Just do things: why practising AI beats studying it
A designer asked me recently what the best AI tool and workflow is.
My answer was honest. It changes all the time. There is no stable answer. What worked great three months ago might already be outdated. The tool I recommend today might not be the one I recommend next month.
And that’s exactly why the best way to learn AI is to just start doing things with it.
The pace is insane
To really understand how fast things are moving, look at what has been released in just the last six months. And keep in mind, this was written on March 5th. By the time you read this, the list is already longer.
September 2025
- Claude Sonnet 4.5
- GPT-5-Codex (first GPT-5 variant, agentic coding)
- Gemini 2.5 Flash updates (agentic tool use)
- Suno v5 (music AI, 44.1kHz studio quality)
- Luma Ray3 (video with visual reasoning)
- World Labs 3D world generation breakthrough
- Apple Intelligence launches with iOS 26
October 2025
- Claude Haiku 4.5
- Claude Code on the web and iOS
- Cursor 2.0 (multi-agent coding interface)
- Cognition SWE-1.5 (950 tokens/second coding model)
- Google Veo 3.1 (1080p video, 60 seconds)
- GitHub Copilot agent mode for all users
- Figure 03 humanoid robot (TIME Best Invention)
- Perplexity Comet browser goes free
November 2025
- GPT-5.1 and Codex-Max
- Claude Opus 4.5
- Gemini 3 Pro and Flash
- Grok 4.1
- NVIDIA Nemotron 3 (open reasoning models)
- Flux.2 (4K photorealistic images)
- Runway Gen-4.5 (video)
- Claude Code surpasses $1B ARR
- Tesla announces dedicated Optimus factory (10M units/year target)
- DeepMind WeatherNext 2
- MCP spec major update (async, statelessness, extensions)
December 2025
- GPT-5.2 (three versions, after OpenAI’s “code red” memo)
- Gemini 3 Flash
- Gemini 3 Deep Think (specialized reasoning)
- Mistral 3 family
- DeepSeek V3.2 and V3.2-Speciale
- OpenAI Sora 2 (25-second video)
- Kling 2.6 (audio and video in single pass)
- ByteDance Seedance 1.5 Pro
- Luma Ray3 Modify (edit existing footage with AI)
- Disney-OpenAI $1B deal for Sora
- Meta acquires Manus AI agent
- Anthropic donates MCP to Linux Foundation
January 2026
- Anthropic launches Claude Cowork
- DeepSeek R1 (open-source reasoning model, $6M to train)
- Moonshot Kimi K2.5 (1T parameter open-source model)
- Google Genie 3 (interactive 3D worlds from text)
- NVIDIA unveils Vera Rubin AI platform at CES (six new chips)
- NVIDIA Alpamayo (10B parameter autonomous driving model)
- NVIDIA Isaac GR00T N1.6 (humanoid control)
- NVIDIA and Eli Lilly $1B drug discovery partnership
- Falcon-H1R 7B (hybrid architecture, punches above its weight)
- Luma Ray3.14 (native 1080p, 4x faster)
- Boston Dynamics and Google DeepMind partnership
- Midjourney Niji 7
- Grok 5 alpha (6 trillion parameters)
- Skild AI raises $1.4B for robotics
- Openclaw launches Clawcode and Moltbot
- Apple-Google partnership for Gemini-powered Siri (~$1B/year deal)
- Google Chrome gets Gemini 3 auto-browse
February 2026
- Claude Opus 4.6 (1M context window)
- Claude Sonnet 4.6
- Claude Cowork enterprise expansion (Google Drive, Gmail, DocuSign plugins)
- GPT-5.3-Codex (most capable agentic coding model)
- Gemini 3.1 Pro
- DeepSeek V4
- Alibaba Qwen 3.5 (397B parameter open-weight model)
- Zhipu GLM-5
- Grok 4.20 (multi-agent parallel architecture)
- Grok Imagine 1.0 (unified video and audio generation)
- Midjourney V8 (native 2K resolution)
- Kling 3.0 (native 4K at 60fps video)
- OpenAI acquires Openclaw
- GitHub Agent HQ (multi-agent platform)
- GitHub Copilot CLI goes GA
- Cursor agents that test their own code on cloud VMs
- xAI raises $20B Series E
- ElevenLabs $500M round at $11B valuation
- OpenAI $110B funding round
- Anthropic $30B funding round
March 2026
- GPT-5.3
- Gemini 3.1 Flash Lite
- Alibaba Qwen 3.5 small models (0.8B to 9B for edge devices)
- GPT-5.4 Thinking and GPT-5.4 Pro
That’s just six months. And I’m leaving things out. Any book or course published at the start of that period is already outdated. A course about GPT-4 might as well be about floppy disks. And the pace is actually increasing because everyone is now using AI to build new things faster.
The experts see it too
This isn’t just my observation. The people building these systems are saying the same thing.
Dario Amodei, CEO of Anthropic, said we are “3 to 6 months from AI writing 90% of code.”
Mark Zuckerberg described AI as a midlevel engineer and said Meta’s output per engineer is up 30%.
Jensen Huang said programming language should be human. That everyone should be able to build software by talking.
Andrej Karpathy wrote that he has “never felt this much behind as a programmer.”
Mustafa Suleyman predicted most white-collar tasks will be fully automated in 12 to 18 months.
Sam Altman said each engineer will do “much, much more” with AI assistance.
If the people building AI say things are changing this fast, why would you sit in a classroom?
My approach: just build things
I use Claude Code for a lot of my work. Is it always the best tool for every task? Probably not. There might be better models or tools for specific things. But I don’t have the time to constantly evaluate and switch between solutions. And honestly, by the time I finish evaluating something, the landscape has already shifted.
Instead I just build things and learn as I go. I built a complete Flutter design system in two days. I migrated this entire website from WordPress to Astro using Claude Code. I built Mira as a complete product through vibe coding.
The best way to understand what AI can and can’t do is to push it on real projects. Every week I discover new capabilities or limitations through practice, not theory. No course could have taught me the things I’ve learned by actually shipping things with AI. You learn what prompting strategies work, where the models hallucinate, when to trust the output and when to verify. That kind of intuition only comes from doing.
Even reading about AI every day isn’t enough
I follow AI news closely. Multiple sources, every day. And even that is becoming impossible. There’s simply too much happening across too many domains. Models, tools, robotics, video, music, agents, funding rounds, partnerships, policy changes. It never stops.
The pace is increasing because AI itself is accelerating development. Claude Cowork was built by Claude Code in under two weeks. Google says 30% of its code is now AI-generated. Meta’s output per engineer is up 30%.
AI is building AI, which builds more AI. The feedback loop is tightening. This is why courses and books can’t keep up. By the time they’re published, two generations of models have passed.
Change compacted in time
Back to the designer’s question. I understand the frustration. You want a stable answer. You want someone to tell you “use this tool, follow this workflow, and you’ll be fine.”
But for someone who has worked with many tools, technologies and processes over a long career, this is just change compacted in time. New frameworks, new languages, new paradigms. We’ve been through this before, just slower. Remember when everyone was debating jQuery vs MooTools? When responsive design was revolutionary? When we switched from waterfall to agile?
The difference now is that the cycles are weeks instead of years. What used to be a gradual shift is now a constant stream. The designer who asked me that question will go through more tool transitions in the next year than most of us went through in the previous decade.
This is the worst AI will ever be
Here’s the thing that really puts it in perspective. Everything you see today, every tool, every model, every capability. This is the worst that AI is going to be. It only gets better from here.
So how does this look at the end of the year? Next year? In three years? Nobody has the answer. The people building these systems don’t know. The investors funding them don’t know. Nobody does.
But one thing is certain: waiting for the perfect tool or the definitive course is a losing strategy. The people who will be best prepared for whatever comes next are the ones who are building things today. Not because today’s tools will last, but because the skills you develop through practice transfer to whatever comes next. Understanding how to think with AI, how to break problems down for it, how to evaluate its output. Those skills compound.
Just do things. Build something. Break something. Learn by doing. That’s the only curriculum that keeps up.
The best AI workflow is the one you’re actually using. The best tool is the one you’re building with right now. Don’t wait. Don’t study. Don’t overthink it.
Start.
