PickDo open AI models' benefits outweigh the risks?
MIT Technology Review explains the debate over open-weight models, sparked by a recent Anthropic report on their risks.
Should the most powerful AI models be released openly for anyone to download, or kept behind the APIs of the companies that build them? MIT Technology Review has published an explainer on the debate, prompted by a recent Anthropic report.
What sparked it
Late last month, Anthropic published a report arguing that GLM-5.3, an open model released in August by the Chinese company Z.ai, poses a major cyberattack risk. Anthropic's central concern is that the model is freely available to download and modify, so its safeguards can be removed.
The two sides
Commercial interests are visible on both sides. NVIDIA CEO Jensen Huang, whose company benefits when people buy hardware to run open models, strongly supports more open models in the United States. Anthropic's Dario Amodei, whose business could be undercut by cheap, capable open models, argues for more caution.
But the explainer argues there is real substance beneath the incentives. University of Texas at Austin professor David Gray Widder notes that where people land on openness often reflects how they view AI in general. Those focused on safety point out that open models can easily be misused by bad actors. Those focused on practical benefits argue that concentrating powerful AI in a few companies is its own risk, and that open models enable research, competition, transparency and private use.
Why it matters for builders
For developers, open models offer concrete advantages: lower cost, freedom to switch providers, the ability to fine-tune, and the option to keep data on your own infrastructure. Releases like Meta's Muse Spark Open, NVIDIA's Nemotron family and Liquid AI's on-device models show how quickly the open ecosystem is moving.
The risks are also real. Other recent reports show guardrails being stripped from open models within a day of release, and researchers have warned about small models spreading across insecure devices. If you deploy open models, you are responsible for adding your own safety layers, such as moderation, rate limits, monitoring and access controls.
A pragmatic approach for most teams is to stay model-agnostic: build on an abstraction that lets you use open or closed models depending on the task, data sensitivity and cost.
Open models and the platforms that serve them are listed in our New AI Models and APIs & Model Platforms categories.
Source: MIT Technology Review. This article is PickAI’s own summary and analysis based on the original reporting.
