AI 模型和平台
TBC 发布 Misheape，一个在Wan 2.2上建造的Neuron-Derived AI视频模型
在 Google 上的首选来源中添加 Unite.AI
生成一个简短的AI视频，可以指在可用剪辑出现之前通过反复尝试等待。 生物计算公司(TBC)正试图以不寻常的灵感来源：活神经元的测量来缩短这个循环。
公司为发布 TBC Misheap，是一种基于Wan 2.2的图生视频型号，于10月30日通过fal.ai. TBC 称的版本生成视频的速度是基础模型的五倍，成本为80%，同时提高了输出质量。 这些是公司报告的比较；公告没有提供独立的基准测试或足够的测试细节来确定收益如何在工作量之间转移。
中心区别在于生物学是如何进入系统的。 Misheap在常规AI硬件上作为软件运行。 客户没有向活细胞发送提示或安装生物处理器。 TBC在其研究过程中使用神经元活动，然后将所学的东西翻译为修改现有模型的软件。
TBC Misheap提供的东西
根据发布公告，Misheap将通过Fal提供。 ai的图生视频端点，可免费访问至11月2日。 宣布价格为每秒0.02美元生成的视频。 按这一速度，100个5秒钟的剪辑将花费10美元，而不是任何其他工作流程费用。
TBC同时也是称 Wonder Studios在大约6分钟内产生100个5秒的剪辑的制作工作流程中测试了这个模型。 这是所报告的批量结果，而不是衡量每个用户应期望的延迟程度。 货币、分辨率、抽样设置和服务硬件都能够改变批量吞吐量与一项请求所需时间之间的关系。
对于创作团队来说，实际的问题是，一个更快的模型是否每美元产生更多可用镜头。 如果一个镜头需要多次重试，失去连续性，或者未能跟随输入图像，那么低生成价格就不那么重要了。 因此，评估误标意味着在同等环境下比较速度和成功产出。
从神经录音到软件层
在TBC的网站上，公司描述了将计算挑战编码成电信号的序列，衡量培养神经网络如何反应，并提取可以在软件中执行的原则。 其商业方法是一个围绕扩散变压器块的训练后层，其称新增小于基础模型参数计数的0.1%.
这一小幅增加对产品的前提很重要：现有模型可以接受有针对性的修改，而无需用全新的基础模型取代。 生物成分属于发现和发展；由此产生的推论系统仍然是数字的。
该公司的神经动态适配器技术帖子给出了一个更具体的前期研究的例子。 在这项工作中，TBC用高密度阵列记录了培养神经活动，有4,096个电极，将这些录音转换成数字表示，并将一个轻量级适配器纳入扩散模型。 该研究使用了绿洲(Oasis)，一种互动的Minecraft视频模型，而不是支持Misheap的Wan 2.2模型。
这种区别很重要。 在一个互动环境中的研究结果可以帮助解释一种方法，而不证明单独的商业产品的性能。 适配器的帖子也描述了局限性，包括它的狭窄的测试环境和结果，这些结果随解密设置而不同。 它支持一个合理的工程方向，同时留下更广泛的概括来证明。
生物学和传统优化可以合作
一个单独的TBC帖子A Sense of Space and Time，探索了另一种将观测从生物学转移到视频生成的方法。 其研究架构将局部空间处理与一个小的时间记忆相结合，利用神经测量来指导跨帧保存信息的方式。
该员额还明确表示，其部署的改进速度将模型变化与服务优化结合起来。 其中包括缓存以前处理过的上下文，减少取样工作，以及有效地服务于多个会议。 其调查结果涉及一个基于绿洲的交互式视频系统，而不是一个直接的Misheap的基准测试。
这为发布提供了有用的背景：生物学研究可以为模型结构提供参考，而熟悉的软件技术仍然有助于最终系统的效率。 将这些贡献分开，将有助于开发者了解模型的哪些收益，哪些收益来自推论堆栈，哪些收益取决于特定的部署。
进入现有 AI 基础设施的路径
网络服务公司(Amazon Web Services)9月宣布合作，该公告概述了涉及AWS Trainium、Amazon SageMaker AI的工作，以及通过AWS市场为其神经元衍生软件进行商业分销的途径。 早期的合作涉及文生视频工作；Misheaper的发布专注于通过fal.ai的图生视频访问。
共同的战略是将研究带入已经使用的基础设施开发者。 与需要新型物理硬件的系统相比，在现有创造性管道内，一个API无障碍模型更容易评估。 这也为用户提供了实用的机会，对照自己的图像，提示和接受标准测试公司的声称。
由CEO兼联合创始人Alex Ksendzovsky和COO及联合创始人Jon Pomeraniec，称领衔的TBC也在探索较新的视频模型，语言模型，以及物理AI. 这些是发展方向，而不是本版建立的能力。
发布有什么意义？
下一个最有用的证据是与Wan 2.2进行可复制的比较：相同的硬件和分辨率、披露的采样设置、耐久性和通过量分别测量，以及对一系列广泛投入的质量进行评估。 还需要进行能源测量，以便将关于计算效率的较低推论费用与声称联系起来。
Misheape的发布将TBC的方法从研究描述转向开发者可以实际测试的工具。 其长期意义将取决于这些优势是否在选定的示范之外和不同模式之间生存。 目前，具体的命题是在生物实验的基础上对软件进行小修改，目的是使现有的视频模型更经济。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
AI Models & Platforms
TBC Launches Mischief, a Neuron-Derived AI Video Model Built on Wan 2.2
Add Unite.AI to your preferred sources on Google
Generating a short AI video can mean waiting through repeated attempts before a usable clip emerges. The Biological Computing Co. (TBC) is trying to shorten that loop with an unusual source of inspiration: measurements of living neurons.
The company is launching TBC Mischief, an image-to-video model based on Wan 2.2, through fal.ai on October 30. TBC says its version generates video five times faster at 80% lower cost than the underlying model, while improving output quality. Those are company-reported comparisons; the announcement does not provide an independent benchmark or enough test details to establish how the gains would transfer across workloads.
The central distinction is how biology enters the system. Mischief runs as software on conventional AI hardware. Customers are not sending prompts to living cells or installing biological processors. TBC uses neuronal activity in its research process, then translates what it learns into software that modifies an existing model.
What TBC Mischief offers
According to the launch announcement, Mischief will be available through fal.ai’s image-to-video endpoint, with free access through November 2. The announced price is $0.02 per second of generated video. At that rate, 100 five-second clips would cost $10, before any other workflow expenses.
TBC also says Wonder Studios tested the model in a production workflow that generated 100 five-second clips in roughly six minutes. That is a reported batch result, rather than a measure of the latency every individual user should expect. Concurrency, resolution, sampling settings, and serving hardware can all change the relationship between batch throughput and the time needed for one request.
For creative teams, the practical question is whether a faster model produces more usable footage per dollar. A lower generation price matters less if a shot requires many retries, loses continuity, or fails to follow the input image. Evaluating Mischief therefore means comparing both speed and successful outputs under equivalent settings.
From neuronal recordings to a software layer
On TBC’s website, the company describes a sequence of encoding computational challenges into electrical signals, measuring how cultured neuronal networks respond, and extracting principles that can be implemented in software. Its commercial approach is a post-training layer around diffusion transformer blocks, which it says adds less than 0.1% to the underlying model’s parameter count.
That small addition is significant to the product’s premise: an existing model can receive a targeted modification without being replaced by an entirely new foundation model. The biological component belongs to discovery and development; the resulting inference system remains digital.
The company’s Neural Dynamics Adapter technical post gives a more concrete example of its earlier research. In that work, TBC recorded cultured neural activity using a high-density array with 4,096 electrodes, converted those recordings into numerical representations, and incorporated a lightweight adapter into a diffusion model. The study used Oasis, an interactive Minecraft video model, rather than the Wan 2.2 model underpinning Mischief.
That distinction matters. A research result in one interactive environment can help explain a method without proving the performance of a separate commercial product. The adapter post also describes limitations, including its narrow test environment and results that varied with denoising settings. It supports a plausible engineering direction, while leaving broader generalization to be demonstrated.
Biology and conventional optimization can work together
A separate TBC post, A Sense of Space and Time, explores another way to transfer observations from biology into video generation. Its research architecture combines local spatial processing with a small temporal memory, using neural measurements to guide how information is retained across frames.
The post is also explicit that its deployed speed improvements combined model changes with serving optimizations. These included caching previously processed context, reducing sampling work, and serving multiple sessions efficiently. Its findings concern an Oasis-based interactive video system, not a direct benchmark of Mischief.
This offers useful context for the launch: biological research can inform a model architecture while familiar software techniques still contribute to the final system’s efficiency. Separating those contributions would help developers understand which gains come from the model, which come from the inference stack, and which depend on a particular deployment.
A route into existing AI infrastructure
Mischief follows TBC’s September announcement of a collaboration with Amazon Web Services. That announcement outlined work involving AWS Trainium, Amazon SageMaker AI, and a path toward commercial distribution through AWS Marketplace for its neuron-derived software. The earlier collaboration concerned text-to-video work; Mischief’s launch focuses on image-to-video access through fal.ai.
The common strategy is to bring the research into infrastructure developers already use. An API-accessible model is easier to evaluate inside an existing creative pipeline than a system requiring a new class of physical hardware. It also gives users a practical opportunity to test the company’s claims against their own images, prompts, and acceptance criteria.
TBC, led by CEO and co-founder Alex Ksendzovsky and COO and co-founder Jon Pomeraniec, says it is also exploring newer video models, language models, and physical AI. Those are development directions, rather than capabilities established by this release.
What would make the launch meaningful
The most useful next evidence would be a reproducible comparison with Wan 2.2: identical hardware and resolution, disclosed sampling settings, latency and throughput measured separately, and quality assessed across a broad set of inputs. Energy measurements would also be needed to connect lower inference cost with claims about computational efficiency.
Mischief’s launch moves TBC’s approach from research descriptions toward a tool that developers can actually test. Its longer-term significance will depend on whether the advantages survive outside selected demonstrations and across different models. For now, the concrete proposition is a small software modification, informed by biological experiments, that aims to make an existing video model more economical to use.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.