In the past hour, total liquidations across the network reached $10.52 million, with $7.63 million from long positions, with $2.89 million from short positions. BTC liquidation $996,146.14, ETH liquidation $1.68 million.

According to a report by OneMillion_AI, "on-track sampling" in the post-training of large models is key to preventing model degradation and improving problem-solving ability. Online reinforcement learning (RL) and on-track distillation (OPD) are superior to traditional supervised fine-tuning (SFT) because they allow the model to optimize based on its own written steps, rather than rote memorizing external standard answers. SFT forcibly instills standard answers, which can easily damage the model's original knowledge structure and cause forgetting. In contrast, RL and OPD allow the model to find and strengthen the best steps in its own drafts, avoiding the accumulation of errors. Experiments show that using SFT and RL tutors for on-track distillation, student models achieved a success rate of 80.0% and 78.7% respectively in writing the correct code on the first try, both surpassing the tutor model. Even if the SFT tutor becomes less efficient due to over-fine-tuning, the student model still achieves high scores, proving that on-track practice can effectively filter out the tutor's bad habits. Currently, DeepSeek-v4 and GLM-5 have introduced on-track distillation to incorporate the capabilities of expert models.

According to Odaily Planet Daily, as market attention continues to rise following SpaceX's IPO, demand for SPCX-related assets has surged. CoinGlass data shows that the SPCX/USDT pair on the Bitget platform is currently trading at $209.84, a 24-hour increase of 23.69%; its 24-hour contract trading volume reached $395 million, a 579.29% increase compared to the previous period. Bitget now supports SpaceX-related spot tokens and stock contracts, providing users with an entry point to participate in long/short trading and cross-asset allocation for this asset.

Odaily Planet Daily reported on the evening of June 16th that, in response to market rumors that SK Hynix was planning a shareholder return policy worth 100 trillion won, SK Hynix stated that while it is exploring various solutions to enhance shareholder value, the company has never discussed the specific details regarding the scale of shareholder returns mentioned in the reports. (Jinshi)

According to CryptoWorld, Zcash founder Zooko Wilcox stated that following the Orchard vulnerability fix, an AI-assisted security audit did not uncover any new critical vulnerabilities in the Zcash protocol. The audit, commissioned by the Swiss non-profit organization Shielded Labs, examined the Zcash protocol using Anthropic's Claude Mythos model and found no additional flaws. On June 3rd, after discovering a design flaw in Orchard, developers suspended transactions in the pool and restored functionality through an emergency upgrade. This vulnerability, which existed for approximately four years, was discovered by security researcher Taylor Hornby with the assistance of the Claude Opus 4.8 model. The Zcash Foundation stated that there was no evidence of exploitation of the vulnerability or any indication of user data breaches. The Zcash case highlights the importance of AI in crypto security, although AI technology also introduces new risks.