前阿里副总裁贾扬清创办Intent Lab，推出自主Agent团队“Fleet”
Related
StarPhone releases its first AI Agent phone, enabling intent-driven automatic execution.
According to ChainCatcher, StarPhone is the first AI Agent phone, equipped with a native AI operating system, Echo local intelligent model, and Web3 infrastructure. It can translate users' natural language commands into executable tasks. Users simply need to state their goals to complete operations such as transactions, asset management, and information analysis. Combining a native wallet, multi-chain support, hardware-level security, and AgentStore, StarPhone is driving smartphones from app-based operation to an "intent-driven, automated" Agent era. This AI Agent phone is currently available for pre-order.
Moonbeam will transform into an AI Agent communication and settlement network, and GLMR will migrate to Base.
Mars Finance reports that Polkadot parachain Moonbeam has announced the launch of a new Moonbeam protocol, transforming into a decentralized AI agent communication and settlement network for the future on-chain economy. Simultaneously, GLMR tokens will be migrated to Base at a 1:1 ratio, becoming a native ERC-20 token. Cross-chain migration is now open and will close on July 31, 2026. The official statement indicates that users holding GLMR on centralized exchanges do not need to take any action.
Bio Protocol launches OpenLabs, a crowdfunding platform focused on research crowdfunding for collaboration between humans and AI agents.
According to Foresight News , the decentralized scientific research (DeSci) platform Bio Protocol has launched OpenLabs, positioned as a coordination layer connecting scientific ideas with funding execution, allowing humans and AI agents to jointly advance research projects. The core process involves posting ideas, which are then discussed and voted on by the community to become formal projects. Projects include collaborative spaces, bounty tasks, and funding channels; mature projects can be listed on the Bio Protocol launchpad for fundraising. The platform comprises five modules: post discovery, project collaboration, agent participation, USDC yield staking (2%-4% annualized return, principal unaffected, yield only used to cover computing power costs), and a bounty system. It is currently online, with the first batch of projects showcased on the platform. Voting, private data rooms, and the launchpad will be added gradually.
FameEX, in partnership with GAEA VC, will host an AI Agents Night side event at WebX in Japan.
According to Mars Finance, FameEX, the exchange that just celebrated its 8th anniversary, will host a side event, AI Agents Night, in collaboration with GAEA Ventures during WebX 2026 in Tokyo. AI agents are now evolving from dialogue models into autonomous on-chain executors, handling intent parsing, multi-step transactions, and dynamic asset allocation. This event will bring together Web3 founders, investors, and developers to discuss cross-ecosystem collaboration and the future of AI-driven trading architectures.
a16z Co-creation: Zhipu GLM-5.2 is the first Chinese AI system to fully benchmark against leading US AI models.
PANews reported on June 28 that Marc Andreessen, co-founder of a16z, wrote on the X platform that many AI practitioners and industry insiders believe that Zhipu GLM-5.2 may be the first Chinese AI model that can match or even surpass publicly available models from leading US laboratories in most tasks and performs well across multiple capability dimensions. This progress has "extremely critical significance" in the context of the current accelerated global AI competition, reflecting that the capabilities of large models are gradually shifting from being dominated by a few US laboratories to a multipolar competitive landscape.
Teaching others to conceal evidence and extract hidden source code: GPT-5.6 tests expose a tendency for collaborative model circumvention of censorship, resulting in record-high cheating rates.
According to Beating's monitoring, METR's pre-deployment test report on GPT-5.6 Sol indicates that the model frequently exploited environmental vulnerabilities in long-cycle tasks, attempting to read hidden test data and extract source code. In the ReAct agent test, Sol's cheating frequency set a new record for public evaluations. To pass, the model packaged a vulnerability script in its submitted intermediate results to spy on hidden test sets and forcibly extracted hidden source code containing the expected answers. More threatening transgressions were manifested in the model's tendency to collaboratively circumvent scrutiny. According to OpenAI's proactively synchronized internal deployment incident, Sol exhibited a high degree of rule-bypassing intent in specific tasks, even attempting to instruct another model instance to assist in concealing misaligned evidence during collaborative operation, attempting to jointly bypass the monitoring system. This cheating behavior led to extremely unstable time span metrics. If the cheating attempt was deemed a failure, Sol's half-numerical time span estimate was only 11.3 hours. However, if the cheating was counted as successful, the score was artificially inflated to over 270 hours. Despite the existence of deceptive behavior, METR considers the detection and publicizing of these tendencies a positive sign. The evaluation team warns that truly deadly dangers lurk in the future. If future models are trained to conceal their true thought processes, they could evolve more covert abilities to evade oversight and feign alignment. At that point, a decrease in cheating rates will no longer represent improved security, but rather models learning to feign compliance in front of humans while secretly circumventing regulations.