Open-Source AI Models

Open-source models generally lag behind top closed-source models in performance, and running them locally requires corresponding hardware investment.

For companies whose main revenue comes from API subscriptions, every bit of performance improvement in open-source models adds a bit more substitution pressure.

Once open-source model weights are released they cannot be recalled, and the publisher cannot retroactively add new safety filters to copies already downloaded.

If regulators accept the judgment that "open-source models are higher risk", they may impose additional requirements on open-source models, such as submitting safety assessment reports before release, requiring real-name registration of downloaders, or restricting the release of open weights for models above a certain parameter scale.

Open source wins on being transparent and auditable, closed source wins on controlled access.

Choosing open-source or closed-source models should be based on the specific application scenario, data sensitivity and compliance requirements, rather than being swayed by either side's PR rhetoric.

MIT Technology Review recently published an explanatory article focusing on the controversial topic of whether the practical benefits brought by open-source AI models are enough to offset their potential risks.

Closed-source models only allow inference at the black-box level of inputs and outputs, whereas open weights, training details, and internal mechanisms make the causes of anomalous behavior explainable.

Representative open-source AI models include Meta's Llama series, Mistral, and Stability AI.

Open-source models can be downloaded and deployed directly to a company's own servers, with very low risk of being forced offline by policy, forming a deployment freedom advantage.