed as founder of theHaze.ai and Industry GTM at Vespa.ai, accepts the need for safety testing and questions the motive behind the proposal, asking how a model leader decides whether a model is safe to release as open weights ^[10][13]. Her objection is about incentives. A company has reason to show its model is safe enough to commercialise, and it can also have an economic reason to argue that a highly capable model is too risky to open, and the column says both can sit alongside genuine safety concerns ^[15]. "A closed model is not automatically safe. An open model is not automatically dangerous," she wrote ^[16].

Her remedy is common benchmarks plus independent assessment involving model developers, researchers, standards bodies, regulators and domain experts, applied to open and closed models with the same rigour ^[5]. Counting those categories, the developer is one participant in five ^[18]. The column does not name an existing body that could run such an assessment.

The analogy she draws is pharmaceutical. Drug discovery carries substantial investment, long development cycles and real risk, so companies get a period of commercial exclusivity, and then patents expire and generic competition spreads the benefit ^[17]. Applied to models, that would let labs monetise the current generation while previous generations get a pathway to broader availability, including open-weight release where appropriate, once they clear an independent safety assessment ^[6]. She calls this "regulated diffusion", and warns against letting temporary safety controls become permanent control over accumulated intelligence ^[7].

There is no purchase order attached to any of this: most enterprises buy an endpoint and a service agreement, and never handle a weight file. Their stake sits in the older generations. If capable superseded models never diffuse, the fallback a buyer assumes, that last year's frontier eventually becomes something you can host yourself, does not arrive. The column also holds the safety line in that direction: "A model with dangerous capabilities does not suddenly become safe simply because a newer generation has been developed," Gopalakrishnan wrote ^[8].

The enterprise half of the argument leans on Microsoft. Satya Nadella has talked about organisations compounding human capital with what he calls "token capital", and Microsoft's version of that future has companies turning their knowledge, workflows and judgment into AI systems that keep improving ^[9]. The column's closing question, whether enterprises are accumulating AI capability or accumulating dependency, breaks off mid-sentence in the published text ^[11].

None of this moves a price list next month, though over several years it bears on where a workload runs, and on whether the option to move that workload in-house is written into a contract or only assumed ^[2].