People say a lot has happened since ChatGPT came out. I beg to differ.
Yes, everything works much better now, and we’ve had model improvements, a fleet of new startups and lots and lots of discussion.
But for anyone hoping to get their head around things, I’d argue there’s really only a handful of big developments.
Here’s the cheat sheet:
Reasoning models made answers 10x better. Nearly two years after ChatGPT launched, OpenAI released o1. Instead of immediately answering, it generates a chain of thought around the question first, adding context and going deeper.
This opened the door to multi-step problem solving – but also arguably, more effective tool use, and therefore the agentic opportunity.
Code generation became real. Building a product is about finding people with a real problem, building features that make a massive difference and working out distribution. The code itself has become a background element, and this is one of the most revolutionary moments in my entire career.
Should you ship your Lovable app? No.
Is it enough to let you validate your business? Yes yes yes yes yes yes yes yes yes yes yes yes yes yes yes.
Open-source quietly thrives. In the AI startup I was working in in early 2023, we thought models will become commoditised. I still believe that. From DeepSeek to Qwen to Llama, unless they really do get an AGI breakthrough, it’s hard to see how it doesn’t go to the mean, with all the consequences that entails. Model companies need advantages elsewhere.
Visual AI crossed the uncanny valley. It works now. I personally don’t find it that exciting – but it’s probably massive in the way visual social media like Instagram and TikTok were 10x bigger than a platform requiring thought like Twitter could ever be.
Honourable mention for voice here – massive companies with big valuations that no doubt will be used all over the place. (But just not very exciting to me personally…)
2026 and beyond – the fifth development
I feel like today, Claude Code does the work, while I do the thinking.
The potential became most clear to me when I had a good chat with Christian Ryan, one of Anthropic’s applied AI team, at a hackathon.
I was running a product idea past him and he suggested just using Claude Code’s agentic harness to bootstrap it. No need to mess around with other agentic systems for prototyping, just create subagents and you’re 90% there.
Not long after, Claude output styles came along, then plugins and – this was the moment where it clicked for me – Claude Skills.
Claude Skills now let you package up subagents, hooks, scripts and docs for templates etc in a zip file that anyone can drag into Claude Desktop and use an agentic workflow.
I’m just wrapping up a project where I helped the Marketing team in a £2bn tech company build a Claude Skill that writes a whole campaign in a box (8+ documents, 16000 words) in 30 minutes. And I think I could get the time even shorter.
It goes without saying that the writing isn’t perfect. Nor should it be. It’s just there to produce the kind of draft that leads to the right kind of feedback – the kind where you can see the opportunities and add the human magic.
This morning the first thing I did at my desk was ask it what I should focus on this week. Yes, I kind of know the answer – but it’s the kind of admin an assistant might do. And it’s the first step to then enriching those thoughts, providing research to accompany it or tying it back into the bigger picture of the month/ year each time.
I would give up my mobile phone before I gave up Claude Code at this point.
And I think across 2026, we are really going to see the benefits of casual agentic work kick in for those who show a little interest. (and certainly for anyone working with me…)