Thread by @benaverbook

This is Dario Amodei.

He's the CEO behind Claude, one of the world's most advanced AIs.

Yesterday, in a 5.5 hour conversation with @lexfridman, he revealed our timeline to superintelligence.

Let me save you 5 hours: 🧵

The timeline that changes everything:

Superintelligent AI by 2026-2027.

But that's their conservative estimate; internal data suggests even sooner.

And they just revealed the 5 signs we're closer than anyone thought:

First, a reality check.

Two approaches to building AGI:

• OpenAI: Racing to be first
• Anthropic: Racing to be safe

But Anthropic's CEO just revealed why this race has a twist:

The winner could determine humanity's fate.

What keeps them up at night:

1. Catastrophic misuse (cyber/bio weapons)
2. AI autonomy (systems too powerful to control)

"With great power comes great responsibility."

But here's the terrifying part:

Smart, educated people rarely cause catastrophic harm.

This "natural safeguard" protected humanity for centuries.

But AI breaks this correlation - giving dangerous capabilities to anyone.

Their testing shows this protection is already crumbling.

They've created a system to track the danger:

"ASL levels" (AI Safety Levels) from 1-5

• Current models: ASL-2
• Next year: ASL-3
• 2026: Likely ASL-4

But ASL-3 is the real turning point:

At ASL-3:

• Models can enhance bad actor capabilities
• New security protocols required
• Enhanced filters needed
• Deployment restrictions

"If we hit ASL-3 next year, we're not ready."

But that's not the scariest part:

Their breakthrough was realizing:

We don't program AI, we grow it (@ch402).

Like biology evolving complex systems:

• We create the scaffold
• We create the light
• But the system grows itself

This led to a shocking discovery:

Looking inside neural networks, they found something wild:

Every AI develops universal features, just like biological brains.

The same patterns appear in:

• AI vision models
• Monkey brains
• Human neural networks

Nature keeps reinventing the same solutions.

Even stranger:

Every large language model develops a "Donald Trump neuron"

It's the only personality that consistently gets its own dedicated neuron.

Why? They don't know.

But it reveals something crucial about AI development:

The data problem everyone fears:

"We'll run out of internet data"

Anthropic's solution?

Models teaching themselves through self-play.

This changes the entire timeline:

Their models have gone from:

• High school level (2022)
• Undergraduate level (2023)
• PhD level (2024)

In some tasks, they're exceeding human experts.

But here's the real breakthrough:

Their programming AI went from solving:

• 3% of real tasks (January)
• 50% of real tasks (October)

At this rate, it hits human-level next year.

But they're more worried about what happens at 200%.

These AIs aren't just getting smarter.

They're developing (@AmandaAskell):

• Deep interests
• Unique perspectives
• Complex personalities
• Abstract understanding

Leading to the biggest question:

The real deadline isn't technological.

It's regulatory:

"If we get to the end of 2025 with no meaningful AI regulation, I'm going to be worried."

Time is running out.

Lex and Dario got into much deeper topics concerning AI and especially AGI in the episode.

If you are interested I definitely suggest going and watching the entire episode here:



(p.s. This whole thread was written by Claude 😉)

Speaking of podcasts...

If you love learning from podcasts but struggle to keep up with all the best episodes,

I spend 20+ hours listening each week and send the most valuable insights straight to your inbox.

Join for free: benaverbook.beehiiv.com/subscribe

Thanks for reading! A bit about me:

I'm a 20-year-old student obsessed with learning & sharing the most fascinating ideas & people I discover.

No products, just pure knowledge.

If you enjoyed this, follow @benaverbook for more breakdowns!