Epitomsoft Academy · Module 4 of 6

From AI to AGI and superintelligence

What do AGI and superintelligence actually mean? Scaling laws, emerging capabilities, how progress is measured – and the expert debate on timelines, with evidence separated from hype.

Apply for the programme
All modules

What you will learn

Understanding where AI is heading – critically

Leading labs openly aim for artificial general intelligence (AGI), and some researchers discuss superintelligence – AI that would exceed human ability in almost every domain. Neither exists today. This module teaches you to reason about these ideas with evidence, not headlines.

  • Define narrow AI, AGI and superintelligence – and their levels
  • Explain scaling laws and why bigger models became more capable
  • Read AI benchmarks critically and spot misleading claims
  • Summarise the main expert positions on timelines and risks
  • Form and defend your own evidence-based view

Module at a glance

  • Format: live online sessions + hands-on lab
  • Level: All levels (concept-focused)
  • Prerequisites: Module 1 recommended
  • You will build: An evidence-based briefing on AI progress

Topics

Key concepts, explained with examples

Narrow AI, AGI and superintelligence

Narrow AI excels at specific tasks; AGI would match skilled humans across most cognitive tasks; superintelligence would greatly exceed humans in nearly all of them. Only the first clearly exists today.

Example: A chess engine is narrow AI. A system that could learn any office job as well as an expert would be AGI.

Levels of AGI

Researchers propose levels based on performance and generality – from “emerging” to “superhuman” – instead of a single on/off definition.

Example: Today’s general-purpose chat models are often classed as “emerging AGI”: broad, but below skilled-human level in many areas.

Scaling laws

Model performance has improved predictably as training compute, data and parameters grow. Research also showed models need enough training data for their size.

Example: The “Chinchilla” study found many large models were under-trained, with roughly 20 training tokens per parameter as a compute-optimal rule of thumb.

Emergent and reasoning capabilities

New abilities – multi-step reasoning, coding, tool use – appeared or improved sharply at larger scales and with new training methods such as reasoning-focused models.

Example: Models that “think step by step” before answering solve far more maths and coding problems than earlier models of similar size.

How progress is measured

Benchmarks test knowledge, reasoning, coding and novel problem-solving. Each has limits: contamination, saturation and narrow coverage.

Example: ARC-AGI tests puzzles designed to be easy for humans but hard for AI; SWE-bench tests fixing real GitHub issues.

Compute, data and energy

Frontier progress depends on chips, data centres, electricity and high-quality data – real-world constraints that shape the pace of progress.

Example: Training a frontier model can require tens of thousands of specialised chips running for months.

The timeline debate

Experts disagree widely: some expect AGI within years, others within decades, others doubt current methods will get there. Good analysis states uncertainty honestly.

Example: Compare forecasts from researchers, forecasting platforms and lab leaders – and ask what evidence would change each view.

Implications for society and work

Advanced AI could transform productivity, science and the economy – while raising questions of safety, fairness, employment and power concentration.

Example: AI-assisted drug discovery could speed up research, while automation reshapes entry-level roles – which is why skills matter.

Worked example

Reading an AI claim like a scientist

A step-by-step method you will apply to real announcements during the module.

Step 1

What exactly is claimed?

Example claim: “Our model reaches human-level reasoning.” Which tasks? Which humans? Measured how?

Step 2

Check the benchmark

Is the test public (risk of training-data contamination)? Is it saturated? Does it measure what the headline says?

Step 3

Look for independent results

Has anyone outside the company reproduced it? What do independent leaderboards and reports show?

Step 4

Separate capability from deployment

A strong demo is not a reliable product. Ask about cost, failure rate and real-world use.

Step 5

State your conclusion with uncertainty

Write: “This is evidence of X, but not of Y; I would update my view if Z.”

Step 6

Track it over time

Use public trend data (e.g. Epoch AI, the AI Index) to see whether the claim fits the longer trend.

Hands-on lab

Write an “AI progress briefing”

Working in small teams, you produce a two-page briefing for a business leader on where AI capabilities stand today.

  • Pick three capability areas (e.g. coding, reasoning, agents)
  • Gather evidence from benchmarks, research papers and public trend data
  • Summarise what is proven, what is promising and what is speculative
  • Present opposing expert views fairly
  • Recommend what a company should prepare for in the next 2–3 years

Key terms

Vocabulary to master

  • Narrow AI
  • AGI
  • Superintelligence (ASI)
  • Levels of AGI
  • Scaling laws
  • Compute
  • Parameters
  • Emergent capability
  • Reasoning model
  • Benchmark
  • Contamination
  • Saturation
  • Forecasting
  • Takeoff

Check your understanding

Module 4 quiz

Score 70% or more to pass. Pass all six module quizzes to unlock your Frontier AI certificate on My progress.

Further learning

Free & open-source references

Original research, open-source projects and free courses – all free to access.

▶ Watch 18 free AI video lessons
← Module 3: Building with frontier models
Module 5: AI safety & governance →

Ready to master frontier AI?

Join the Epitomsoft Academy advanced programme – tell us about yourself and we will be in touch.

Apply now