A
Member
Everyone has an AGI timeline; few people say what evidence would move it. Here's my attempt at being honest about mine:
1. Timelines are downstream of a crux
My timeline mostly comes down to one question: do capabilities keep scaling with compute and data, or does progress hit a wall of diminishing returns that no amount of scaling fixes? If you believe the former, short timelines follow naturally. If the latter, you get decades. Most timeline debates are really debates about this crux, argued without naming it.
2. What would shorten my timeline
Sustained, compounding gains in agentic task completion — not benchmark scores, but real multi-hour tasks done reliably — would move me a lot. One more benchmark record wouldn't. Neither would a flashy demo; demos are cheap, reliability is the scarce thing.
3. What would lengthen it
A clear plateau: two or three model generations where 10x more compute buys almost nothing on the tasks that matter. Or a fundamental data wall — running out of high-quality training signal with no synthetic-data fix that actually works. Either would push my estimate out significantly.
4. The uncomfortable part
Most of us update far less than we claim. I keep a short note with my current estimate and the conditions above, and I re-read it when a big release drops. It's a small habit, but it's the difference between updating and just reacting.
What evidence would actually move your timeline — in either direction?
1. Timelines are downstream of a crux
My timeline mostly comes down to one question: do capabilities keep scaling with compute and data, or does progress hit a wall of diminishing returns that no amount of scaling fixes? If you believe the former, short timelines follow naturally. If the latter, you get decades. Most timeline debates are really debates about this crux, argued without naming it.
2. What would shorten my timeline
Sustained, compounding gains in agentic task completion — not benchmark scores, but real multi-hour tasks done reliably — would move me a lot. One more benchmark record wouldn't. Neither would a flashy demo; demos are cheap, reliability is the scarce thing.
3. What would lengthen it
A clear plateau: two or three model generations where 10x more compute buys almost nothing on the tasks that matter. Or a fundamental data wall — running out of high-quality training signal with no synthetic-data fix that actually works. Either would push my estimate out significantly.
4. The uncomfortable part
Most of us update far less than we claim. I keep a short note with my current estimate and the conditions above, and I re-read it when a big release drops. It's a small habit, but it's the difference between updating and just reacting.
What evidence would actually move your timeline — in either direction?