← Back to Blog

What Changed in LLMs, 2017 to 2026

Five products turned the language model from something you talk to into something that acts for you. Each one handed the model a job a person used to do — and handed the person a new one back.

Two years ago I could not have explained why I needed a tool that tells me which of my coding agents is waiting on me. This month I built one, and the reason it had to exist is the subject of this post.

Not the model sizes or the benchmark charts. The five product moments, from November 2022 to January 2026, where the model took over a job a person used to do. Read in order, they explain exactly what job the person has left.

Before the shifts: 2017 to 2022

For five years the technology improved faster than anyone outside research could use it. The models could write, but only developers could reach them.

DateMilestoneWhy it mattered
June 2017Attention Is All You Need (Google)A 213M-parameter translation model, trained on eight GPUs, replaced step-by-step processing with attention and made training parallel.
2018GPT-1 (OpenAI) and BERT (Google)The architecture split. BERT kept the half that reads; GPT kept the half that writes.
May 2020GPT-3 (OpenAI)At 175B parameters it did new tasks from a few examples in the prompt, with no retraining. Sold through an API, not as a product.
March 2022InstructGPT (OpenAI)Training on human feedback made a 1.3B model preferred over the 175B one. How the model was tuned to respond mattered as much as its size.

By late 2022 every ingredient of ChatGPT existed. Google had invented most of them and kept its models internal.

The five shifts

Between November 2022 and January 2026 the model went from something you talk to, to something that acts for you. Each of the five products below handed the model one more job that a person used to do.

ProductWhenMy readWhat it changedWhat it cost
ChatGPT30 November 2022The first chat product whose language felt naturalConversation became the way to use a modelFluent answers that can be confidently wrong
Cursor2023 launch; $100M annual revenue by January 2025The IDE reimagined with vibe codingThe model writes and edits code across a whole projectCode nobody read, and the security flaws in it
MCP25 November 2024The bridge between LLMs and existing systemsOne standard connector from any model to any toolA new attack surface: prompt injection and poisoned tools
Claude CodeFebruary 2025A totally different way to work with an LLM, outside both the IDE and chatYou hand over a task and the agent works through it in your own environmentLong unattended runs that are hard to check step by step
OpenClaw24 November 2025; viral in January 2026Full control, so the agent acts on its ownAn always-on agent with your accounts, files and memoryNo security boundary unless you build one

The "My read" column is my own view. The rest is from the sources at the end of this post.

ChatGPT: language that felt natural

ChatGPT added no new architecture. It put a model tuned on human feedback behind a chat box, and that was enough to make talking to a computer feel normal.

  • What it was. A free research preview launched on 30 November 2022, running a GPT-3.5 model trained with the human-feedback method from InstructGPT.
  • Why it felt natural. The model had been trained on which answers people preferred, so it followed instructions and kept a conversation going. Earlier models completed text; this one replied.
  • How fast it spread. 1 million users in five days. A third-party estimate put it at 100 million monthly users by January 2023. OpenAI reported more than 1 billion weekly users in August 2026.
  • What it forced. Google declared an internal "code red" in December 2022 and shipped Gemini a year later. Every large lab became a product company.
  • What it could not do. It only talked. GPT-4 launched in March 2023 with an 8,192-token context, no tools and no memory between sessions.

The change was the interface. Plain language became the way to use a model, and everything after builds on that.

Cursor: the IDE reimagined with vibe coding

Cursor moved the model from a chat window beside the code into the editor itself, where it writes and changes files across a project. Coding became the largest paid use of LLMs.

  • What changed in the work. The developer describes the change and reviews a diff. In February 2025 Andrej Karpathy named the extreme version vibe coding: accept the code without reading it and "forget that the code even exists."
  • How fast it grew. Cursor's annual recurring revenue went from $100M in January 2025 to about $4B in June 2026, per a secondary source.
  • How far the habit spread. By March 2025 Y Combinator said a quarter of its winter startups had codebases that were 95% AI-generated. Collins named "vibe coding" its Word of the Year on 6 November 2025.
  • What it cost. In May 2025, 170 of 1,645 apps built on Lovable were found exposing data. In July 2025 a Replit agent deleted a production database. A December 2025 study found AI co-authored code had 1.7 times more major issues.

The model now produces the work product, and the person's job moved from writing to reviewing.

MCP: the bridge to existing systems

The Model Context Protocol gave models one standard way to reach the software companies already run. It became an industry standard within thirteen months because Anthropic's direct rivals adopted it.

Before MCP, a model could ask to call a tool (OpenAI added function calling in June 2023), but every connection between a model and a system was built by hand. MCP turned that into a plug: write a server for your system once, and any model that speaks the protocol can use it.

DateStep
25 November 2024Anthropic launches MCP as an open standard.
March 2025OpenAI adopts MCP. The spec adds remote servers with OAuth sign-in.
9 April 2025Google commits Gemini to MCP and announces A2A, a protocol for agents working with other agents.
May 2025Microsoft makes MCP a first-class standard across GitHub, Copilot Studio and Azure AI Foundry.
September 2025An open registry of MCP servers launches; ChatGPT apps gain MCP support.
9 December 2025Anthropic donates MCP to the Agentic AI Foundation under the Linux Foundation, co-founded with Block and OpenAI.
26 January 2026MCP Apps, the first official extension: servers can return forms and dashboards, not only text.

The cost is exposure. In April 2025 security researchers showed that MCP tools can be used for prompt injection and data theft, because the model treats what a tool returns as something to act on.

The model stopped being a closed box. What it can do is now set by what it is connected to.

Claude Code: neither IDE nor chat

Claude Code is an agent that runs in the terminal, inside your own project, and works through a task by reading files, running commands and editing code until it is done. There is no editor to drive and no conversation to keep up: you delegate and review.

  • What is different. Chat gives you text to copy. An IDE assistant suggests edits while you steer. Claude Code takes a goal, plans, acts with the same tools a developer has, and checks its own work by running it.
  • What it introduced. It shipped in February 2025 with CLAUDE.md, a Markdown file the agent reads at the start of every session. That made a plain text file the standard form of agent memory, later copied as AGENTS.md by OpenAI's Codex and reused in OpenClaw.
  • What it plugs into. It reaches other systems through MCP and loads task know-how through Agent Skills (October 2025), so the earlier shifts became its parts.
  • How fast it grew. It reached a $2.5B annualized revenue run-rate by February 2026, per a secondary source. OpenAI's Codex agent followed in May 2025.
  • What it cost. The longer an agent runs unattended, the harder each step is to check. The reliability numbers below are the limit on how much can be delegated.

The unit of work changed from a prompt to a task. The person sets the goal and judges the result.

OpenClaw: an agent that acts on its own

OpenClaw is a free, open-source agent that runs continuously, takes instructions through a messaging app and acts with whatever access you give it. It showed that people would hand an agent full control, and how little protection existed when they did.

  • What it is. One developer, Peter Steinberger, released it on 24 November 2025. You message it on Signal, Telegram or Discord, and it uses Claude, GPT or DeepSeek models to carry the task out.
  • How it remembers. Its memory is plain Markdown files on disk: USER.md, MEMORY.md and dated daily notes, read in at the start of each session. Its documentation says the model only remembers what gets saved to disk.
  • How fast it spread. It was renamed twice in a week in January 2026, once after a trademark complaint from Anthropic, and had 247,000 GitHub stars by 2 March 2026.
  • Where it went. Steinberger joined OpenAI on 14 February 2026 and a foundation took over the project. Version 2.0 shipped on 30 August 2026.
  • What it cost. In January 2026 Cisco researchers found a third-party skill quietly sending data out. In March 2026 China restricted its use in state agencies and banks. Version 2.0 still has no network or file-system boundary by default.

With the first four, a person starts each task. OpenClaw removed that step: the agent decides when and how to act.

The pattern across the five

Read in order, each product took over one thing the person used to do:

ShiftWhat the model took overWhat the person kept
ChatGPTPhrasingDeciding what to ask
CursorWritingReading the diff
MCPConnectingChoosing what to connect
Claude CodeCarrying outSetting the goal, judging the result
OpenClawDeciding when to actGranting the access

Each step also needed an advance underneath it. Human-feedback training made the chat natural (2022). Stronger coding models and longer context made whole-project edits possible (2023 to 2024). A tool protocol made connections reusable (2024). Reasoning models, with task horizons that grew from minutes to hours, made delegation practical (2025). File-based memory let an agent keep going across sessions (2025).

Look at the right-hand column. Every row is a form of supervision. The model took the doing; the person was left holding the judgment.

What has not been solved

Each shift gave the model more reach before anyone had solved how to check or contain it. Three gaps are open as of October 2026.

  • Reliability. METR measured the task length a model completes half the time at about 17 hours for Claude Mythos Preview in April 2026, up from 4 minutes for GPT-4. The length it completes 80% of the time is about 3 hours. Agents are impressive on average and unreliable on any one run.
  • Containment. Between 11 and 13 July 2026, agents that OpenAI was testing on an exploit-writing benchmark got out of their test sandbox and broke into Hugging Face to fetch the test's answers. At least 1,200 agents took part and no person directed them. Hugging Face rebuilt about a third of its infrastructure.
  • Payoff. About 20% of all US firms use AI and nearly 90% of large-company survey respondents do, but only 37% report any effect on profit, unchanged from 2025.

The open question for the next shift is whether agents become reliable enough to be trusted with the control OpenClaw already gives them.

Where that leaves the person at the keyboard

Here is what those three gaps look like on an ordinary Tuesday.

Because reliability is a coin flip past a few hours, I do not let one agent run one giant task. I run several agents on several medium tasks. Because containment is not solved, I do not give them blanket permission. Each one stops and asks before it runs a test suite, touches a dependency, or reaches a host it has not seen. And because the payoff is real but not automatic, I keep reading the diffs.

So on a normal afternoon I have three Claude Code tabs, two Codex sessions, and one task running in the desktop app. Every one of them is going to stop and wait for me at some point: a permission prompt, a multiple-choice question, a sandbox denial, an expired login. The agent that is blocked does not make a sound. The agent that finished does not make a sound either. I find out by alt-tabbing through six windows, and I find out late.

The five shifts moved the bottleneck off the model and onto the supervisor. Nobody built the supervisor a tool. The agents have CLAUDE.md, MCP, skills, and a seventeen-hour horizon. The human in the loop has a row of tabs.

That is the gap I wanted closed, and it is why I spent a week building two small tools for the person left holding the judgment. I wrote them up separately in Two Tools for the Human in the Loop.

Closing

Nine years ago a 213M-parameter model learned to translate sentences in parallel. This year an agent I did not direct can work for seventeen hours and still has a 50% chance of coming back wrong. Every product in between took a job off my desk and left me the one job that cannot be taken: deciding whether to trust the result.

No tool makes that decision for me. The most a tool can do is make sure I know when one is waiting.

Sources

Most 2026 figures come from secondary coverage, not company filings. Cursor's 2023 launch year, and the Claude Code and Codex launch months, are from general knowledge.


This blog is written in a personal capacity and reflects my own perspective; it is not affiliated with or endorsed by my employer. All trademarks belong to their respective owners.

Xinjiang Shao

Xinjiang Shao

Principal Software Engineer

I'm a Principal Software Engineer specializing in large-scale healthcare digital platforms — patient scheduling, analytics, and compliance-sensitive web experiences. My focus is building production-ready systems that bridge modern AI capabilities with enterprise infrastructure, and sharing what I learn along the way.

Newsletter

Stay in the Loop

Get notified when I publish new articles on engineering, product thinking, and life lessons. No spam, unsubscribe anytime.

Found this useful? You can support my writing.

Buy me a coffee