3 ms·
There's nothing new here in terms of architecture. Whatever secret sauce is in the training.
by redox99 1y ago
There's nothing new here in terms of architecture. Whatever secret sauce is in the training.
- BoorishBears 1y agoPart of the secret sauce since O1 has been accesss the real reasoning traces, not the summaries. If you even glance at the model card you'll see this was trained on the same CoT RL pipeline as O3, and it shows in using the model: this is the most coherent and structured CoT of any open model so far. Having full access to a model trained on that pipeline is valuable to anyone doing post-training, even if it's just to observe, but especially if you use it as cold start data for your own training.
- anticensor 1y agoIts CoT is sadly closer to that sanitised o3 summaries than to R1 style traces.
- BoorishBears 1y agoIt has both raw and summarized traces.
- anticensor 1y agoI mean raw GPT-OSS is close to summarised o3.