Why is Covenant AI Fleeing Bittensor?

marsbit发布于2026-04-10更新于2026-04-10

文章摘要

Covenant AI announced its exit from the Bittensor network, accusing co-founder Jacob Steeves (Const) of centralized control and governance abuses. The team claimed Steeves unilaterally suspended subnet emissions, revoked community management permissions, deprecated infrastructure without due process, and exerted economic pressure through token sales. Covenant AI highlighted its achievement in developing Covenant-72B—a 72-billion-parameter language model trained collaboratively by over 70 contributors on consumer hardware—as proof that decentralized AI training is viable. However, they argued that Bittensor’s governance is effectively controlled by a small group, with Steeves retaining ultimate authority, contradicting the network’s promised decentralized and permissionless nature. The incident sparked community debate, with some critics pointing to Covenant AI founder Sam Dare’s prior sale of 37,000 $TAO tokens. Meanwhile, $TAO’s price dipped amid the controversy. Covenant AI reaffirmed its commitment to decentralized AI and will continue its mission independently, emphasizing that true decentralization cannot coexist with centralized control.

Editor's Note: Covenant AI has issued a statement announcing its withdrawal from Bittensor, accusing its co-founder Jacob Steeves (Const) of centralizing power in network governance, including suspending its subnet emissions, revoking community management permissions, and applying pressure through token sales.

Previously, Covenant AI was responsible for operating several key subnets, covering pre-training, compute scheduling, and fine-tuning, and had completed a large-scale language model trained collaboratively by multiple parties on general-purpose hardware. This achievement had garnered industry attention and was mentioned by figures such as Jensen Huang and Jack Clark.

The controversy is not one-sided. Some critics have pointed the finger at Covenant AI founder Sam Dare, claiming he sold approximately 37,000 $TAO tokens; meanwhile, supporters believe this conflict may push the network towards a more community-oriented governance model.

On the market front, the price of $TAO fell from around $340 to $286 during the event, before recovering to $291, accompanied by a significant increase in trading volume.

Discussions surrounding governance structure, power boundaries, and incentive mechanisms are ongoing.

Below is the original text:

Covenant AI was founded on a simple, steadfast belief: the training of frontier AI models should not be controlled by any single entity.

For the past two years, our team has been putting this vision into practice. Covenant-72B—a model with 72 billion parameters, trained collaboratively by over 70 independent contributors on general-purpose hardware without requiring permission—has become the largest decentralized pre-training practice for a large model in history. It received recognition from NVIDIA's CEO, was cited by the co-founder of Anthropic, and contributed to a 90% surge in the ecosystem we helped build.

We never actively sought attention. We simply wanted to prove that decentralized training is feasible. And when the results speak for themselves, attention naturally follows.

We want to be perfectly clear about what all this means.

When a single actor can suspend a subnet's emissions, strip node owners of control over their community spaces, publicly "deprecate" projects without due process, and even apply economic pressure through token sales to force compliance—this is no longer decentralization, but centralized control disguised as decentralization.

Every participant in the ecosystem—miners, validators, and investors—should be aware that this power exists and has been exercised by Jacob Steeves (Const). These actions were not taken to safeguard the network's health, but to regain control over a team that had become too independent and difficult to manage. A subnet owner capable of autonomously building a community, making independent decisions, and operating without permission is, in itself, a threat to those whose power is built on the premise that "everyone must depend on them."

The problem with decentralization goes beyond individual incidents.

Bittensor effectively operates under a "triumvirate"—three individuals jointly manage the multi-signature permissions required for network upgrades, packaged as "distributed governance." But that is not the reality. It is more like a "decentralization performance." Jacob Steeves effectively controls this structure, refuses to implement any substantive decentralization of power, and, at his discretion, bypasses processes and consensus to unilaterally deploy changes. Other involved parties act more like legal "shields"—they bear responsibility and face litigation risks, while the true controller remains insulated.

This network constantly talks about governance and decentralization but has never truly implemented it. Power has never left the hands of one individual.

This is the core of the problem.

Bittensor's fundamental promise, the key premise that attracted developers, miners, validators, and investors into this ecosystem, is that it is not controlled by any single entity. But this promise is not true.

Given this reality, we can no longer continue building on this network responsibly. The fundamental representation we made to investors—that this infrastructure is decentralized and permissionless—contradicts the actual governance. We cannot raise funds, attract talent, or ask our community to invest resources on a foundation that can be shaken at any moment by a single will. We are unwilling to pass this risk on to those who have trusted us.

Therefore, with deep disappointment, we announce: Covenant AI will exit the Bittensor network.

Over the past few weeks, Jacob Steeves (aka Const) has taken a series of actions against Covenant AI's operations that are fundamentally incompatible with the principles the network proclaims. These include suspending our subnet's emissions, stripping our administrative rights to our own community channels, unilaterally declaring our subnet infrastructure "deprecated," and applying direct economic pressure through large-scale, public token sales at critical moments of operational conflict.

These were not governance decisions made through transparent, decentralized consensus. They were punitive actions imposed by someone who never truly relinquished control, while publicly claiming he no longer controls the network.

Covenant AI's mission remains unchanged. Decentralized, permissionless AI training is not a feature exclusive to Bittensor; it is a technical capability we will continue to advance. Our research, team, models, and vision will move forward together. We are already progressing on some very important new projects and will share updates with the public soon.

To our community, miners, and everyone who contributed compute, time, and belief to Covenant-72B: you have proven something once thought impossible. This achievement does not belong to a Discord server, nor is it dependent on a network's governance structure. It resides in the research itself, the model itself, and within this team. Wherever we go next, we will continue to earn your trust.

Decentralization is not a marketing narrative that can be overturned when inconvenient. It is a promise to every builder, miner, and investor—a promise that no one can do to others what we have experienced. Either be truly decentralized, or stop pretending.

—Sam Dare, Founder of Covenant AI

相关问答

QWhy did Covenant AI decide to leave the Bittensor network?

ACovenant AI left Bittensor due to concerns over centralized control by co-founder Jacob Steeves (Const), including actions like unilaterally suspending subnet emissions, revoking community management permissions, and applying economic pressure through token sales, which contradicted the network's promised decentralized and permissionless principles.

QWhat was Covenant AI's key achievement on Bittensor before the conflict?

ACovenant AI developed Covenant-72B, a 72-billion-parameter language model, which was the largest decentralized pre-training practice of a large model to date. It was collaboratively trained by over 70 independent contributors on commodity hardware and received recognition from industry figures like NVIDIA's CEO and Anthropic's co-founder.

QHow did Jacob Steeves (Const) exert control over Covenant AI, according to the article?

AJacob Steeves allegedly exercised centralized power by pausing Covenant AI's subnet emissions, removing their administrative control over community channels, unilaterally declaring their subnet infrastructure 'deprecated,' and applying economic pressure through large, public token sales during operational conflicts, bypassing transparent governance processes.

QWhat impact did the conflict have on the $TAO token market?

ADuring the event, the price of $TAO dropped from approximately $340 to $286, later recovering to $291, accompanied by a significant increase in trading volume, reflecting market reactions to the governance dispute and Covenant AI's exit.

QWhat is Covenant AI's stance on decentralized AI training post-exit from Bittensor?

ACovenant AI remains committed to decentralized, permissionless AI training, emphasizing that it is not exclusive to Bittensor. They plan to continue their mission, research, and development of new projects independently, upholding the promise of a truly decentralized infrastructure free from single-entity control.

你可能也喜欢

突发!Anthropic呼吁全员停止AI研究

人工智能公司Anthropic在其官方博客中发布重要观点,指出其AI模型Claude已展现出显著的“自进化”能力,即递归自我提升(RSI)的早期迹象。 核心数据显示,截至2026年5月,Anthropic代码库中超过80%的代码由Claude编写,而在其代码工具发布前,这一比例仅为个位数。工程师的代码交付量达到2024年的8倍。在编程质量上,Claude处理最复杂模糊任务的成功率在半年内从26%跃升至76%,其代码质量被认为年内有望超越人类。 Anthropic提出了“AI能独立完成的任务时长”这一新衡量维度:从2024年3月的4分钟,增至2025年的1.5小时,再到2026年的至少16小时,翻倍速度已加快至每4个月一次。若趋势持续,2027年可能达到数周。 在研究层面,Claude展现强大能力:将训练小模型的代码运行速度优化了52倍,远超人类水平;在一项AI安全研究中,其将效果差距缩小了97%,而人类研究员仅缩小23%。 Anthropic认为,人类在AI开发中的角色正不断收窄,最后优势可能仅剩研究品味与方向判断。公司描绘了三种未来:能力增长停滞;AI加速但人类主导;或AI实现完全递归自我提升,自主设计下一代AI,这可能带来巨大福祉,但也存在对齐失败、最终失控的风险。 为此,Anthropic呼吁,如果存在可验证的机制确保全球AI实验室能同步暂停竞争,其愿意减速甚至暂停研发。OpenAI近期也发表了类似观点,认为自进化迹象将加剧竞争与治理挑战。这表明AI发展的“奇点”可能正在加速逼近。

marsbit25分钟前

突发!Anthropic呼吁全员停止AI研究

marsbit25分钟前

价格里的概率:世界杯赔率是怎么算出来的

本文探讨了世界杯夺冠概率数字的来源与可信度。市场预测和模型模拟是两大主要途径。 市场预测(以Polymarket、Kalshi等平台为代表)通过交易形成价格,价格直接隐含概率。例如法国合约价格17美分,即市场认为其有17%概率夺冠。该价格由做市商报价和交易者成交共同形成,并经过跨平台成交量加权平均得出。截至5月底,相关合约累计成交额超5亿美元。但市场存在“热门-冷门偏差”,即系统性地高估冷门、低估热门,且流动性深度影响价格可靠性。 模型模拟(以Opta超算为代表)则基于球队数据(如Elo评级算法)计算每场比赛胜平负概率,并通过上万次模拟得出各队夺冠频率。例如西班牙夺冠概率为16.1%。需注意,这类模型部分输入数据本身已包含博彩市场赔率信息。 文章指出,目前尚无严谨的跨届学术研究直接对比这两类方法在世界杯预测上的准确度。以2022年为例,阿根廷赛前是第二或第三热门,并非真正“冷门”,但不同来源给出的概率差异显著。此外,加密预测市场的所有交易记录公开透明,利于审计,但也可能因流动性浅而受小额资金影响。 最后,监管环境也是重要变量。美国明尼苏达州已将预测市场运营定为重罪,CFTC与各州司法管辖权存在争议,这种不确定性本身也影响着市场。 总之,“法国17%”和“西班牙16.1%”这两个数字的生产逻辑完全不同。在决赛结果揭晓前,面对任何概率数字,都应追问其生产方式。

marsbit27分钟前

价格里的概率:世界杯赔率是怎么算出来的

marsbit27分钟前

交易

现货
合约

热门文章

加密市场宏观研报:原油飓风、AI巨浪与比特币的十字路口

全球金融市场正经历一场由地缘冲突引发的系统性重估:霍尔木兹海峡封锁导致原油一度暴涨30%,G7紧急释放储备后涨幅收窄,滞胀风险取代通胀成为核心担忧,美元成为“唯一避风港”并逼近100大关,亚太及美股遭遇“黑色星期一”全线重挫;AI领域则冰火两重天,国家发改委提出“十五五”末10万亿规模目标,OpenClaw项目火爆推动概念股狂飙;比特币在宏观风暴中跌破70000美元关键防线。

544人学过发布于 2026.03.12更新于 2026.03.12

加密市场宏观研报:原油飓风、AI巨浪与比特币的十字路口

相关讨论

欢迎来到HTX社区。在这里,您可以了解最新的平台发展动态并获得专业的市场意见。以下是用户对AI(AI)币价的意见。

活动图片