Thinking Machines argues for staged access, using Inkling as the test case
- Source
- x.com
- Date
Releasing weights indiscriminately isn't safe. Neither is keeping capable models inside a few labs. We think there's a path between them. We haven't mapped all of it. Our new post covers the part we can see: how we assessed Inkling, and why access should widen in stages. https://t.co/zGFbuXNr0Y
We’re also expanding our safety team. If you’re interested in joining, please reach out! https://t.co/yS2F8gZqhh
Context
Thinking Machines proposes widening model access in stages because downloadable weights cannot be recalled once distributed. Hosted access offers intermediate choices: people can use or a model while the provider retains the ability to monitor usage and revoke access. The proposed stages are meant to generate evidence and give defenders time to prepare. The framework.
For Inkling, the company reports internal testing, external assessments and experiments that fine-tuned away refusal behavior. It concluded that release was unlikely to add material risk beyond existing models. That is the company's comparative risk assessment. Its broader proposal leaves thresholds for advancing or stopping a release to be specified. Assessment and unresolved criteria.
- fine-tuning — Taking a trained model and training it a bit more on your own examples so it gets better at one specific job.
- open weights — A model whose trained parameters are published for anyone to download and run — unlike API-only models you can access but never possess.
Checking sign-in…
Loading comments…