Inference is mostly matrix math + a few standard ops, and the behavior isn’t hand-coded rule-by-rule. The “algorithm” part is more like instincts in animals: it sets up the learning dynamics and some biases, but it doesn’t get you very far without what’s learned from experience/data.

Also, most “knowledge” comes from pretraining; RL-style fine-tuning mostly nudges behavior (helpfulness/safety/preferences) rather than creating the base capabilities.

Heres my script if anyone is interested in as I find it to be incredibly useful.

find . -type f \( -name ".tf" -o -name ".tfvars" -o -name ".json" -o -name ".hcl" -o -name ".sh" -o -name ".tpl" -o -name ".yml" -o -name ".yaml" -o -name ".py" -o -name ".md" \) -exec sh -c 'for f; do echo "### FILE: $f ###"; cat "$f"; echo; done' sh {} +

On the energy side, Google recently estimated that an average Gemini inference consumes around 0.24 Wh, which is roughly the same as running a microwave for a single second. Older rule-of-thumb comparisons put the figure closer to 3–6 seconds of microwave use, or about 0.8–1.7 Wh per prompt. If you apply those numbers to U.S. usage, you get somewhere between 79 MWh and 550 MWh per day nationally, which translates to only a few to a few dozen megawatts of continuous load. Spread across the population, that works out to between 0.09 and 0.6 kWh per person per year — just pennies worth of electricity, comparable to a few minutes of running a clothes dryer. The bigger concern for the grid is not individual prompts but the growth of AI data centers and the energy cost of training ever-larger models.

For example, I work in operations, so most of what I touch is bash, Ansible, Terraform, GitHub workflows and actions, and some Python. Recently, our development team demonstrated a proposed strategy to use GitHub Copilot: assign it a JIRA ticket, let it generate code within our repos, and then have it automatically come back with a pull request.

That approach makes sense if you are building web or client-side applications. My team, however, focuses on infrastructure configuration code. It is software in the sense that we are managing discrete components that interact, but not in a way where you can simply hand off a task, run tests, and expect a PR to appear.

Large language models are more like asking a genie. Even if you give perfectly clear instructions, the result is often not exactly what you wanted. That is why I use Copilot and Gemini Code Assist in VS Code as assistive tools. I guide them step by step, and when they go off track, I can nudge them back in the right direction.

To me, that highlights the gap between management’s expectations and the reality of how these tools actually work in practice.