What’s happening in the world of AI:
Highlights
The UK Government Spent 4 Million Quid on a Website I'm Rebuilding for Free
The UK government's AI Skills Hub (aiskillshub.org.uk) cost taxpayers 4.1 million pounds (about $5.5 million). I've logged in, used it, and I'm annoyed.
@iamkylebalmer What even is this? the UK government just gave PwC £4.1m to build a terrible AI skills site that should have taken a week. the more i look... See more
Where to start. The registration process is broken. The verification email lands in spam and comes from some random address. Once you're finally in, you get asked to pick your industry from a list of about five options: agriculture, transport, construction, creative industries, or "other." If you work in healthcare, education, retail, finance, or literally anything else? Other. Cheers.
The site itself is a catalogue of pre-existing courses from Microsoft, Google, Amazon, and others. Courses that have been freely available online for years. All they've done is dump them on a page with some Shutterstock images slapped on top.
Want to know how much a paid course costs? Tough. It just says "fee payable." Could be fifty quid, could be five grand. They didn't bother to find out. The "personalised pathway" system is equally useless. I set myself up as a beginner leader with no specific industry. It sent me an agriculture-focused data concepts course. Marked as "fee payable" when it's actually free. Brilliant work all around, great job guys.
Oh yeah did I mention this cost £4.1 million?
Kyle's take: I'm rebuilding this. And going to improve it.
Then make it open and accessible. Then spin off versions for other countries.
Codex already grabbed all 500 courses for me. I'm enriching the data, adding actual prices, building a proper assessment system, and making it openly available without a registration wall that doesn't even work.
If the government can charge 4 million for a glorified link directory, I can and will do better for nothing. I also want to expand this internationally. Every country should have access to something like this, done properly, in their own language. So that other nations don’t get scammed like this.
Claude Goes for the Jugular with Super Bowl Ad
Anthropic dropped a Super Bowl ad mocking ChatGPT's upcoming adverts. It’s superb:
A bloke's talking to his therapist about communication problems with his mum. The therapist gives him the usual "start by listening" advice, then suddenly pivots into flogging a dating site for older women. It's a spot-on impression of what ads in AI conversations might feel like.
They also published a statement: Claude is a space to think. "There are many good places for advertising. A conversation with Claude is not one of them." They're committing to keeping Claude ad-free. No sponsored links, no advertiser influence on responses, no product placements. Full article here.
This is a direct attack on OpenAI, who are starting to roll out ads in ChatGPT's free and low-cost plans.
Kyle's take: The ad is brilliant marketing. But Sam Altman's response actually makes a strong case. He pointed out that more Texans use ChatGPT for free than the total number of Claude users in the entire US. OpenAI serves nearly a billion people. That costs serious money to keep running, and ads help keep it free for everyone.
Anthropic can afford to say "no ads" because they don't have the traffic to make ads worthwhile. Their monthly visitors are about 15 million compared to ChatGPT's 400 million. Claude's money comes from developers and people like me paying 200 quid a month for the max plan. Different business models for different problems.
The fact that Altman wrote a whole essay in response to a joke ad tells you it really got under his skin. But his argument holds up. I think he's dead to rights here, even if his essay was a bit much.
Google's Quiet Domination: 750 Million Gemini Users and Counting
Google dropped their quarterly results and there are some massive numbers in here. Annual revenue exceeded $400 billion for the first time. Their AI stack is firing on all cylinders. But the headline that caught me was this: the Gemini app has 750 million monthly active users. Excusez-moi??
ChatGPT sits at about 900 million. Google could overtake them as the biggest AI app on earth this year.
While Anthropic and OpenAI have been publicly scrapping over Super Bowl ads, Google have been quietly building in the background. They make their own TPUs. They have their own cloud infrastructure. Chrome has 70-80% browser market share. Android has similar penetration. They've got decades of data from Search, YouTube, Maps, and everything else. They have money coming out of their ears from the existing business.
As a fun aside I was at Google’s offices yesterday and the only photos I took were TPUs… I know right? I’m that cool.
They've also (quietly) rolled out full AI agent capabilities in Chrome. I saw it in action at the Google office in London. Someone's Gemini sidebar took over their browser and booked a train ticket, going all the way to the checkout. It was like a game of chicken watching how far he trusted it -and whether it would just plug his credit card details in! Luckily it stopped and waited for his input!
People have been releasing agentic browser all through 2025. Comet. Atlas. Remember those? And I always said that as soon as Google add AI into Chrome it’s over. I stick to that. 70%+ market penetration remember.
The Gemini sidebar can now also use Nano Banana to generate images based on what's on the page. Looking at a flat on Rightmove? Ask it to show you the place fully furnished.
Kyle's take: I've said it before and I'll say it again. Google will probably win the AI race because they own the whole vertical, top to bottom. The 750 million figure is the one to watch. If that keeps growing at this rate, they overtake ChatGPT this year. And they've got the cash reserves to sustain it without needing ads in the chatbot, at least for now. They will add ads eventually. You don't see them saying they won't. But they're in no rush.
New Model Rumours: Sonnet 5 and GPT-5.3
Rumours swirling about Anthropic releasing Sonnet 5. Leaked benchmark tables have been floating around showing it beating everything from 2025 across the board. Take these with a grain of salt. They're probably fake.
The way Claude's model releases work: Opus is the premium model but expensive. Then Sonnet catches up and overtakes it. Then a new Opus comes along. Rinse and repeat.
Sonnet 5 was apparently meant to drop on Tuesday but Anthropic had a massive outage. The rumour is they tried to launch, the servers collapsed, and they rolled it back. A bunch of vibe coding apps like V0 and Replit had been teasing "something big coming" around the same time, then went quiet. Adds up.
Meanwhile, someone asked Sam Altman about GPT-5.3 and got a cheeky side-eye response. So that's on its way too.
Kyle's take: There will always be a new best model. Then another one a few weeks later. And another. As long as the trajectory keeps going up, that's all you really need to know. You don't need to chase every release. You don't need exact benchmarks. It's all getting better. That's enough.
Last minute update: Opus 4.6 dropped. No sign of Sonnet 5 yet.
One Year of Vibe Coding
Turns out “vibe coding” was a shower thought he fired off without much consideration, and somehow it stuck. It's on his Wikipedia page now as a "major memetic contribution" and the article about vibe coding is apparently longer than his own biography. He's the co-creator of ChatGPT and has done loads of other things, but he'll probably be remembered for this.
A year ago, vibe coding was for throwaway weekend projects. Fun little toys. Now? Programming via LLM agents is becoming a default workflow for working professionals. Even Linus Torvalds, who created the Linux kernel, has been using it on side projects.
Karpathy's suggesting we rename it to "agentic engineering" to differentiate the professional use from the amateur stuff. Agentic because you're orchestrating agents rather than writing code. Engineering because there's a skill to it.
Kyle's take: "Agentic engineering" doesn't have the same ring. I don't think it'll stick. Sorry Andrej! Your first naming was just too good!
What will happen is that vibe coding just becomes... coding. Same way we don't call it "non-punch-card coding" anymore. The abstraction layer moves up. The name drops away. Give it a couple of years.
Codex vs Claude Code: I'm using both. Codex is slower (20-40 minutes per task) but the results come back polished. Claude Code is faster and better for the back-and-forth stuff. I'm using Codex for big projects and security audits. Claude Code for iterative daily coding. Still working out how they fit together.
AI Authority Accelerator: I'm launching a 30-day programme to help you build a following in AI content for your industry or demographic. About 150 quid. Launches Monday. Can learn more here.
Want the full unfiltered discussion? Join me tomorrow for the daily AI news live stream where we dig into the stories and you can ask questions directly.
Streaming on YouTube (with full 4k screen share) and TikTok (follow and turn on notifications for Live Notification).
