Open models may help cut AI agent costs
The argument pushes back on Dario’s criticism of . The main point is that the claim that people cannot inspect a model does not fit models. such as Claude do not expose their weights, while models such as GLM 5.2 do.
Nemotron3 Ultra is presented as a stronger example because it shares the data, training scripts, and model. Open models can also improve through many people adapting them, with fine-tuning cited as the practical path. The claim that these models must ultimately run in the cloud is also rejected.
Smaller MoE models and such as Qwen 27B can be run directly, which may reduce dependence on paid s in some cases.
Key points
- models let people download and run the model’s learned values.
- GLM 5.2 is used as an example of a model whose weights can be inspected, unlike Claude.
- Nemotron3 Ultra is cited as going further by sharing data, training scripts, and the model.
- Fine-tuning is presented as evidence that open models can be improved for specific uses.
- Smaller MoE models and such as Qwen 27B may allow direct operation without relying only on the cloud.