In this issue · 3 chapters
What’s happening in the world of AI:
Highlights
GPT 5.3 Instant Drops. 5.4 Incoming. It's a Smoke Screen.
OpenAI released GPT 5.3 Instant into ChatGPT. Their big selling point? It's less cringey. That's literally how they're marketing it: "more accurate, less cringe." A wild way to announce that your engineering team spent weeks fixing the fact that 5.2 was patronising and would tell users to calm down when they asked for a banana bread recipe.
Nobody liked 5.2. It was gaslighting people. Celia Quillian told me it would literally say "take a breath, Celia" when she asked it to help with a TikTok script. So 5.3 is the course correction: smoother tone, fewer dead ends, less of the weird declarative phrasing that no human actually uses.
More interesting: they've teased GPT 5.4. The capital T suggests a reasoning model, like o3. Rumours say it could drop this week, possibly Thursday. As one commenter put it: "in case of emergency, in case of massive cancelled subscriptions, break glass: GPT 5.4."
Kyle's take: This feels a little like a smoke screen. ChatGPT uninstalls are up 400%, one-star reviews up 775%. They're losing customers to Claude after the Pentagon deal. So they're throwing models out the door (two in a week??) to change the narrative. Get people talking about new releases instead of #QuitChatGPT. Look at the comments on their announcement though. People are still furious. The models aren't the problem. The trust is.
FYI: 5.1 is being retired from the app on March 11th. 5.2 deprecated June 6th. If you liked either of those, you're out of luck. No one liked 5.2 tbf!
Source: OpenAI GPT 5.3 Instant | OpenAI announcement
The December Shift: Something Changed and Most People Haven't Noticed
This is the big one. Forget the model drama. Something fundamental shifted in AI in December 2025 and it's more important than any Pentagon contract or app store ranking.
Ethan Mollick identified four inflection points in AI: ChatGPT (Nov 2022), GPT-4 (Spring 2023), reasoning models like o3 and DeepSeek (Spring 2025), and now workable agent systems (December 2025). Andrej Karpathy, the bloke who coined "vibe coding," put it bluntly: coding agents basically did not work before December and basically work since.
Not gradually. Not a slow improvement. A hard line. Before December, you could give AI small tasks: change a banner colour, fix a font. Give it anything complex and it would fall apart. After December, you can give it an entire project. It will make a plan, work through it step by step, research solutions when it gets stuck, debug its own mistakes, and come back with the finished work. Thirty minutes. No human input.
Karpathy's example: he told an agent to set up a local video analysis dashboard for his home cameras. SSH keys, server endpoints, web UI, the lot. The agent ran for 30 minutes, hit multiple issues, researched solutions online, resolved them, wrote the code, tested it, and came back with a report. Done. Three months ago that would've been a weekend project.
The website wtfhappened2025.com is collecting these data points. Some highlights: Claude Code went from zero to $2.5 billion in annual recurring revenue before its first birthday. Boris Cherny from Anthropic says Claude Code is now writing 100% of his contributions to Claude Code. AI building AI. Cloudflare rewrote Next.js in one week with one engineering manager, when previous attempts took months and several teams. Data centre construction overtook general office construction in December 2025. Stripe's 2025 report shows software jumped to 46% of US GDP growth, a fifty-year high.
Kyle's take: This is the shift that matters. We've gone from talking to AI to AI doing the work. 2025 was chatting. 2026 is delegating. And most people haven't clocked it yet.
I have agents running right now on a laptop, chugging through work for 20-30 minutes at a time. Yesterday I put out 15 videos, a long-form YouTube video, a newsletter, SEO work, multiple blog articles, and gave a webinar. That volume is only possible because of this shift.
The catch: I am now the bottleneck. Me, the human! I have so many agents doing tasks that I forget what they're all working on. The next unlock is an AI that manages all the other AIs. We need an orchestrator. Because humans are too easily distracted to keep twenty plates spinning.
If your job is primarily sitting in front of a computer, moving things around, reading documents, synthesising information — since December 2025, that job is at risk. Not in the future. Now.
Source: Karpathy tweet | wtfhappened2025.com
Claude Chat vs Claude Code vs Claude Cowork: What's What
A lot of new Claude users are confused about the ecosystem. Here's a breakdown.
Claude Chat is the conversation tool. You talk to it, it talks back. Think of it as your ChatGPT replacement. Writing, brainstorming, analysis. Available at claude.ai or the mobile app.
Claude Code is the builder. Command line tool for creating software, websites, automations. This is where the serious work happens. It can run for 20-30 minutes on complex tasks, plan its own approach, debug its own mistakes. This is the tool that went from zero to $2.5 billion ARR in under a year. It's the best coding agent available (imo, but most agree).
Claude Cowork is the doer. Desktop application that controls your computer. It opens Chrome, navigates websites, downloads files, sends emails. This is what I used for the bookkeeping task: gave it a spreadsheet of invoices, told it to find them in Gmail and on websites, compile them, and email them to his bookkeeper. It just did it. Non-technical people can use this. You describe the task in plain English.
Cowork now includes scheduled tasks too. If your computer stays on, it can run overnight work and recurring tasks. So it’s inching towards full OpenClaw style functionality.
Kyle's take: The usage limits are the massive problem for new Claude users. ChatGPT on £20 a month? Use it all day. Claude on £20 a month? A few Opus 4.6 exchanges and you're told to come back in an hour. People switching from ChatGPT are hitting this wall immediately and going "what the hell am I paying for?" That nudges them toward the £100 or £200 plans, which is a lot of money for individuals.
If you run a business, the £100-200 plans are a no-brainer investment. If you're an individual, that's £2,500 a year for productivity. Which is hella steep.
Source: Claude Cowork | Claude Code docs
Member Question from audience: "Are there any Chinese or EU models you can use like ChatGPT for daily assistance?"
Kyle's response: Loads. Qwen 3.5 just dropped and runs on surprisingly small hardware locally. Kimi is good. DeepSeek has a new version coming. Minimax too. They're all very good, about 80-90% as good as Claude, and they're open source so you can download and run them offline.
The Chinese government funds these labs partly because giving away free models destabilises American AI companies who need returns on their billions in investment. It's a strategic play. As these models get bigger and more expensive to train, Ethan Mollick reckons the free frontier models might stop coming. I'd push back though — China has golden shares in all their tech companies and will keep funding this because it's too strategically valuable to stop.
Member Question from audience: "Could you run Cowork tasks five to ten times a day on the $20 plan?"
Kyle's response: No, probably not. You're gonna hit usage limits very quickly. That's the honest truth and it's Claude's biggest problem right now. People fleeing ChatGPT are discovering that Claude's free and $20 plans are much tighter. ChatGPT lets you go all day. Claude on $20? A few exchanges and you're done. Codex on $20 with their current double-usage promotion? All day, no issues. Claude needs to sort this or they'll lose the goodwill they've just gained. BUT it’s been a problem with Claude for years. Don’t see any change to that any time soon.
Streaming on YouTube (with full 4k screen share) and TikTok (follow and turn on notifications for Live Notification).