生物运算初创公司 The Biological Computing Company（TBC）近日获得亚马逊重大支持，其利用大鼠脑细胞神经模式开发的 AI 模型已向部分 AWS 客户开放预览，并计划进一步推广至所有企业客户。该公司采用务实方法，将图像等资讯编码至生物材料上，透过观察生物处理资讯的过程，打造出能够增强现有生成式 AI（特别是影片生成模型）效率的软体工具。
TBC 由两位神经科学家兼外科医生于四年前创立，总融资额已超过 5,000 万美元，并在旧金山设立了研发实验室。研究人员将活体大鼠脑细胞与人类干细胞置于瑞士 3Brain 制造的多电极矽阵列上，透过电刺激模式刺激神经元并记录其反应，进而分析神经活动以转译为计算模式。在投资人 Jeff Dean 的建议下，TBC 初期专注于视觉资讯与影片生成领域，以便透过既有行业基准验证其技术进展。
TBC 声称其技术相较于基础开源模型，能将影片生成速度提升高达五倍，并大幅降低推论成本。尽管透过 AWS 上架有望让该技术触及更广泛的客户群体，但亚马逊代表亦指出，该技术的扩展性仍面临考验，特别是在产生十分钟甚至一小时的长片时，模型能否维持画质逼真度及前后连贯性，将是其未来面临的关键挑战。（关键数字： ten minutes, an hour）
The Biological Computing Company (TBC), a startup using neural patterns from rat brain cells to advance artificial intelligence, has received a major boost from Amazon by launching its AI model as a limited preview for select AWS enterprise customers. Rather than completely reinventing large language model architectures, TBC takes a pragmatic approach by encoding visual information onto biological materials, observing how living neurons process the data, and designing software tools that mimic these biological processes to enhance existing video generation models.
Founded four years ago by two neuroscientists and neurosurgeons, TBC has raised over $50 million in total funding and established a research lab in San Francisco. Researchers place living rat brain cells and human stem cells onto multi-electrode silicon arrays from Swiss biotech firm 3Brain, sending electrical stimulation into neurons and recording their computational patterns. Prompted by an early suggestion from prominent AI researcher and investor Jeff Dean, TBC initially focused on visual mapping and fine-tuning video generation models to benchmark its scientific progress against established industry standards. (Key numbers: 5,000)
TBC claims its model can generate video up to five times faster and significantly cut inference costs compared to the open-source frontier model it builds upon. While partnering with Amazon allows the ambitious startup to reach a significantly wider customer base, AWS leadership notes that its true scalability remains to be proven, particularly regarding whether the model can maintain visual fidelity and narrative consistency when pushed to generate longer videos lasting ten minutes or an hour.