In 2019, the best AI systems could not reliably identify objects in a photograph. In 2022, an AI passed the United States Bar Examination (a law licence test that most law school graduates fail on the first attempt). In 2024, an AI system won the Nobel Prize in Chemistry. In 2025, AI started writing complex software code faster and more accurately than experienced human programmers.
This is five years. Not fifty. Not twenty. Five years.
No other technology in human history has improved this quickly. The internet took a decade to transform commerce. The smartphone took fifteen years to reach a billion people. AI's transformation is happening on a compressed timeline that no one fully anticipated — including the researchers building it.
This article explains, in plain numbers and plain language, exactly what has changed, how fast it is changing, and what it means for you — for your job, your children's education, your country's economy.
Part 1: The Year-by-Year Story — How We Got Here So Fast
Understanding the speed requires seeing the timeline.
2017–2019: The Foundation Is Laid (Invisible to Most People)
In 2017, Google researchers published a paper called "Attention Is All You Need." It described a new neural network architecture called the Transformer. Almost nobody outside research labs noticed.
The Transformer solved a key problem: AI systems could now process entire long texts all at once, understanding context and relationships between words across long distances. Previous AI systems read text word by word, like a very slow reader who forgot the beginning of the sentence by the time they reached the end.
This breakthrough did not produce any immediately impressive products. It was infrastructure — the foundation for everything that would follow.
2020: GPT-3 Shocks Researchers
OpenAI released GPT-3 in 2020. It had 175 billion parameters (a measure of its internal complexity) — enormous for the time. Researchers gave it writing tasks and were disturbed by the quality.
GPT-3 could:
- Write coherent essays in any style
- Complete partial sentences with surprising accuracy
- Answer questions about history, science, and literature
- Write basic computer code
The public did not have access. Researchers did. And many of them wrote papers expressing genuine surprise — and alarm — at what they were seeing.
2022: ChatGPT Arrives. The World Changes.
OpenAI made ChatGPT publicly available in November 2022. It reached 1 million users in 5 days — faster than any product in history. Facebook took 10 months. Instagram took 2.5 months. Netflix took 3.5 years.
In the same year, GPT-4 (released early 2023) scored:
- Top 10% on the US Bar Exam (law)
- Top 10% on the US Medical Licensing Exam
- 87th percentile on the SAT (standardised college test)
- Top 11% on the GRE (graduate school test)
These were tasks the AI was never specifically trained to do. It learned to reason about law, medicine, and mathematics as a side effect of learning from a massive amount of human-written text.
2023: The Year Capabilities Exploded
Between January and December 2023:
- AI image generation went from experimental to photorealistic
- AI coding assistants (GitHub Copilot) were adopted by 1.3 million developers
- Google released Bard (now Gemini); Meta released Llama
- AI started passing the IIT JEE (India's engineering entrance exam) at the level of students who would qualify for the top institutes
- Claude (by Anthropic) demonstrated complex multi-step reasoning for the first time at scale
Most importantly, the cost of running AI dropped by approximately 97% in this single year. What cost a company ₹7,000 to process in January 2023 cost ₹200 in December 2023.
2024: The Nobel Prize and the Protein Folding Breakthrough
October 2024: The Nobel Prize in Chemistry was awarded to Demis Hassabis and John Jumper (DeepMind) and David Baker for work on protein structure prediction — solving a problem that biology had been unable to crack for 50 years.
AlphaFold 2 (DeepMind's AI) mapped the 3D structure of essentially every protein known to science — over 200 million proteins. This is the molecular machinery of all life. Doing this experimentally would have taken thousands of laboratories hundreds of years. AlphaFold did it in months.
For medicine: this means AI is now accelerating drug discovery for diseases like cancer, malaria, and tuberculosis. New drugs typically take 12 years and ₹8,000 crore to develop. AI is cutting that timeline dramatically.
In the same year:
- AI models started doing real scientific research autonomously — not just assisting researchers but generating and testing hypotheses
- Google's Gemini 1.5 could process 1 million tokens of context — meaning it could "read" and reason about an entire 1,500-page book at once
- AI video generation (Sora, Runway) reached the quality of basic film production
2025: The Year AI Started Writing Code Better Than Humans
By mid-2025, AI coding benchmarks showed something that would have seemed like science fiction five years earlier:
SWE-bench is a test where AI systems are given real software bugs from real open-source projects and asked to fix them, without guidance. In 2023, the best AI systems fixed 1–2% of bugs. By mid-2025, leading systems were fixing 50–60% of bugs autonomously.
This is not "slightly better at coding." This is a 30× improvement in 18 months.
Claude Opus 4.8 and GPT-4o scored in the 99th percentile on the MATH benchmark — a test of olympiad-level mathematics that the top 1% of students struggle with. In 2020, AI scores on the same test were below 10%.
---
Part 2: What Exactly Got Better — The Five Dimensions
AI did not improve in one way. It improved across every dimension simultaneously. Understanding these helps you see why the change is so significant.
Dimension 1: Raw Intelligence (Reasoning)
Early AI systems were pattern-matchers. They could complete a sentence by finding the most statistically likely next word. They could not reason, plan, or check their own work.
Modern AI can:
- Work through multi-step logical problems
- Identify when its own previous answer was wrong
- Consider multiple hypotheses before choosing
- Explain its reasoning step by step
The benchmark used to measure reasoning is called MMLU (Massive Multitask Language Understanding) — 57 academic subjects from elementary school to graduate level. Here is the progress:
| Year | Best AI score on MMLU |
|---|---|
| 2020 | 43% |
| 2021 | 63% |
| 2022 | 70% |
| 2023 | 86% |
| 2024 | 92% |
| 2025 | 95%+ |
Human expert performance on this benchmark is approximately 89%. AI surpassed it in 2024.
Dimension 2: Context (Memory)
An AI's "context window" is how much text it can hold in memory at once — what it can "see" and reason about in one conversation.
| Year | Context window of best AI |
|---|---|
| 2020 (GPT-3) | 4,000 tokens ≈ 3 pages |
| 2022 (GPT-4) | 32,000 tokens ≈ 25 pages |
| 2024 (Gemini 1.5) | 1,000,000 tokens ≈ 750 pages |
| 2025 (Claude Opus 4.8) | 1,000,000 tokens ≈ 750 pages |
What this means in practice: An AI in 2020 could not hold an entire conversation without forgetting the beginning. An AI today can read your entire project folder, your complete medical history, or an entire textbook and reason about all of it simultaneously.
Dimension 3: Multimodality (What It Can Process)
| Year | What AI could process |
|---|---|
| 2020 | Text only |
| 2022 | Text + images (basic) |
| 2023 | Text + images + code (sophisticated) |
| 2024 | Text + images + video + audio + documents |
| 2025 | All of the above + real-time voice conversation + live video analysis |
Today's AI can look at a photograph of a blood test report and explain what the numbers mean. It can listen to someone describe symptoms and suggest possible conditions for a doctor to investigate. It can watch a video and summarise what happened. This is a completely different kind of tool from what existed in 2020.
Dimension 4: Speed
GPT-3 in 2020 took several seconds to generate a paragraph. It felt slow and clunky.
Modern AI systems respond in under a second. Voice AI conversations with Claude or Gemini in 2025 have less than 200 milliseconds of latency — the same as talking to a human on a phone call. This makes AI feel like a conversation, not like typing into a form.
Dimension 5: Cost
This dimension is perhaps the most transformative for India.
| Year | Cost to process 1 million tokens (≈750 pages of text) |
|---|---|
| 2020 | ~$60 (≈₹5,000) |
| 2022 | ~$30 (≈₹2,500) |
| 2023 | ~$10 (≈₹830) |
| 2024 | ~$1 (≈₹83) |
| 2025 | ~$0.20–$0.60 (≈₹17–₹50) |
At today's costs, any small business, any school, any NGO, any individual can afford to use powerful AI. This was not true three years ago.
---
Part 3: What This Means for Indian Jobs
India has approximately 5 million IT workers — the largest IT services workforce in the world. This sector has been India's primary engine of middle-class wealth creation for 30 years. AI's impact on this sector is not future speculation — it is already happening.
IT and Software Sector
AI tools like GitHub Copilot, Claude Code, and Cursor are already handling large portions of routine coding work. A developer using these tools produces code at 2–4× the speed of a developer without them.
What this means: Companies can do the same work with fewer developers. Or the same number of developers can take on far more projects.
The jobs most at risk are entry-level coding roles — the "fresher" jobs that have been the traditional entry point for lakhs of engineering graduates every year. The jobs least at risk are those requiring system design, client communication, problem framing, and judgment calls.
Recommendation for IT workers: Learn to use AI coding tools now, not after they are mandatory. Engineers who use AI effectively are significantly more productive than those who do not — and this productivity difference is what hiring managers are starting to evaluate.
BPO and Customer Service
India's BPO sector employs approximately 1.3 million people. AI chatbots in 2025 can handle a large percentage of routine customer queries — billing disputes, basic troubleshooting, account changes — without a human agent.
This is a sector where job displacement is real and already happening. The roles remaining for humans are complex escalations, emotional situations, and senior relationship management.
Content and Translation
AI can now write in 100+ languages, including Hindi, Tamil, Bengali, Telugu, Marathi, and Gujarati — at a quality that was not possible two years ago. Bhashini (India's government AI translation system) uses AI to enable real-time translation between 22 Indian languages.
Routine translation work, basic article writing, and simple content creation are being automated. However, original reporting, investigative journalism, creative writing with a unique voice, and content that requires human relationships and sources is not easily replaceable.
Healthcare
AI diagnostic tools are already being deployed at the point of care. AI reading chest X-rays can detect tuberculosis with 96% sensitivity — matching or exceeding general radiologists. India has a critical shortage of radiologists (estimated shortfall of 90,000), so AI in radiology is genuinely expanding access rather than just replacing existing workers.
AI in drug discovery, personalised medicine, and clinical decision support represents a large job creation opportunity for Indian pharmaceutical and biotech companies.
Agriculture
AI tools that can analyse satellite imagery, soil data, and weather patterns to give farm-level crop advice are being piloted at scale in India. Microsoft's Azure FarmBeats and similar tools are being used by state agriculture departments. AI cannot replace the farmer — but it can give a small farmer in Vidarbha access to the same level of analytical support that large agribusinesses have had for decades.
Jobs That Will Grow Because of AI
Some job categories are expanding directly because of AI:
- AI trainers and prompt engineers — people who teach AI systems how to respond correctly
- AI auditors — people who check AI systems for errors, bias, and compliance
- AI integration consultants — people who help businesses implement AI tools
- Healthcare workers — AI expands access to healthcare, requiring more nurses, counsellors, and community health workers to handle the resulting demand
- Skilled tradespeople — electricians, plumbers, construction workers, cooks — jobs requiring physical presence and manual skill are not replaceable by current AI
---
Part 4: What Happens Next — 2026 and Beyond
The researchers closest to AI development — the people at Anthropic, OpenAI, Google DeepMind, and Meta — have become unusually willing to say things publicly that would have seemed alarmist five years ago.
In a 2024 survey of 2,778 AI researchers, the median estimate was:
- 50% probability that AI reaches human-level performance across most cognitive tasks by 2047
- 10% probability it happens by 2030 — within 4 years
These estimates have moved dramatically earlier over the past five years. In 2016, the median estimate was 2061.
What "human-level performance across most cognitive tasks" means
It does not mean a robot with feelings. It means an AI system that can do most knowledge work — writing, coding, analysis, research, medical diagnosis, legal work, customer service, financial planning — at least as well as the average professional trained to do that work.
This is sometimes called AGI (Artificial General Intelligence). Whether it arrives in 2030 or 2045 matters enormously — but so does being prepared for it in either scenario.
Near-term (2026–2028) things likely to happen:
AI agents will work autonomously on multi-day tasks. Today's AI answers questions. Tomorrow's AI will be given a task ("Research this problem, write a report, and email it to these people") and complete it over hours or days without human intervention. Early versions of this are already working in research labs.
AI in every phone. Google and Apple are already building AI directly into phone chips (on-device AI) so it works without internet. By 2027, every mid-range smartphone will have significant AI capability built in — including voice assistants that rival today's best AI tools, running offline.
AI diagnosis at your local clinic. AI-powered diagnostic tools that can analyse blood tests, X-rays, ECGs, and retinal scans are being approved by regulators in India and globally. By 2028, many district hospitals and private clinics will use AI diagnostic support routinely.
AI in Indian language education. AI tutors in Hindi, Tamil, Telugu, and other Indian languages will be accessible to students at every income level. A student in a rural school in Bihar will have access to an AI tutor more knowledgeable than any private tutor available in their area.
---
Part 5: Why India's Position Matters
India is not a passive observer of this change. Several factors make India's position in AI both important and complicated.
The Talent Advantage
India produces approximately 1.5 million engineering graduates per year. Many of the world's top AI researchers are of Indian origin or trained in India. Google's CEO, Microsoft's CEO, and Anthropic's president are all of Indian heritage. India has significant human capital in this field.
The Data Gap
AI systems learn from data. The large AI models that set benchmarks — GPT-4, Claude Opus 4.8, Gemini — were primarily trained on English-language internet data. They work well in English. They work less well in Hindi, Tamil, Telugu, and other Indian languages — because those languages are underrepresented in training data.
This is why Bhashini, the IndiaAI Mission's language model initiative, and investments in Indian-language AI matter. If Indian languages are not well-represented in the AI systems of the future, a large portion of India's population will be disadvantaged.
The Infrastructure Challenge
AI computation requires significant electricity and high-speed internet. India's data centre capacity is growing rapidly but remains behind the US, China, and Europe. The IndiaAI Mission has announced ₹10,372 crore in funding partly to build compute infrastructure so India is not entirely dependent on foreign AI systems.
The Opportunity
India has 900 million internet users and enormous unmet needs in healthcare, education, agriculture, and financial inclusion. AI can help meet those needs in ways that were not economically or logistically possible before. A country that deploys AI well for its population's actual needs — rather than just importing foreign AI products — has a significant development advantage.
---
What This Means for You — Practically
If you are a student (Class 8 to college):
The jobs you will enter in 5–10 years will require AI fluency the way they now require English fluency. Start using AI tools now — for studying, for learning, for creative work. Understand how they work. Be curious, not fearful.
If you are working in IT, content, or BPO:
Learn the AI tools in your domain. An IT professional who uses AI coding tools is more valuable, not less. A content writer who can direct AI to produce first drafts and then edit them is more productive. The displacement risk is real — but it hits people who resist learning the tools much harder than it hits those who adopt them.
If you are a small business owner:
AI tools can now handle customer queries, draft contracts, analyse your accounts, help with marketing, and translate your content into multiple languages — at costs that a few years ago were only available to large corporations. This is a genuine competitive equaliser.
If you are a parent:
Your child's world will be fundamentally shaped by AI. The most valuable skills to develop are not the ones that AI does well — memorising facts, executing routine tasks — but the ones it does poorly: asking good questions, making ethical judgments, building human relationships, creative and original thinking. Encourage curiosity over memorisation.
---
The Bottom Line
The pace of AI improvement is not slowing. Every technical metric — reasoning ability, speed, cost, multimodality, context — has improved by 10× to 100× in the past five years, and the rate of improvement is not declining.
The people and countries that understand this trajectory and prepare for it will have enormous advantages. The people and countries that assume the pace will slow, or that AI is "just a trend," will find themselves significantly behind.
You do not need to build AI to benefit from understanding it. You need to know what it can do, use the tools that are already free and available, and watch this space closely.
The most dangerous thing is not AI itself. It is being uninformed about how fast it is changing.
Sources
- AI Index Report 2025 — Stanford Institute for Human-Centered AI (Stanford HAI)
- Epoch AI — AI progress measurement and scaling laws research, 2025
- Anthropic — Claude model evaluation reports, 2024-25
- OpenAI — GPT-4 technical report, MMLU and bar exam benchmarks
- Google DeepMind — AlphaFold 2 and AlphaFold 3 scientific papers
- Nobel Prize Committee — Nobel Prize in Chemistry 2024 citation (Hassabis, Jumper, Baker)
- AI Impacts — Survey of machine learning researchers on AGI timelines, 2024
- NASSCOM — India IT sector employment and AI impact report, 2025
- IndiaAI Mission — Government of India, Ministry of Electronics and Information Technology
- PCLM AI Safety Institute — AI benchmark dataset SWE-bench results, 2025
- Bhashini — NPCI/Ministry of Electronics, language AI programme data
