How Fast Is AI Actually Moving?
Lesson 9.1 — How Fast Is AI Actually Moving?
AI is moving fast. That much is uncontested. But "fast" means different things to different people, and the pace of change is often described in ways that are either breathlessly optimistic or catastrophically alarming — neither of which is particularly useful for making decisions about your own life and work.
This lesson gives you a grounded view: a factual timeline of what has happened, a clear explanation of what exponential change actually feels like from inside it, and a realistic (not hyped, not doomy) two-to-three year view of where things are likely to go.
A Timeline: 2020–2024
| Year | Development | Significance |
|---|---|---|
| 2020 | GPT-3 released (OpenAI) | First model to demonstrate convincingly human-like text generation at scale. 175 billion parameters. Limited access. |
| 2021 | DALL-E (OpenAI); Codex powers GitHub Copilot | AI image generation becomes a reality; AI begins assisting programmers meaningfully |
| 2022 (Jan) | ChatGPT-precursor models; Midjourney launches | Image generation goes mainstream; early versions available to public |
| 2022 (Nov) | ChatGPT launched | Reached 1 million users in 5 days; 100 million in 2 months. Fastest product growth in technology history. |
| 2023 (Mar) | GPT-4 released; Claude 1 released | Significant capability jump; multi-modal (image + text); professional exam performance |
| 2023 (May) | Google launches Bard (now Gemini) | Google enters competitive AI assistant market |
| 2023 (Jul) | Meta releases Llama 2 (open source) | Powerful models available to run locally, outside proprietary platforms |
| 2023 (Nov) | OpenAI board crisis; GPT-4 Turbo; Assistants API | AI governance crisis; significant capability and price improvements |
| 2024 (Feb) | Sora (text-to-video, OpenAI); Gemini 1.5 Pro | Video generation arrives; 1 million token context window (can read an entire book) |
| 2024 (Mar) | Claude 3 (Haiku, Sonnet, Opus) | Anthropic's strongest release; Opus matches or exceeds GPT-4 on many benchmarks |
| 2024 (May) | GPT-4o; Google I/O AI announcements | Multimodal real-time conversation; AI integrated into Google search at scale |
| 2024 (Sep) | OpenAI o1 "reasoning" models | New paradigm: models that "think before answering," dramatically improving complex reasoning |
| 2024 (Oct–Dec) | Claude 3.5 Sonnet v2; Gemini Flash 2.0; Llama 3 | Continued rapid capability improvements; efficiency gains dramatically reduce costs |
What this timeline shows: The pace from GPT-3 (impressive but limited) to genuinely multimodal, reasoning-capable AI assistants was approximately four years. Each year involved improvements that, in previous decades of technology, would have taken a decade or more.
What Exponential Change Feels Like
Exponential growth is famously difficult for humans to intuit. We expect progress to be linear — a steady improvement from year to year. When growth is exponential, the early stages feel slow and the later stages feel overwhelming. By the time most people notice the change, a lot of it has already happened.
The classic illustration: If you take 30 linear steps, you travel 30 metres. If you take 30 exponential steps (each step doubling), you travel over a billion metres — enough to circle the Earth 26 times.
AI does not follow a perfect doubling curve, but the general dynamic applies. Capabilities that seemed out of reach (coherent long-form writing, complex code, natural conversation, image generation, reasoning about novel problems) have crossed from "impossible" to "routinely available" in a span most people did not expect.
What this means practically:
- Things that seem impossible today may not take as long as you expect
- Assumptions about what AI "cannot do" have a short shelf life — check them regularly
- The rate of change in the field means that tutorials, comparisons, and capability assessments from 18 months ago may already be significantly outdated
The right mental model is not "AI improves steadily, like software updates" but "AI capabilities are broadening and deepening in ways that regularly cross new qualitative thresholds."
A Realistic 2–3 Year View
Forecasting AI is genuinely difficult — researchers who have been working in the field for decades consistently report that progress has surprised them. That said, some directions are clearer than others.
Near-certain in the next 2–3 years:
-
AI becomes embedded in most professional software. The period of "AI as a separate tool you visit" is transitioning to "AI as a feature of every tool you already use." Microsoft Copilot, Google Workspace AI, Notion AI, and similar integrations are early examples of a general trend.
-
Costs continue falling significantly. The cost per token for AI inference has fallen by approximately 100x in two years. This trend is likely to continue, making AI economically accessible in contexts where cost was previously prohibitive.
-
AI agents become more prevalent. Rather than responding to a single prompt, AI agents complete multi-step tasks autonomously — browsing the web, writing and running code, interacting with external services. Early versions exist; more capable and reliable versions are coming.
-
Multimodal capabilities expand. Models that can understand and generate text, images, audio, and video — and reason across all of them — will become standard rather than frontier.
Genuinely uncertain:
-
Artificial General Intelligence (AGI). AGI is loosely defined but usually means AI that can perform any cognitive task a human can, with comparable or greater capability. Estimates from leading researchers range from 2–3 years to decades. Treating AGI timeline forecasts as reliable would be a mistake.
-
Economic disruption. The degree to which AI displaces versus augments work is genuinely unknown and probably varies enormously by sector, role, and geography. Confident predictions in either direction should be held loosely.
-
Regulatory effects. The EU AI Act is now in force; US, UK, and other regulations are developing. How these shape the pace and direction of development remains to be seen.
Key takeaway: The next two to three years will bring meaningful capability advances and broader integration into everyday tools. Confident predictions of either utopian transformation or catastrophic disruption within a specific timeframe are likely wrong. Steady, informed attention is more useful than either dismissal or anxiety.
What This Means for You
Stay curious, not anxious. The people who thrive alongside rapid technological change are not those who predicted it most accurately — they are those who stayed engaged, continued learning, and adapted their mental models when evidence required it.
Update your assumptions regularly. If you concluded six months ago that "AI can't do X," check again. That conclusion may still be right, or it may have been overtaken by developments. The capability assessments that matter are current ones, not past ones.
Focus on fundamentals over features. Tools will change. The underlying skills — prompting effectively, thinking critically about outputs, knowing when AI helps and when it doesn't — are durable. Invest in those.
Avoid apocalyptic or utopian narratives. Both have proven reliably wrong as guides to near-term AI development. The reality is more mundane and more genuinely interesting: useful tools, arriving faster than expected, with real trade-offs and real benefits that vary by person and context.
Practice Task
Look back at your own assumptions about AI from one or two years ago. If you made predictions — things you thought AI would be able to do or not do by now — how did they turn out? This exercise is not about being right or wrong. It is about calibrating how to hold AI predictions lightly, regardless of who makes them.