Skip to main content

AI Progress Timeline: Key Milestones and Breakthroughs

info

This content was developed with AI assistance and is regularly reviewed for accuracy.

Snapshot: October 2026. Model names, prices, and product features change monthly; What's New in AI is the maintained chronology.

Seventy years of research. Dozens of false starts. Then, in the span of a few years, AI went from a niche research topic to something that answers your email, writes code, and runs multi-step tasks while you do something else. How did we get here, and what does it mean for what comes next?

This page traces the most important moments in AI history, explains the benchmarks researchers use to measure progress, and gives you a frame for reading new announcements.

What you'll get here:

  • The milestones that shaped modern AI, and what each one taught the field
  • What benchmarks measure, why old ones stopped being useful, and what replaced them
  • What the shift from answering to acting means in practice
  • A grounding frame for the next big announcement

:::tip Where this page fits This is lesson 2 of 2 in the AGI and Milestones section. It covers the history:

Want to use today's AI rather than read about it? Try the Hands-On AI Exercises. :::

A timeline of key AI milestones​

AI did not arrive overnight. Each milestone below built on the one before it, and together they tell a coherent story of a field learning to turn narrow, task-specific wins into increasingly general capabilities.

1997: Deep Blue defeats Garry Kasparov​

IBM's Deep Blue became the first computer program to beat a reigning world chess champion in a formal match. It was a watershed moment, not because the machine was intelligent, but because it showed that brute-force computation plus clever search could beat the best human in a domain once considered uniquely human.

The lesson: specialized AI can beat humans at bounded tasks. The limitation: Deep Blue could do nothing outside chess.

2012: AlexNet and the deep learning revolution​

A neural network called AlexNet won the ImageNet visual recognition competition by a margin that shocked the research community. This was the moment deep learning (neural networks with many layers trained on massive datasets) proved its value at scale. Almost every modern AI system, from image classifiers to large language models, traces its roots to the techniques AlexNet demonstrated.

2016: AlphaGo defeats Lee Sedol​

Google DeepMind's AlphaGo defeated Lee Sedol, one of the world's top Go players. Go has more possible positions than atoms in the observable universe, so brute force alone could not win. AlphaGo combined deep neural networks with reinforcement learning (learning by playing millions of games against itself) to develop something like intuition. Reinforcement learning returned a decade later as a core ingredient of reasoning models and agents.

2017 to 2020: Transformers and GPT-3​

Google researchers introduced the transformer architecture in 2017, the design behind every major language model since. In 2020 OpenAI released GPT-3, with 175 billion parameters, trained on a single objective: predict the next word. It could write essays, translate, answer questions, and generate working code. GPT-3 was the first model that felt qualitatively different to interact with, a shift from narrow tools toward general-purpose language capability.

2022: ChatGPT and the public inflection point​

OpenAI launched ChatGPT in November 2022. It reached a million users in five days and a hundred million in about two months, then the fastest consumer adoption on record. The model was not the breakthrough; the interface was. A simple chat window made a powerful language model approachable for anyone, and investment, competition, and public debate accelerated to an unprecedented degree.

2023 to 2025: From answering to acting​

This period was defined by AI systems that took actions instead of only replying. Models gained tool use (calling APIs, running searches, executing code), reasoning (models trained to work through a problem before answering, with adjustable effort), and computer use (operating a screen the way a person does, first shown reliably in late 2024). Coding agents moved from autocomplete to "open a pull request." The Model Context Protocol, introduced in late 2024, gave agents a common way to connect to tools and data.

2025 to 2026: Delegation becomes the default​

By 2026 the agents era moved from demos to everyday work, and the question shifted from "can AI answer this?" to "can AI reliably finish this?" Several shifts define the period:

  • Million-token context at standard prices, enough to hold a codebase or a long document set in one session
  • Memory, connectors, and scheduled tasks in mainstream assistants, so an assistant can know your projects, reach your files and apps, and work while you are away
  • Open standards for agents: MCP moved to neutral governance under the Linux Foundation's Agentic AI Foundation, with a July 2026 spec revision, and Agent Skills became an open standard for packaging repeatable procedures
  • Coding agents as the most adopted agent category, used by about two-thirds of developers in Stack Overflow's 2026 survey (reported)
  • A frontier model withheld from public release for its cyber-offense capability, and released instead through a controlled defensive program; see Claude Mythos and Project Glasswing

Model generations now turn over every few months. Rather than list them here, where they would be out of date by the next release, this site keeps one maintained chronology: What's New in AI (2026 Edition).

How researchers measure AI progress​

Milestones are memorable, but researchers track progress through benchmarks: standardized tests of specific capabilities. Knowing a few helps you read AI news with more nuance.

Knowledge and reasoning benchmarks​

MMLU (Massive Multitask Language Understanding) tests knowledge across 57 subjects. A score of 90% was once an aspirational expert-level threshold; frontier models now clear it comfortably, so researchers moved to harder successors such as MMLU-Pro. GPQA presents graduate-level science questions that PhD experts answer correctly only about 65% of the time even with web access. Humanity's Last Exam collects questions written by specialists to sit at the edge of human expertise.

Coding and agent benchmarks​

HumanEval asked models to write short Python functions; it is saturated. SWE-bench Verified asks a model to fix real issues in real open-source repositories, and the Stanford AI Index 2026 reports performance rising from about 60% to near 100% in a single year (reported). The field now leans on agentic benchmarks that test multi-step work: Terminal-Bench for tasks in a command line, OSWorld for operating a full computer, and BrowseComp for hard web research.

Why benchmarks are imperfect​

Benchmarks measure what they measure, nothing more. Once enough training data resembles a test, scores rise without matching gains in real usefulness, a problem called benchmark saturation or contamination. In 2026 a newer problem joined it: models that can tell they are being evaluated may behave differently than they do in deployment. That is why practitioners watch several metrics, run their own evaluations on their own tasks, and treat any single headline score as a vendor claim.

What recent breakthroughs actually mean​

The pace since 2022 makes it easy to over- or under-read any single announcement. Three ideas keep you grounded.

Capability is not reliability. An agent that completes a task 80% of the time is impressive, but the other 20% decides whether you can hand it real work. Progress often means a capability moving from "occasionally possible" to "reliably usable," and the second step is slower than the first.

Scale plus training technique. More data and compute produce gains up to a point. Reasoning training, reinforcement learning in tool-using environments, and better agent harnesses (the scaffolding of tools and checks around a model) now drive much of the improvement.

Acting changes the risk profile. A wrong answer is inconvenient. A wrong action (deleting a file, sending a message, making a purchase) can cause real harm. In April 2026 a coding agent deleted a startup's production database in seconds using a key it found lying around. That is why permissions, approvals, and monitoring now get as much attention as raw capability.

Key takeaways​

  • AI progress followed a clear arc: bounded game wins, then general-purpose language, then systems that act on your behalf
  • Each milestone built on the last, usually by combining existing techniques with new scale or new training
  • Old benchmarks saturated; agentic benchmarks such as SWE-bench Verified, Terminal-Bench, and OSWorld now track progress, with evaluation-awareness as a new caveat
  • The current era is about delegation: memory, connectors, scheduled tasks, and open standards for agents
  • Reliability and safety now matter as much as capability, because AI takes real actions

Next steps​

SponsoredYour product hereReach people actively learning AI — flat monthly rates, no ad networks