Lab cells, videos on GPUs: the strange recipe for Mischief 1

With Mischief 1, TBC leverages discoveries made on living neurons to accelerate Wan 2.2. An AI video announced five times faster and 80% less expensive.

What if the best way to speed up AI video generation was to study how living neurons process information? That's the bet The Biological Computing Co. (TBC) is making. The San Francisco startup has turned several years of neuroscience research into its first commercial product. Called Mischief 1, the image-to-video model builds on Wan 2.2, promising comparable quality while dramatically reducing generation times and costs. What makes the approach unusual is where the optimization supposedly comes from: electrical activity recorded from living neuron cultures, translated into software improvements.

The proposition is as intriguing as it deserves scrutiny. TBC claims its technology generates video five times faster while cutting inference costs by 80% compared with the original model. Those figures, however, come from the company's own evaluations, without detailed independent benchmarks establishing exactly how much of the improvement can be attributed to its biological research. One thing is clear: Mischief 1 does not run on living cells. The neurons remain in the laboratory, while customers use an entirely software-based model running on conventional GPUs. When Living Neurons Become a Source of Computational Efficiency

TBC was co-founded by Alex Ksendzovsky and Jon Pomeraniec, whose backgrounds span neuroscience and neurosurgery. Their company is developing a research platform designed to understand how biological neural networks process information and identify principles that could make artificial intelligence models more efficient.

The process involves growing living neurons on devices equipped with electrodes. Researchers convert visual information into electrical signals, stimulate the neural networks, and record their responses. This allows them to observe how electrical activity spreads, how neurons interact, and how those responses evolve over time.

The goal is not to recreate the entire human brain, but to identify particular computational behaviors worth investigating. Researchers compare the stimuli delivered to the neurons with the resulting activity, looking for relationships between the timing of electrical signals and the processing of information.

Those observations are then translated into software modifications and tested on existing AI architectures. TBC describes an approach that goes beyond drawing abstract inspiration from biology: it uses direct experimental measurements to guide its optimizations.

On its technology page, the company outlines the main stages of this process. The publicly available information, however, does not yet provide enough detail to reconstruct the exact mathematical transformations involved or independently measure the contribution of each component. Accelerating Wan 2.2 Without Rebuilding the Entire Model

Mischief 1 is an image-to-video model capable of animating a starting image based on a text instruction. It builds on Wan 2.2, an open-weight video architecture already used across many generative creative environments.

Rather than training an entirely new system, TBC added a proprietary software component derived from its neural research. According to the company, this modification accounts for less than 0.1% of the original model's parameter count.

That lightweight design is central to the proposition. Conventional optimization methods often aim to reduce numerical precision, decrease the number of generation steps, compress internal representations, or improve hardware utilization. TBC claims to introduce an additional source of efficiency based on the study of biological neural networks.

The result is reportedly a version of Wan 2.2 capable of generating comparable sequences in substantially less time. The company also says it is applying the same approach to newer video models, with early results suggesting even greater efficiency improvements.

What remains difficult to determine is how much of that improvement comes exclusively from the component derived from biological experiments. Without a detailed technical publication or independent comparisons that control for the full inference setup, the distinction between a scientific breakthrough and conventional software optimization remains unresolved. Five Times Faster, but the Benchmarks Need Context

On the official Mischief 1 website, TBC highlights a generation speed increase of at least five times and an 80% reduction in inference costs compared with the base model.

One demonstration shows Wan 2.2 taking 142 seconds to generate a video, while Mischief 1 completes the same type of task in 12 seconds. That represents an approximately twelvefold difference, substantially higher than the fivefold improvement emphasized in the company's broader messaging. The figures may reflect different testing conditions or measurement methods, but the published information does not fully explain the discrepancy.

TBC notes that its generation times come from internal GPU benchmarks and that actual performance may vary depending on hardware and operating conditions. Its claim of visual quality comparable to Wan 2.2 also requires independent evaluation across complex scenes, fast movements, and different types of source images.

That uncertainty does not diminish the potential significance of the reported gains. Video generation remains computationally demanding, especially when creators need to produce numerous variations before finding a satisfactory sequence. A substantial reduction in generation time could directly change how creative teams work, even if it does not guarantee better creative results. One Hundred Five-Second Videos for $2.50

Mischief 1's other major selling point is price. TBC advertises a starting rate of $0.005 per second of generated video, or half a cent.

On fal, that rate applies to 480p generation. A five-second clip therefore costs $0.025, while 100 attempts add up to $2.50 in generation fees. At 720p, the listed price rises to $0.01 per second, or $0.05 for a five-second video.

These amounts