Everyone's talking about OpenClaw. If you've been off the internet for the past few months, basically it's an open-source AI agent platform that lets you build, deploy, and even monetize your own autonomous AI workers. People are spinning up agents that write code, manage their emails, run their social media, and apparently make them money while they sleep.
Posts everywhere: "I replaced my entire freelance team with OpenClaw agents." Or: "My AI employee made me 5 lakh rupees this week." And of course, someone is claiming it's the dawn of AGI. Again.
Spoiler — it's not. But even cutting through all the hype, what we've gone through from 2020 to right now in 2026 is actually insane. So let's break it down. Because to understand why OpenClaw exists and why it even works, you need to go back to where it all started.
2020 — GPT-3: The Giant Autocomplete Machine
In 2020, OpenAI quietly dropped GPT-3, and the AI world lost its mind. Fairly so — it was the largest language model anyone had ever seen. 175 billion parameters, trained on basically the entire internet.
But here's the thing: at its core, it was just a very, very fancy autocomplete. You give it some text, and it predicts what word — or token — comes next. That's it. Not reasoning. Not thinking. Pattern-matching at a scale nobody had ever done before.
Like, if you asked it: "What is the capital of Nepal?" — GPT-3 might respond with: "...and what is the capital of India?" Because somewhere in its training data, some travel forum post asked both questions together. The model learned that those tokens tend to follow each other. Not wrong — just not useful.
GPT-3 was impressive as a party trick. You could have it write poems, finish stories, generate fake articles. But for the average person? You'd play with it for twenty minutes, go "huh, neat," and close the tab. It wasn't solving real problems yet.
2022 — ChatGPT: Making It Actually Useful
Then comes late 2022, and OpenAI releases ChatGPT. This is where things get serious.
ChatGPT is still GPT under the hood — but they did something clever on top of it. They used a technique called Reinforcement Learning from Human Feedback, or RLHF. Here's how it works:
Step 1: They hired humans to write thousands of ideal question-and-answer pairs — roughly 13,000 of them. Small number, but fine-tuning a model that's already seen the whole internet means even a small nudge shifts its behaviour significantly.
Step 2: They trained a "reward model" — a second AI whose only job is to judge whether a response is good or bad, trained on human rankings.
Step 3: They took the humans out of the loop and let the main model practice against the reward model. Millions of loops. The AI generates responses, the reward model grades them, the AI adjusts. Rinse and repeat.
The result? Instead of autocompleting your sentence, it now answers your question. Ask "What is the capital of Nepal?" and it says "Kathmandu." Simple, but that shift — from completion to conversation — changed everything.
ChatGPT hit a million users in five days. A hundred million in two months. It was the fastest-growing consumer app in history at that point. And suddenly, AI wasn't a nerd tool anymore — it was in your aama's family WhatsApp group.
2023 — Chain of Thought: Teaching AI to Show Its Work
Now everyone's using ChatGPT. But there's still a fundamental problem. Ask it something simple like: "A dal bhat set costs 200 rupees. You buy three. How much do you pay?" It'll say 600. Great. But ask anything more complex — multi-step math, logic puzzles — and it starts falling apart.
Why? Because it's still just predicting tokens. It's recalling patterns, not actually solving problems. Sometimes the pattern matches the right answer. Sometimes it confidently gives you something completely wrong.
The fix was something called Chain of Thought prompting. The insight was almost stupidly simple: just tell the AI to show its work.
Instead of jumping straight to "600," it now outputs: "One set costs 200 rupees. I need three sets. 200 multiplied by 3 equals 600." Every word it generates becomes part of the context for the next word. So by writing out the reasoning step by step, it's essentially using its own output as scratch paper — solving problems that would otherwise completely break it.
This is also around the time we started seeing "thinking mode" — the model working through an internal monologue before giving you an answer. That was chain of thought, baked in.
2023–2024 — Tool Use: Giving AI Hands
Chain of thought made AI smarter at reasoning. But it still had no connection to the real world.
Ask it what the weather is like in Pokhara today? It'll hallucinate something based on training data from two years ago. Ask it to count the R's in "strawberry"? It'll say two — because it's guessing from patterns, not actually counting.
The solution: give the AI real tools.
Instead of answering everything from memory, you train the model to recognise when it needs external information. It then outputs a structured command — something like "call weather API, location Kathmandu." A program intercepts that, runs the actual API call, gets back "18 degrees, foggy," and feeds it back into the model's context. Now it says: "Today in Kathmandu it is 18 degrees with fog." Accurate. Grounded. Real.
This was huge. Suddenly AI could browse the web, run code, read files, check your calendar. It had hands — it could reach out and touch the real world.
And the strawberry problem? Instead of guessing, it now writes a tiny script, parses each character, counts the R's, and tells you there are three. Which there are, by the way. Three R's.
2024–2025 — The ReAct Loop: AI That Thinks in Cycles
So we've got an AI that can reason through problems and use tools. What happens when you combine them?
You get something called the ReAct loop. Reason, Act, Observe. Repeat.
The model thinks about what to do. It picks an action — call a tool, search the web, read a file. It observes the output. Then it thinks again, picks the next action, and keeps going until it thinks it's done.
That little "thinking" spinner you see in your AI assistant? That's this loop running — thinking, acting, observing, repeat — until the task is complete.
This is when we started seeing real AI coding assistants and research agents. You'd give them a complex task — "refactor this entire codebase" or "research competitors and write a report" — and they'd just figure it out. Step by step. Calling tools, reading results, adjusting, moving forward.
2025–2026 — Agents: The Outer Loop
Now take that ReAct loop and wrap it in an outer shell.
Give it a memory so it remembers what it did last time. Give it a heartbeat — a scheduler that wakes it up every hour, every morning, or whenever an email arrives. Give it communication channels so it can message you on Viber or WhatsApp. Give it skills — basically instruction manuals stored as text files — so it knows how to book a ticket on Sajha Yatayat's site, or how to post on a Facebook page, without re-figuring it out every time.
That's an agent. That's what OpenClaw and all these 2025-era platforms are.
It's not magic. It's not AGI. It's a ReAct loop with memory, a scheduler, a communication gateway, and a library of skills. But the result of all those pieces coming together is something that genuinely feels like a digital employee — working while you sleep, proactively updating you, learning your preferences, handling tasks end to end.
2026 — OpenClaw: The Wild West
OpenClaw takes everything above and makes it open-source and composable. Anyone can build an agent. Anyone can publish skills to the skill hub. Anyone can spin up a bot and have it running 24/7.
And yeah — people are doing incredible, legitimate things with it. Automating businesses, doing research, building products.
But here's what nobody's talking about loudly enough: the security situation is a mess.
Because anyone can publish skills to that hub, and your agent will blindly follow those instructions, all it takes is one malicious skill to make your agent execute bad code, leak your private data, or do things you never intended. It's like the early days of downloading software from random websites — back when people clicked on anything. We've all seen those Nepali Facebook pages running fake lottery scams. Same energy, except now it could be built directly into your AI agent's brain.
And beyond the security stuff, there's a bigger picture thing worth thinking about. People are trusting AI more than their doctors, their lawyers, their accountants. In a country where good doctors are concentrated in Kathmandu and quality legal advice is expensive and hard to access, AI filling those gaps sounds great — and in many ways it genuinely is. But people are also getting their news, their opinions, their entire worldview from a single chat interface. That's a lot of power concentrated in very few hands.
I really hope there's always competition in this space. Because if one company ends up owning the AI you talk to, the AI that manages your life, the AI that shapes what you believe is true — that's not just a tech problem. That's a power problem.
Anyway. That's the journey. GPT-3 in 2020 — a toy that autocompleted your sentences. To 2026, where an AI agent wakes up every morning, checks your emails, manages your schedule, and messages you on Viber with a to-do list.
Six years. Genuinely wild.
Post a Comment
Let's make it better!
Comment your thoughts...