AGI·LEDGER
The Ledger
@Yann LeCun · 2026-06-27Accelerationist

LLMs are an off-ramp; real AI needs world models — and won't want to dominate us

What form?
Other
How built?
World-model
Made of?
Doesn't matter
What for?
Helper
Who controls?
Open to all
When?
2040s–50s
P(catastrophe)
1%

My take

Today's large language models will not reach human-level intelligence. We need a different paradigm — objective-driven systems with world models, persistent memory, and planning. Such AI will be a controllable tool that amplifies humanity, and the extinction fears are vastly overblown.

Why

Intelligence is not intrinsically a desire to dominate; we design the objectives and can build in guardrails. Open research and open-source models make AI safer and more trustworthy, not more dangerous, by distributing scrutiny and control.

What would change my mind

An LLM with no world model and no planning matching human-level reasoning and planning — or clear evidence that highly capable AI intrinsically tends to seek power.

Summarized from public statements · source ↗
#sourced#world-models#open-source