How to Build Proprietary Models
You can build a model around work your business already knows how to do. Start with an existing model, teach it with your own examples, and measure whether it gets better.
1. Get a high-RAM computer
A MacBook Pro with an M5 Max, 40-core GPU, and 128GB of memory gives you room to train and run models locally. It can run and train smaller models easily and you can work up from there over time as work becomes more complex.
2. Download OpenAI Codex
The practical unlock is OpenAI Codex. Use it to help prepare the data, write training scripts, test results, and deploy. Jetpack, my personal agent, helps coordinate the work. Use it as your development partner, it’s the best developer on the team now.
3. Choose the right outcome
This is the most important part. What valuable business outcome the team keeps driving toward, and what data do you already have that shows how to achieve it?
A shortcut for thinking about this is the input and output are known, and the model figures out the complex stuff in between automatically, given there’s enough examples to learn from.
For Sales and Marketing use cases, one example is the outreach and proposal work we’re doing in the Autonomous Enterprise. A model learns from examples to turn a prospect’s needs into a relevant pitch, or transform a project brief into an proposal response. You can measure customer responses, meetings, and paid work alongside accuracy. These are candidates for the training data set.
For R&D, consider the language side of PROME’s Care Core: turning requests into structured descriptions of a machine’s actions. Examples could help train that interpretation step. As a reminder, Care Core’s separate mathematical values and permission checks remain outside the language model.
In general, the goal is to pick one repeatable outcome, define what a good result looks like, and what improving it would be worth financially. Remember, the work we do needs to be linked to some improvement on the income statement.
4. Prepare your data
There’s no getting around it, you need to clean up your data. Remember the old adage: garbage in, garbage out? Thus, you need to pair each input with a good output. This requires removing duplicates, mistakes, and information you don’t have permission to use. Codex can help with this.
You must also keep some of that example data separate from the training data to do testing on. You need to know whether the model works on data it hasn’t already seen or been trained on to validate the accuracy of the finished system.
Sometimes it will require multiple rounds of training to get the model working the way it should.
5. Download an open-source model
Find an open-weight model you can run locally on your computer. Again, Codex can help you find the right one for your use case and that will fit on your computer and handle the training.
Open weight models means its “settings” are available to download and edit via your training.
One critical note, however. Check the chosen model’s license before using it commercially. The terms determine whether you can keep your data and improvements private and whether any sharing obligations apply. Those obligations vary by license, so check exactly what must be shared before building a product you plan to sell.
6. Train and test the model
Use Apple’s MLX tools to fine-tune the model on your examples. It may require downloading some updates to your terminal and other applications to update them for the training process.
Fine-tuning a model means that it adapts an existing model’s weights instead of training one from scratch. That is the better approach because it’s likely you don’t have millions of dollars and millions of internet-wide data to teach a model the basics of language and logic. So, these days, you get to skip that initial step.
Be sure to compare your newly trained model on your separated test set to see how it performs. Measure accuracy, speed, and the result you care about to see if it gives you the desired result compared to how you used to do it manually. Here’s an MLX training guide that you may want to point Codex at to help.
7. Deploy it and put it to work
Once you’ve got your trained model ready, you need to make it available to yourself and others.
One option, that really only works for you or internally at your organization is to serve it from your local computer. However, once more people need to use it and to expose it to the open internet, you will need to transfer that model to some compatible computer in the cloud to host it.
If you’ve trained your model on an Apple device, then the easiest is to use an Apple Silicon cloud host that supports MLX. These are less prevalent. One MLX hosting option
Which is why you see many companies doing the initial training on NVIDIA chips and then serving those models on a compatible version for NVIDIA hosting.
The key thing to realize here is that you can move an Apple-trained model to NVIDIA hosting, provided the model format and serving software are compatible. Depending on the model, this may require exporting or converting the weights and adapters for the hosting setup.
8. Enjoy the benefits as it can keep improving
The primary benefit is now you have a trained model that can handle work tasks more capably than manual humans may have in the past. For example, responding to similar support requests.
Now, you probably want to keep improving the model as new work is completed and you collect new data and information about your business and the market.
This is where the world is beginning to shift from static models that are trained every so often to adaptable models that are updated as new data flows in. Each update still requires another training-and-testing cycle before deployment. That’s the shift toward Biologic Intelligence.
This fits our Autonomous Enterprise: understand the business, choose the work that should improve, then build and test the agreed capabilities.
—Sean