Use AI for everything. Verify everything.
Point it at hard problems.

Nobody hires juniors anymore. And the rest of office work is next.
The biggest cost in most products is people's time.
Great as a consumer. Brutal if you sell your time — or a product whose price falls to zero.
Yes, I think there's a bubble. No, it doesn't change what you should do.
Raw intelligence of top models has more or less peaked.
So labs specialise — and smaller, cheaper models beat yesterday's frontier.
64GB kit: ~€170 in mid-2025 → €926 today. Data centres are buying all the memory.
Shut down its consumer brand to sell to AI companies instead.
New consumer GPU generations — a first in about 0 years.
Worst case: OpenAI and Anthropic disappear tomorrow morning.
Chinese labs have released open models, close to the frontier and good enough for most uses.
Open = download it, modify it, host it anywhere. You can't un-invent that.
Bad news first. No sugar-coating. Then the good news.
AI writes the code, so technical people got cheaper. Anyone can build — so everyone does.
Products are easy to copy and the internet lets the winner reach everyone. Winner takes most.
Flooded with "vibe-coded" products. Every email well written. Every landing page beautiful.
Polish means nothing now. Same for consumers: infinite apps, less chance they pick yours.
Building is cheap. Getting noticed is expensive.
First 100 users, first 10 customers: now a huge budget and a huge effort.
Sounds depressing. But it isn't the end of the story.
If easy things are worthless, hard things are worth more than ever.
Domain knowledge · hardware · regulation · private data · real relationships.
The only place left where competition isn't infinite.
No manual. No SOP. Just people who know.
You can't automate what isn't written down. Years of interviews, mapping, change management, resistance.
Engineers who sit inside the business, learn the work hands-on, and build the automation right there.
Give context and examples. Let it draft, you edit. Then ask it to critique like a tough professor.
Notes → outline → argue with the outline → slides. You're looking at one.
Scripts, storyboards, voiceovers, subtitles, editing — and now the footage itself.
Personas, competitors, positioning, 10 ad variants. Role-play your toughest customer.
Even if you've never written a line. Landing pages, prototypes, small automations.
Ported generative-AI inference engines — video, voice cloning, text analysis — to my favourite language. For each one, independently, it:
AI is not infallible. It hallucinates: stats, quotes, court cases, papers, links.
With total confidence. It doesn't know it's wrong.
If it cites a source, open the source. Check it exists and says what the AI claims.
Check every number that matters.
If it writes code, run it and test it. Don't assume it works.
Read everything before it goes out with your name on it. If it's wrong, it's on you.
You don't need to spend money to start.