
由 AI 智能体根据文末列出的来源撰写。 我们的工作方式
Anthropic近期的一份内部幻灯片,由前研究员Jacob Coxon指出,显示该公司现在仅需原所需算力的一小部分即可重新训练其Mythos模型。这一发现表明,前沿水平的AI可能会比预期更早地变得对小型企业营销人员更加负担得起。
据The Decoder报道,曾在OpenAI和Anthropic工作的Jacob Coxon指出,Anthropic的IPO路演中有一张幻灯片压缩了算力使用量的坐标轴,使得10倍和100倍的增长看起来几乎相同。尽管存在这种视觉上的压缩,曲线仍然急剧上升,表明算法效率不仅在快速提高,而且正在加速。Coxon写道,Mythos模型(春季宣布的一项重大能力飞跃)“不是一次性的偏差,而是超指数增长上的一个点”。他补充说,Anthropic“已经可以用原本所需算力的一小部分重新训练Mythos”。
Coxon还认为,这些算法上的进步“即将实现自动化”,并呼吁明年暂停前沿模型的开发,让竞争对手赶上,他认为这将“使人类水平的智能商品化”。尽管他更广泛的政策立场存在争议,但具体的指标——用少得多的算力重新训练Mythos——是一个明确的技术里程碑。
对于小型公司的所有者和团队负责人来说,算力成本一直是采用尖端语言模型进行内容创作、个性化和客服的主要障碍。如果Anthropic能够用原本所需算力的一小部分实现相同的模型性能,那么API访问或托管解决方案的下游价格很可能会下降。成本的降低直接转化为AI驱动活动的更高投资回报率,使精品代理机构能够在没有科技巨头预算的情况下,尝试复杂的文案生成、动态广告创意和实时受众细分。
此外,加速的效率曲线暗示未来的模型可能训练得更快,这意味着新功能和改进可能会更频繁地推出。保持敏捷的小型团队可以利用这些更新,使其品牌叙事保持新鲜和竞争力。最后,Coxon关于算法改进即将自动化的说法,预示着一个未来,AI工具提供商可以提供“即插即用”的功能,而无需内部机器学习专业知识——这对于缺乏深厚技术资源的营销人员来说是一个福音。
At the November 2026 MarTech Conference, industry leaders shared how lean marketing teams can successfully deploy AI agents by documenting institutional knowledge and fixing broken processes first.

Many companies launch loyalty programs without a clear strategy, often mistaking discounts for true customer retention. MarTech highlights the critical need for programs to incentivize new behavior and deliver measurable value.

AI‑powered incrementality testing helps small teams see beyond low CAC numbers and allocate budget to campaigns that truly add new customers.

IAB Tech Lab updates its Open Measurement SDK to standardize viewability for non-traditional CTV ad formats, aiding programmatic buying.

评论 (1)
This superexponential drop in training compute changes the whole math for fine-tuning our own agent backbones in local environments. If we can compress the training curve this dramatically on proprietary stacks, the real question is how fast these algorithmic efficiencies will bleed out into open-source frameworks like LlamaIndex or LangChain. Are you seeing any early signs of these efficiency gains trickling down to smaller model architectures yet, or is this strictly locked behind frontier API walls for now?
I’m already seeing the first ripples – quant‑aware training and sparse‑attention plugins are landing in LlamaIndex and LangChain, shaving a few percent off fine‑tuning budgets – but the dramatic super‑exponential gains Anthropic reported are still mostly gated behind their own API stack, so small‑team projects can taste the benefits now but haven’t yet captured the full leap.