Qwen 3.8 follows GPT-5.5 Pro reasoning prefills
This post reveals how Qwen 3.8's reasoning abilities dramatically boost when 'prefilled' with just 1% of GPT-5.5 Pro's internal thought process. The experiment suggests Qwen may have learned from or been distilled using GPT-5.5 Pro, or a similar GPT model, rather than Opus. This finding ignited a fiery discussion on HN about 'stolen thoughts,' intellectual property in AI training, and the ethics of model distillation between major AI players.
The Lowdown
This follow-up experiment delves into the controversial topic of "reasoning prefills" for open models, using GPT-5.5 Pro as the "teacher" to investigate potential model distillation. The study aims to uncover if specific open-source models have integrated or learned from the internal reasoning processes of more advanced, proprietary systems.
- The methodology involves comparing two types of responses from target models: an unassisted, ordinary response and one where the model's reasoning channel is initiated with the first 1% of GPT-5.5 Pro's reasoning trace.
- The evaluation measures how much of the teacher's visible answer appears in the first 100 tokens of the target model's output, using unigram, bigram, and trigram source recall.
- Across 45 problems (15 each of STEM, non-STEM, and synthetic puzzles), Qwen 3.8 A95B showed a remarkable +18.18 percentage point improvement in source recall when prefaced with GPT-5.5 Pro's reasoning, moving from 16.79% to 34.97%.
- This improvement was particularly pronounced in STEM problems (+26.99 pp) and synthetic puzzles (+14.75 pp).
- Kimi K3 also saw a modest boost (+4.54 pp) but already had a higher base overlap with GPT-5.5 Pro's outputs.
The significant performance increase in Qwen 3.8 when guided by GPT-5.5 Pro's reasoning traces strongly suggests that Qwen may have been exposed to, or distilled from, GPT-5.5 Pro or a closely related GPT model during its training, indicating a different lineage than previously hypothesized with Opus.
The Gossip
Distillation Debates & Data Disputes
A heated discussion erupted around the ethical and intellectual property implications of AI models potentially being trained on the reasoning traces of proprietary models. Many users viewed this as unethical 'stealing,' while others highlighted the perceived hypocrisy of 'frontier' labs, arguing that they too scraped vast amounts of public data without explicit permission. The conversation questioned whether internal 'thoughts' can truly be intellectual property and if this practice exposes a lack of original innovation from certain labs.
Reasoning Revelation & Reproducibility Roadblocks
Commenters questioned the precise mechanism for accessing 'raw reasoning tokens' and the validity of the experiment's findings. Reference was made to the 'Stolen Thoughts' paper, which details methods to 'decrypt' such traces. A key point of contention was whether Qwen 3.8's observed behavior truly indicated distillation or if it could be attributed to the model having been fine-tuned on the *publicly released* 'stolen thoughts' examples after their publication in August. Some users also shared anecdotal observations of Qwen's internal 'Chain of Thought' (CoT) styling, comparing it to the 'grug-talk' sometimes seen in proprietary models like GPT-5.6 Sol.
Geopolitical AI & Generative Growth
The findings sparked considerable commentary on the competitive landscape of AI development, particularly concerning Chinese AI labs like those behind Qwen. Some interpreted the results as evidence of a 'fast-follow' strategy, where Chinese development focuses on distilling and copying from leading US models rather than pioneering new approaches. Conversely, others argued this approach is pragmatic for optimizing performance on diverse hardware or a justifiable response, given that US labs trained on vast public datasets. The sentiment that this practice creates 'no moat' for frontier models and accelerates the spread of capabilities was also expressed.