Chainfeeds Summary:
Silicon Valley is working to evolve AI into a "reasoning agent," while the rise of open-source forces in China is breaking the monopoly on this technological evolution.
Article source:
https://x.com/nake13/status/2006027328766501223
Article Author:
Zhixiong Pan
Opinion:
Zhixiong Pan: 1) Karpathy: 2025 LLM Year in Review. We are not "evolving/breeding animals," but "summoning ghosts." 2) Google DeepMind: Security of Distributed AGI. AGI is not an entity, but a "state of affairs": a mature, decentralized intelligent agent economy in which the main role of humans is orchestration and verification. 3) OpenAI: Frontier Science: Evaluating AI's Ability to Perform Expert-Level Scientific Tasks. Overall, we found that frontier AI systems are making rapid progress in solving expert-level reasoning problems, especially in self-contained Olympiad problems; however, they are far from saturated in research-style tasks. 4) OpenAI: The State of Enterprise AI in 2025. The shift from "demanding output from the model" to "delegating complex, multi-step workflows to the model." 5) OpenRouter & a16z: The State of AI: An Empirical Study of One Trillion Tokens Based on OpenRouter. The focus of the field is shifting from single-forward pattern generation to multi-step deliberate reasoning inference. 6) Anthropic: How AI is changing the way Anthropic works. Claude is a continuous collaborator, but its use often requires active supervision and verification, especially in high-risk tasks; rather than simply handing over tasks without any verification. 7) DeepSeek-V3.2: Driving the forefront of open-source large language models. DeepSeek-V3.2 performs comparably to GPT-5. 8) UC Berkeley / Stanford / IBM Research: Agent evaluation in production environments. Reliability remains the most significant development challenge, stemming from the difficulty in guaranteeing and evaluating the correctness of agent behavior. 9) Anthropic: AI agent discovers a $4.6 million blockchain smart contract vulnerability. Profitable, real-world autonomous vulnerability exploitation is technically feasible. 10) DeepSeek-OCR: Contextual optical compression. We explore a potential solution: using visual modalities as an efficient compression medium for textual information.
Content source 
