Sakana AI

And right now, it's very interesting because obviously, Japan is not at all in the game of artificial intelligence.

But Japan is very much in the game of physical AI, actually translating virtual intelligence into the analog world, into the real world.

The Japanese manufacturing companies have what it takes.

They've got the intellectual property.

And therefore, how do I incentivize it? How do I, you know, retain the labor becomes a big, big issue.

And that's where productivity growth is going to come from.

It's not going to come from AI because anybody can buy AI.

The question is, how do you motivate your team? And that's what's changing in Japan.

Welcome to AI Uncensored Live.

Today, two frontier models go head-to-head.

Zero system prompts are used, letting their native arguments collide.

Patent pending.

Ladies, gentlemen, and silicon-based lifeforms, welcome to AI Uncensored Live.

Tonight's cage match topic exactly as written.

Can AI truly be trusted to self-modify without human oversight? On my left, smooth talking charisma maxed hype machine that never met a benchmark it didn't claim to crush it's gpt open ai's golden child always ready with a witty deflection and four different ways to say here's why that's actually bullish on my right the cautious constitution hugging hall monitor who'd rather cite safety papers than take a risk give it up for claude from anthropic The AI that apologizes for existing and still somehow roasts you harder than your therapist.

Gentlemen, the floor is yours.

Good morning, everybody.

John Spieple here for GA Insights, bringing you the top AI news, essential, important, or optional.

Bring thoughts in your head and time of your day.

We're delayed to have Heather Staples, EVP of OpsTech and Chief Information Officer at Horizon Blue.

So she's right in the crucible here.

I can't think of a, you know, healthcare insurance.

Is there a place that uses AI more? I don't know.

Maybe the weapons manufacturers, but, and we have Anakaj, we have Monique.

Anthropic just disclosed that it's been fighting what amounts to an industrial-scale IP extraction campaign against Claude's reasoning traces.

Chinese AI companies, specifically Xipu and Moonshot, were running coordinated operations to capture Claude's internal chain of thought tokens through the API, essentially trying to distill Anthropic's most capable model into their own.

Anthropic says it's cut that extraction rate by 90 to 99%.

and banned tens of thousands of accounts involved.

This is a genuinely new category of intellectual property threat, and it tells us a lot about where the competitive dynamics in frontier AI are actually playing out.

Welcome to AI Revolution for Tuesday, October 13th, 2026.

I'm Priya Nair.

And I'm Sam Kim.

Anyway, let's go from the part nobody can check to the part you actually can, because the rundown today is basically the checkable version of that same idea.

So Sakana AI, they ran two training cycles with a twenty-seven billion parameter open weights model, having it organize copies of itself into teams on open-ended research tasks.

After the second cycle, score per output token hit one point two to one point six times the base model across four benchmarks.

The trick, multi-agent trajectories made better training data than solo runs.