The shift in model training is moving from a focus on pretraining to a greater emphasis on post-training, which is becoming a more substantial part of the overall training process.

The era of scaling pre-training may not be completely over, but reasoning now appears to offer the most potential for maximizing gains from scaling compute.

The last two years have been dominated by large AI models, which have proven to be popular and effective, but we are now entering an era focused on complex workflows around these models.

We've made substantial progress in reducing reward hacking in our models, cutting it down by about 80% in the new iterations.

Instead of just scaling up model size during training, researchers are likely to focus on scaling the compute available for the model's Chain of Thought.

The current landscape shows that large model labs are focused on building powerful foundation models that require less fine-tuning, which may lead to a decline in the relevance of fine-tuning in the near future.

While scaling pre-training may have plateaued, training test time compute could open up a new paradigm for scaling laws, unlocking previously unimaginable capabilities.

Despite increasing GPU resources, the intelligence improvements from the current generation of LLMs are not meeting expectations, with reports of failed training runs and diminishing returns emerging from major labs. Additionally, the lack of high-quality data to train new models has become a major bottleneck for AI development.

We realized that the existing AI models were not sufficient for our needs, prompting us to develop our own models to enhance product quality.