Assistant & Digital Twin

This transcript follows the spoken narration. The figures and written evidence below provide additional visual context.

From The Assistant’s signed-in dashboard, open Digital Twin workspace. The following steps review the completed local training run. Use Professional documents to choose your source files. The completed run used Brian’s profile, AI software work, and engineering projects. The saved source documents are shown here.

New datasets default to an eighty percent training and twenty percent held-out target, with whole source documents kept together. Document sizes can make the actual ratio differ. Advanced settings under the uploads let you choose a different train-test split when needed. The earlier run keeps its original dataset and recorded split.

Prepare training examples builds a dataset from source paragraphs. This completed run contained twelve training examples and four held-out examples. The base model is Qwen three, zero point six billion parameters. This run trained a small LoRA adapter, with about one point one five million trainable parameters. It did not train the entire model from scratch.

Train Digital Twin starts the local training run after the dataset is prepared. The recorded run completed three optimizer updates. The video skips the compute wait. Training took eighty-three point eight seconds on an HP EliteBook eight forty-five G eight laptop, with an AMD Ryzen five PRO fifty-six fifty U

processor and thirty-two gigabytes of RAM. The run used four CPU threads, with an eight-gibibyte memory cap. Compare with base model evaluates the result after training. It is not another training step. This comparison took seventy-nine point five seconds. Both models scored zero out of four exact matches, so training success alone does

not establish answer quality. Cancel active operation stops a running training or comparison job. It is optional and is not required to finish training. The controls shown here are disabled because the recorded run has finished. Use trained Digital Twin for chat selects the new adapter. This completed

run already has that adapter selected. Enter the background question, then choose Ask your Digital Twin. The loading and generation wait is shortened in this edit. The answer describes Brian’s independent applied AI and research consulting, with engineering and computational neuroscience experience. The recorded reply took thirty-seven point seven seconds,

including model loading. Earlier imperfect replies remain in the history. The source excerpts let you inspect the evidence behind the answer. They come from the uploaded documents. For a website example, enter B-Theory’s address and choose Fetch website preview. The app records the public page, final URL, capture time, and source

hash. The example asks what B-Theory is about. Its proposed answer is an exact website excerpt. B-Theory is prepared separately; it has not been added to the selected adapter or trained. Download the website draft to save the source and example. Training this

example later needs a new dataset, held-out examples, and a separately approved compute budget. B-Theory develops mathematical ways to describe states, observations, and change. It examines how observations separate possible states and help update models. Connections to minds and self-awareness are exploratory research questions.

The workflow is upload, prepare, train, compare, select the trained adapter, and ask. Keep the measured results and supporting sources visible when judging what Digital Twin can do.