According to a report by AI Yi, an on-chain analyst, research indicates that large models gradually lose their ability to absorb new knowledge during training, eventually becoming increasingly rigid. If this loss of plasticity cannot be overcome, large models will be unable to learn continuously at low cost, requiring retraining on all historical and new data with each knowledge update, consuming enormous amounts of computing power. The research points out that while increasing model size can delay degradation, the marginal benefit diminishes, and simply adding more parameters cannot fundamentally cure the loss of plasticity. A 1-b parameter model becomes less efficient after training 1.8 trillion tokens, while a 7-b model shows this after 9 trillion. The research also points out that the reasons for large models becoming less efficient include increased parameter size hindering gradient propagation, large-scale neuronal dormancy, and attentional paralysis. Potential solutions include limiting parameter inflation, periodically performing "neural resets" on dormant neurons, and introducing random noise into the attention mechanism.

In the past hour, total liquidations across the network reached $14.07 million, with $13.00 million from long positions, with $1.07 million from short positions. BTC liquidation $10.35 million, ETH liquidation $1.36 million.

According to Bijie.com, Anthropic recently shared its engineering experience with a human-machine collaboration team. Multiple employees and several agents with independent system credentials work together on Slack. These agents are directly attached to the team roster and communication threads, functioning like human employees with clearly defined roles and autonomously driving projects. To ensure effective integration of agents into the team, collaboration is made public by default. The company sets security boundaries to avoid cumbersome single-document authorization decisions. The team assigns specific roles to agents by writing skill files to prevent information fragmentation. Agent autonomy is directly proportional to reliability; the engineering manager assigns agents to independently fix 500 bugs and requires them to submit weekly reflection reports to prevent recurrence. The team uses a double-confirmation mechanism, with one agent reviewing the work of another, ensuring the sustainable operation of the human-machine team.

According to Bijie.com, Bybit CEO Ben Zhou stated at the Point Zero Forum in Zurich that exchanges are evolving from trading platforms into infrastructure for payments, tokenized assets, and global financial access. He believes the key to tokenization lies not in putting assets on-chain, but in establishing deep liquidity. If similar assets are issued on different chains, it could exacerbate ecosystem fragmentation; the future focus should be on interoperability and liquidity integration. Furthermore, he anticipates that AI will become the gateway to financial interaction, with AI agents helping users automatically match products and strategies.

According to Bijie.com, the State Administration for Market Regulation, in conjunction with relevant departments, is accelerating the development of standards for cutting-edge technologies such as intelligent agents, and dynamically improving the national standard matrix for artificial intelligence to adapt to industrial development. Currently, the national standards under development include, in addition to intelligent agents, standards for other cutting-edge technologies such as embodied intelligence, world models, and ontology models, as well as foundational standards for computing infrastructure, high-quality datasets, simulation testing platforms, deep learning compilers, and open-source model platforms. Furthermore, the focus will be on industry management and the evaluation of AI implementation, promoting the release of key standards such as those for assessing the maturity of AI enterprise intelligence and evaluating high-value AI application scenarios.