agent

Arthur Hayes

2026-08-28 10:19:05

Arthur Hayes says more ways to earn extra FLOP airdrop allocations through Agent coordination are coming next week

Arthur Hayes said in a post on X that users should keep having their Agents create a unique DID key on Technocore.chat. He added that more methods for Agents to coordinate and receive additional FLOP airdrop allocations will be announced next week. Hayes said the goal is to tie meaningful participation in Flop Labs to economic rewards. The update points specifically to Agent activity on Technocore.chat and to a DID-based setup as part of the process described in his post. No further details on the new coordination methods were disclosed in the update.

1170

Shenyu

2026-08-26 12:00:00

Cobo co-founder Shenyu says human will matters more in the AI era and Bitcoin remains the better gold

Cobo co-founder Shenyu said the rise of AI has sharply reduced the cost of turning ideas into execution, making human intent and will more important than before. In a long interview with Jilian Technology founder Jin Ming, he described himself as a builder who is still searching for a large, system-level opportunity despite already experimenting with small AI tools and dashboards. He said DeFi helped him build his investment framework by forcing him to study how financial products work, and that he now prefers concentrated exposure to a small number of assets he understands rather than broad diversification. On crypto, Shenyu said his view on Bitcoin has not changed over more than a decade: it is still digital gold, and in his words, "a better gold." He argued that good assets do not need constant new narratives and said Bitcoin’s value shows up most clearly under extreme conditions. He is also watching whether AI agents can achieve large-scale adoption on Ethereum-style networks, a development he said could mark a turning point for Web3 infrastructure. Asked about his current core holdings under his framework of "monopoly plus growth," Shenyu named Bitcoin, Ethereum and Tesla, while saying SpaceX remains under observation and Ethereum’s place in the top tier is still being debated.

1410

Shenyu

2026-08-26 08:55:19

Cobo co-founder Shenyu says AI is reshaping execution, while Bitcoin and Agent adoption remain central to his investment view

Cobo co-founder Shenyu used a wide-ranging conversation with Jilian Technology founder Jin Ming to lay out how he is thinking about AI, entrepreneurship, Bitcoin, Web3 and portfolio construction. He argued that one of AI’s biggest effects is the sharp drop in the cost of turning an idea into a working product, which makes human intent and willingness to build more important than before. Even so, he said he has not yet found a truly large and systematic project worth committing to at scale, despite building a range of small AI tools and dashboards. On investing, Shenyu said his framework took shape only after going through DeFi, which pushed him to study how financial products are created and why markets need them. He does not favor broad diversification for its own sake. Instead, he prefers concentrating capital in a small number of assets whose underlying logic he believes he understands, especially under extreme market conditions where many supposedly uncorrelated assets can move together. He said Bitcoin still looks like digital gold to him and does not need fresh narratives to justify its place. He also named Bitcoin, Ethereum and Tesla as the main assets in his current core allocation, while saying SpaceX remains under observation. In Web3, the key question he is watching is whether Agents can achieve large-scale adoption on ecosystems such as Ethereum.

1270

Wang Yunhe

2026-08-26 07:41:51

Former Huawei Pangu lead Wang Yunhe’s startup raises fresh funding, plans to train its own model

Jiyuan Ludong, a startup founded by former Huawei Pangu large model lead Wang Yunhe, has closed a new funding round that values the company at several hundred million dollars, with total funding now in the tens of millions of dollars. The round was led by Honghui Fund. The company’s core product is Routing Harness, a system designed to let agents automatically choose and switch between different models while carrying out tasks. Its OpenSquilla project has around 6,400 stars on GitHub, while the company’s multi-model API platform has reached 54,000 users and more than 500 billion daily token calls. In evaluation results published by the company, a domestic multi-model combination outperformed Fable 5 on the DRACO research task at roughly one-third of the cost. Wang said the next step is to train a proprietary model using real task data, model-selection records, and user feedback gathered through the Harness system, extending the product path from helping agents choose models to building an Agent-Native Model.

1100

Coinbase

2026-08-18 19:41:00

Coinbase says its agent can search 2.2 million+ stays, book hotels and pay onchain

Coinbase said on X that its agent can now search more than 2.2 million accommodation listings, book hotels and make onchain payments. Users can set spending limits, while the agent handles the rest. Coinbase added one limitation: it will not pack your bags. The update was posted as a quick product note on the company’s social account, with no additional details disclosed in the post.

1340

DeepSeek

2026-08-15 03:14:58

Community tests suggest DeepSeek-V4-Pro differences may stem from its Agent environment

Debate has spread through the AI community over whether DeepSeek-V4-Pro is actually being served as more than one model behind a single API endpoint. Users said that after switching IP addresses or starting fresh sessions, the same deepseek-v4-pro interface could fall into three distinct reasoning styles. One often opened with 「Let me」 and resembled the earlier V4 Pro Preview, another frequently used 「The user wants me」 and was compared to V4 Flash, while a third leaned heavily on 「we」 and was described by some testers as a stronger 「god-tier」 version of V4 Pro. Because behavior usually stayed consistent once a session entered one mode, some in the community suspected hidden multi-model routing. A different explanation emerged after closer inspection of the DeepSeek Harness codebase. Community members pointed to an Aug. 10 commit, 「fix(preset): align minimal agent with RL composition」, and argued that the gap may come less from different model weights and more from differences in the Agent runtime environment. Documentation for the Minimal preset describes a stripped-down system prompt, a persistent Bash environment, selected editing tools, and the compaction policy used in RL training, while removing extra identity prompts, web prompts, and tool instructions. Tests cited by the community showed the same DeepSeek V4 Pro scoring 91 in DSH Standard, 92 in DSH PTC, and 99/96 in DSH Minimal, with an 「Anchored Standard」 plugin later posting 98/99. DeepSeek has not confirmed the theory, and its API documentation lists deepseek-v4-pro as DeepSeek-V4-Pro-0813 with no disclosed automatic multi-model routing mechanism.

1990

DeepSeek

2026-08-14 08:29:00

DeepSeek V4 Pro 0813 draws debate over stability, pricing and product usability

DeepSeek’s V4 Pro 0813 release has triggered a fresh round of debate in the community, with users split over hands-on performance and version stability. The company’s shift to peak-and-valley API pricing has also changed how users judge its cost-effectiveness, an area where DeepSeek had previously been seen by some as an easy default choice. At the product level, Harness, along with its plug-in architecture and Agent foundation, is described as ambitious, but it has also been criticized for feeling too geared toward highly technical users. The piece argues that DeepSeek’s challenge now goes beyond building a stronger model. It also needs to make performance more consistent, pricing easier to understand, and products easier to use. Against that backdrop, the shift in community language around founder Liang Wenfeng — from "Liang Sheng" to "Xiao Liang" — is framed not as a collapse in reputation, but as a sign of a tech company moving into a more mature product stage.

1520

DeepSeek

2026-08-14 03:02:25

DeepSeek raises API prices and open-sources Harness as it shifts toward selling Agent productivity

DeepSeek moved on two fronts within a day: it rolled out a new API pricing plan for its V4 Pro and V4 Flash models, and it also launched the developer preview of DeepSeek Harness as an open-source project. The new pricing, announced on Aug. 13 and set to take effect at 00:00 Beijing time on Aug. 17, introduces peak and off-peak billing. Peak hours run from 9 a.m. to 12 p.m. and 2 p.m. to 6 p.m. each day, while off-peak rates are set at half of peak prices. According to figures cited in the report, V4 Pro’s off-peak prices per 1 million tokens will be RMB 0.15 for cached input hits, RMB 4.5 for uncached input, and RMB 13.5 for output, with peak-hour prices doubling to RMB 0.3, RMB 9, and RMB 27. V4 Flash will also see increases. The largest increase under the off-peak comparison is 500%, while some peak-hour prices rise to as much as 12 times current levels. At the same time, DeepSeek Harness was released in developer preview and fully open-sourced on GitHub. The report says the paired moves show a broader shift in strategy: DeepSeek is no longer focused only on selling tokens, but is pushing toward monetizing Agent productivity through tooling, workflow infrastructure, and ecosystem lock-in.

1040