Expertise plus AI can equal powerful custom capabilities, and Ben Affleck, AI’s latest guru, just showed how an open weight model can be custom trained to perform a specific function with high utility for a vertical industry.
The star explaining how to unfreeze a video model’s weights was not the AI industry crossover many people expected. In a clip from his Bloomberg Screentime interview, the filmmaker described taking existing models, training them to meet cinematic standards, and allowing filmmakers to adapt them further using material from their own productions.
Thomas Wolff’s response captured the surprise at the depth of Affkeck’s expertise and joked about how, with Andrej Karpathy absent from X, Affleck is taking over technical AI tutorials. From a Hollywood star outside the industry, this is unexpected, but this is not the first time Affleck his demonstrated his command over and versatility with the subject matter of AI. He’s even a bona fide exited AI entrepreneur, having recently sold his AI production startup to Netflix for $600 million.
Underneath the joke and the narrative is a practical business opportunity. A general model can supply capabilities that would be expensive to build from scratch. An experienced practitioner can identify where those capabilities fall short, supply relevant examples, and define what a useful result should look like. Additional training can bring the two together.
For B2B companies, that opens a route to custom models built around the patterns of a particular industry, organization or task. The expertise already inside the business becomes part of the development process.
What Affleck is describing
In the clip published by Bloomberg Live, Affleck describes two stages of specialization. His team first takes existing video models and trains them toward cinematic requirements. Filmmakers can then adapt a model further using material associated with their own production. He says they retain the proprietary benefit of that additional learning.
The distinction is between general video capability and a system suited to a particular production. A filmmaker needs consistency, control and outputs that serve a creative intention. Those requirements provide a direction for training and a standard for evaluating the results.
Affleck’s account establishes the approach he says his team took. Affleck’s team builds on an existing video model by training it further on material selected for cinematic requirements. During training, the software repeatedly compares the model’s predictions with the examples and adjusts its trainable numerical settings (the weights) to reduce errors.
Filmmaking expertise guides the selection of material and the evaluation of results, helping shape the model toward the needs of a production. By running additional training on examples of the kind of footage they want the model to handle, the filmmaker helps select and describe those examples; the training software uses them to adjust the model’s weights.
For a typical video-generation model, that can work like this:
- Supply a video clip, often paired with a description of its content or camera movement.
- Add artificial noise to the footage’s numerical representation, obscuring some of the information.
- Ask the model to recover the underlying visual information.
- Compare its prediction with the known training example.
- Adjust the unfrozen weights to reduce the error, then repeat across many examples.
Over those repetitions, the model can learn patterns specific to the new material, such as particular camera movements, lighting or visual characteristics. This is a common training approach; Affleck’s excerpt doesn’t establish his team’s precise procedure. (The excerpt doesn’t identify the underlying models, disclose the full training method or quantify production savings. His reference to a final cinematic layer may describe a stage of specialization rather than one literal layer of the neural network.)
The broader opportunity is nevertheless clear: build on an existing foundation, then use specialist knowledge to make it useful for specific work.
What “unfreeze the weights” means
Weights are the numerical settings a model acquires during training. Together, they influence how it processes inputs and produces outputs. Think of them as a vast collection of adjustable knobs whose positions have already been set through learning.
There are several ways to adapt a model. A team can update all its weights, change selected parts, or keep the original weights frozen and train small added components. One established method, LoRA, reduces the number of parameters that need training.
The training data supplies examples. A training objective measures error, and an optimizer calculates adjustments intended to reduce it. The software applies those adjustments to the weights that have been marked as trainable. Freezing weights tells the training software to keep selected settings unchanged. Unfreezing permits the software to adjust them again. The model continues to use frozen weights; they just remain fixed during that training run.
This is different from prompting because it is a “permanent” not contextual change and improvement. A prompt gives an existing model instructions for a particular use. Fine-tuning changes model settings through additional training. As Hugging Face’s training guide explains, the process starts from an existing model and uses new data to change the outputs it produces.
Affleck’s reference to “moving from training step 8,631 to 8,632” illustrates continuing from an already trained starting point. The team doesn’t have to repeat the entire original training process.
Direct control requires access to the weights and permission to modify them. Open-weight models provide that access, subject to their licenses. Proprietary providers can also offer managed fine-tuning without releasing their weights, as AWS documents for Amazon Bedrock.
Expertise is the necessary ingredient
Training software cannot independently determine what a producer considers a usable shot. Someone has to define the requirements, select relevant examples and identify failures.
Affleck is an Oscar-winning director with knowledge and technique. A filmmaker can recognize a continuity problem, an inappropriate camera movement or lighting that undermines the scene. He also understands how to codify addressing solutions to these issues into an LLM. Knowledge guides both the training material and the evaluation process. An ML engineer implements the training; the filmmaker determines whether the output serves the production.
The potential benefits are commercial as well as creative. Faster iteration and more controllable variations could reduce manual work on selected tasks and give producers more options within a budget. Whether they actually make production cheaper depends on the cost of training, generating, correcting and integrating the results. Versatility is valuable when the variations are usable.
This is a useful model for enterprise adoption: select a function, define its standards, and measure whether specialization improves the work.
Two B2B applications
Manufacturing and industrial services
A manufacturer inspecting one family of components may see a recurring set of defects. A suitable vision model could be adapted using inspection images labeled by quality engineers. The training material would need examples of genuine faults, acceptable variation and difficult borderline cases.
The expertise is essential. A mark that looks alarming to an outsider may be harmless; a subtle deviation may indicate a serious manufacturing problem. Engineers translate that distinction into examples and evaluation criteria.
The potential value is faster inspection triage and less repetitive screening. Performance would need testing on unseen production batches, changing lighting and different equipment conditions.
Video-model customization offers a related opportunity in simulation. NVIDIA documents adapting its Cosmos models using a developer’s own videos for particular physical environments. Generating simulation footage and detecting defects are separate functions, requiring appropriate models for each. Both illustrate how industrial knowledge can shape a more specific application.
Financial services
FinGPT provides a documented example of adapting existing models to specific financial tasks. The project publishes models based on foundations including Llama 2 and uses LoRA for specialization. Its tasks include financial sentiment analysis, identifying entities and extracting relationships from text.
That demonstrates a route from a general language model to selected financial functions. It does not establish profitable deployment at a named financial institution.
A B2B financial-services company could apply a similar approach to extracting defined fields from financing documents or classifying financial communications. Experienced analysts would prepare examples and distinguish meaningful categories from superficially similar language.
Published experiment costs should be read carefully. A training run is only part of implementation. Data preparation, expert review, evaluation, integration and ongoing operation also consume resources.
Five implications for businesses
First, expertise becomes a model-development asset. Experienced employees help define what the model should learn and whether it has learned enough to be useful. Their accumulated knowledge can inform examples, labels and tests.
Second, well-curated data can support differentiation. A large archive is useful only if the company understands its contents. Relevant examples, reliable labels and coverage of difficult cases give training a stronger foundation.
Third, narrow functions make value measurable. A company can compare a model’s ability to classify a defined defect or extract a specified field with the current process. Accuracy, review time and total cost provide concrete measures.
Fourth, control becomes a commercial decision. Licenses, hosting arrangements and contracts determine what a company can modify, retain and move. Open weights can increase technical control, but access alone does not settle every ownership question.
Fifth, specialization still requires oversight. Additional training can improve selected behavior while introducing errors or weakening other capabilities. Evaluation needs to include realistic failures and unfamiliar cases, with human review where consequences justify it.
How a B2B company can get started replicating this
The first step is choosing a frequent, bounded task with a clear standard of success. Measure the current process, then test whether prompting or retrieving relevant documents already delivers an adequate result. Fine-tuning earns its place when it produces a meaningful improvement.
| Requirement | What the company needs |
| Data | Representative examples with expert-approved labels or desired outputs, plus a separate evaluation set. Hundreds to thousands of examples can be an initial planning assumption for some narrow text tasks; vision and video requirements vary substantially. |
| Expertise | A domain expert who can judge the results, an ML engineer or experienced delivery partner, and the person responsible for the operational workflow. |
| Compute | GPU capacity or a managed training service sized to the model and method. Some smaller language-model adaptations fit on one GPU; video work can require substantially more capacity. |
| Cash | A budget covering data preparation, expert review, engineering, training, evaluation, integration and ongoing operation. Training compute alone is not the project budget. |
| Time | As a planning estimate, roughly four to eight weeks for a narrow text pilot with usable data. Collecting data, training video models and integrating production systems can take longer. |
A practical pilot begins with explicit success criteria: acceptable accuracy, unacceptable errors, review time and total cost. The team establishes a baseline, prepares the data, trains a small adaptation and tests it on examples kept out of training. A supervised trial then determines whether the improvement justifies expansion.
The importance of using your own, proprietary data
It’s your biggest competitive advantage. We’ve explored the importance of using your own data, and understanding which data you use, in our Nudgment research. The ADP case study makes that relationship concrete: accumulated operational data creates a valuable foundation, while human expertise determines the knowledge sources, interprets the signals and validates the results.
For custom models, that means understanding where the training material came from, what it represents, what it leaves out and why it belongs in the dataset. An archive can contain outdated practices, recording errors and historical biases. Training on it indiscriminately can reproduce those patterns.
Using company data gives a business material it can examine and contextualize. It does not erase the influence of the base model’s earlier training, and it does not make every company record suitable for reuse.
Affleck’s account illustrates how expertise can guide the adaptation of a general capability to a specific purpose. For B2B companies, the opportunity begins with knowing the work, knowing the data and knowing how to judge the result. Powerful custom patterns emerge when those forms of knowledge direct the training.
