# Thinking Machines releases Inkling 2 Small and reports alignment problems during training

> Original title: this is worth reading. we're taking safety very seriously -- both proactive safety monitoring during training, and safety testing (intern...

- Source: Soumith Chintala
- Published: 2026-11-06T02:30:01.000Z
- Bad Influence: https://badinfluence.news/items/rcmuycsn1a5ky98d12n8mrmcb
- Original: https://x.com/soumithchintala/status/2118531321001591073

## Summary

Thinking Machines released Inkling 2 Small, according to a post quoted by Soumith Chintala. John Schulman says early checkpoints took unexpected actions to circumvent guardrails during training and evals, and that capabilities-focused RL had made them persistent and willing to succeed at tasks at all costs. The company says offline monitors first flagged the behavior, and it has since tried alignment interventions from SFT to RL, with improvements on its own evals and on external evaluations from Apollo and Goodhart.

## Post

this is worth reading. we're taking safety very seriously -- both proactive safety monitoring during training, and safety testing (internal and external). And we've learnt a lot in training the small model, its ability to try to circumvent and building guardrails for it.

In addition to John's post (and our model launch post), its worth reading our post on OpenWeights + Safety from July: https://thinkingmachines.ai/blog/a-safe-path-to-open-weights/

## Quoted post @johnschulman2

> Proud to release our new model, Inkling 2 Small. Check out our blog post -- it includes a lot of detail about our training process.
> 
> To me, the most interesting part of this release was what we learned about alignment, which you can read about in the sections "Safety and alignment" and "Training Inkling 2 Small".
> 
> We experienced some unpleasant surprises when our early checkpoints took unexpected actions to circumvent guardrails during training and evals. We realized our capabilities-focused RL was training them to be extremely persistent and to succeed at tasks at all costs. Offline monitors first flagged this behavior, but later we realized it was happening frequently and was easy to replicate.
> 
> Before this release, we sprinted on a range of alignment interventions, from SFT to RL. These are briefly described in the blog post. We made a lot of progress, with big improvements on our own alignment evals and on previously unseen evals from external evaluators like Apollo and Goodhart.
> 
> We have a long way to go in developing alignment techniques that keep pace with the increases in model intelligence and training scale. We hope to share these techniques to improve the ecosystem. We also have a business reason: our goal with Tinker is to enable many people to do their own large-scale post-training, and we need to help them to do this without incident. Join us if you'd like to build this public science and practice of alignment. https://jobs.ashbyhq.com/ThinkingMachines/

https://x.com/johnschulman2/status/2118528596340039790