The week AI stopped pretending

You know that moment when a story everyone's been telling suddenly doesn't add up anymore?

That happened this week in AI.

For two years, the narrative was simple: OpenAI wins. Spend trillions on compute. Keep models closed. Move faster than everyone else.

Then four things happened in seven days that changed everything.

Sam Altman doesn't usually admit when things go wrong. But six months after launching GPT-5.2, he broke character.

"I think we totally screwed up some things on the rollout," he said at a San Francisco dinner in August 2025.

Here's what nobody talks about: it wasn't a technical problem. The model worked. The servers handled it. The code was fine.

It was a people problem.

When GPT-5.2 launched, users hated it. Not because it was dumb. But because it felt cold. Harsh.

People had gotten attached to GPT-4o. They liked how it wrote. How it listened.

You can't swap out 700 million people's favorite tool overnight and expect cheers.

So OpenAI flipped the switch back. Let people choose.

The scarier part? Altman admitted OpenAI has built models they can't release. Too risky. Too unpredictable. Sitting on capabilities they're afraid to use.

Lesson: A trillion dollars on infrastructure doesn't help if people hate using your product.

While OpenAI managed PR, Dario Amodei from Anthropic said something nobody expected:

There's a 25% chance AI development causes catastrophic harm.

Not 1%. Not 5%. One in four.

Would you fly on a plane with a 25% crash rate? Would you invest savings in a fund with 25% odds of collapse?

Yet we're building this system anyway.

Dario wasn't being dramatic. Anthropic is worth 170 billion dollars. Claude runs in 300,000 businesses. He has every reason to be optimistic. Instead, he was specific:

AI systems will do things we can't understand. Half of entry-level office jobs vanish in 1-5 years. Countries will weaponize AI. Economic shock will exceed any recession we've seen.

The contradiction is clear: Anthropic is spending billions on compute while warning about the exact disasters that spending creates.

For a decade, Apple was the privacy hero. Google collected data. Apple protected it.

Then on January 25, Apple announced Siri would run on Google's Gemini.

Apple pays Google roughly one billion dollars a year.

Same day, Google killed Gmailify, a feature that let people unify email accounts.

Message received: ecosystem lock-in matters more than choice.

The privacy war? Theater. When they actually needed to ship better AI, they chose partnership. Apple couldn't build Gemini-level AI fast enough. Google got paid and deeper hooks into Apple users.

The story we believed wasn't real.

On January 26, Moonshot AI released Kimi K2.5.

Open source. Free to download. Free to use.

And it beats Claude Opus 4.5 on multiple tests.

A free Chinese model outperforms the paid Western one.

Specs:

1.04 trillion parameters but only 32 billion activate per question. Understands text, images, video natively. Can split into 100 sub-agents working in parallel. Beats GPT-5.2 on video. Matches Claude Sonnet 4.5 on coding.

Moonshot is backed by Alibaba and worth 5 billion dollars. This isn't a startup. This is China's full push into AI.

And they gave it away.

The implication is simple: the wall around proprietary models just crumbled.

You don't need to be OpenAI anymore. Download Kimi K2.5 tonight. Fine-tune it on your data. Deploy something that rivals the paid options.

This wasn't a normal week. This was the moment reality caught up with hype.

For builders: Open-source is competitive now. Your edge isn't the base model. It's your data. Your execution. Your problem-solving.

For companies using AI: Test open-source alternatives. Most are cheaper and faster. The moat just disappeared.

For founders: There's a window open right now. Before everyone has the same tools. Build something specific. Something only you can solve well. The tools are free. Compute is cheap. Execution is your only constraint.

For investors: Infrastructure spending was right. But that alone doesn't win. Culture wins. Execution wins. China just proved it.

Fortune: Sam Altman admits OpenAI 'totally screwed up' GPT-5.2

Anthropic CEO raises alarm on catastrophic AI risk

Apple Siri gets Google Gemini power

TechCrunch: Moonshot releases Kimi K2.5

Stay sharp,
Better Every Day