CHATAKE INNOWORKSJigyasa by MindForgeAI · Lecture field guide

THE LEARNING ROADMAP

Machine Learning
in the Age of AI

Equations, data, learning algorithms and intelligent systems—connected through a small experiment you can reproduce.

What you should be able to explain

• How a mathematical model becomes a fitted predictor.
• Why EDA and data timing come before algorithm choice.
• How KNN, SVM, XGBoost, K-means and LSTM differ.
• How to interpret scores, check generated claims and reason about agents.

Opening film: BabyX First Words — University of Auckland Laboratory for Animate Technologies. An expressive interface prompts questions; evidence establishes capability.

CHATAKE INNOWORKSJigyasa by MindForgeAI · Lecture field guide

01 / THE MODEL

A differential equation with an unknown rate

Liquid begins at 80°C in a 25°C room. Assume the room temperature and effective vessel properties remain constant. Temperature changes in proportion to its difference from the room.

t is in minutes; T is in °C; k is in min⁻¹. The structure comes from the physical model. The unknown k comes from fitting observations.

A low error cannot establish that a different cup, liquid or room will follow the same equation.

CHATAKE INNOWORKSJigyasa by MindForgeAI · Lecture field guide

02 / THE EXPERIMENT

Fitting, training and a new prediction

1. Initial guess: k = 0.015 min⁻¹. Calculate predicted temperatures at the training times.
2. Loss: mean squared error = average of (prediction − observation)².
3. Training: repeatedly update k to reduce this loss. The notebook optimises log(k), so k remains positive.
4. Check: compare the fixed fitted model with the two held-out observations.
5. Inference: put a new time into the fitted equation without changing k.

A new input: 45 minutes

≈ 34.18°C

Changing k is additional training. Changing t while keeping k fixed is inference. This estimate extrapolates beyond the 0–30 minute training interval.

Run the supplied cooling notebook or use the interactive experiment. Training uses learning rate 0.001 and 180 gradient updates.

CHATAKE INNOWORKSJigyasa by MindForgeAI · Lecture field guide

03 / CHOOSING THE METHOD

Data, tasks and model families

The separate campus-energy CSV has 600 fictional independent scenarios. Temperature and expected occupancy each have 24 missing entries. EDA checks meanings, units, ranges, missingness, duplicates, relationships and availability at prediction time.

These are different tasks, methods and architectures. Compare models only after making inputs, timing and evaluation conditions comparable.

Use later-period tests for a real time series. The campus teaching rows are independent fictional scenarios.

CHATAKE INNOWORKSJigyasa by MindForgeAI · Lecture field guide

04 / INTELLIGENT SYSTEMS

Foundation models, scores and agents

GPT and Gemini are model families. As of 17 October 2026, official catalogues list GPT-6 Astra and Gemini 4 Argon. A typical development story includes pretraining, further tuning, evaluation and deployment; exact proprietary recipes may be only partly public.

What does a score mean?

Ask which benchmark, dataset, metric, denominator, date and conditions produced the result. A high score does not certify a different local task. Accuracy, MAE, pass rate and human preference measure different things.

Hallucinations

A fluent answer can contain an unsupported claim, invented citation or false number. Open the source and verify the exact claim. Use deterministic calculations for arithmetic. Retrieval helps but does not guarantee truth.

Agentic systems

A complete agent includes tools, state, permission boundaries, stop rules and review. A safe teaching example reads a synthetic CSV, counts missing entries and drafts a report for a person to approve.

Engineering work

AI can assist with bounded tasks. Engineers still frame the problem, test failures, secure integrations and take responsibility for deployment. Avoid assuming that a capability demonstration predicts job replacement.

Primary references: OpenAI models · Google DeepMind model cards · Attention Is All You Need

CHATAKE INNOWORKSJigyasa by MindForgeAI · Lecture field guide

05 / WORKSHEET AND RESOURCES

Your next experiment

Choose a useful ML application. Write a small plan that another person can test.

Questions for discussion

1. Why is fitting k training, while substituting a time is inference?
2. Why can a lower training error still fail in a different room?
3. What separates K-means from KNN?
4. Why might a simple baseline beat an LSTM on a small task?
5. How would you check a generated citation?

Method references: scikit-learn · XGBoost · Keras LSTM. All numerical cooling and campus examples in this guide are synthetic.