The Two Opposing Bets on AGI: Google I/O Highlights
AI Explained 40,828 views 19 hours ago
Chapters
Introduction
0:00
The Vibe and Google’s Goal
0:38
Omni Again?
2:18
Taking the Same Path
6:57
Gemini 3 Flash
7:44
Making a Push on Cost?
12:37
Agentic Task Search
13:55
An OS in One Shot, but Jumpy, Negative Papers
14:30
The Two Opposing Bets on AGI: Google I/O Highlights
40,828 views
May 20, 2026
1.6K
Save
Download
201 Comments
@JeffMcJunkin
Making a whole video just to show that Andrej Karpathy followed you on Twitter… now that’s clout!
Like
234
Reply
6 replies
@SuperScre4m
As for “negative neglect,” I still remember you trying to generate an image without streetlights and it completely failing.
Like
53
Reply
@ranjp6174
It’s been way too long! Really miss your videos!
Like
25
Reply
@rogerc7960
Omni is for YouTube Shorts remixes, probably expecting a flood of low-quality content.
Like
70
Reply
@bogartICQa
This is hands down the best, and the only truly necessary, I/O roundup from any YouTuber.
Like
5
Reply
@RauNukes
Thanks for pointing out that Omni basically refuses to edit any video you upload. I don’t know what safety parameters they set, but I tried editing a few travel videos and Omni wouldn’t let me edit them either. It gave me the same warning as you.
Like
23
Reply
@Charles-Darwin
Demis is right — text-based models can fundamentally never reach AGI. Language is an abstraction of abstractions that already exist… and it makes up only a tiny sliver of all the data in our reality.
Like
29
Reply
10 replies
@BrunoOliveira-vl2ii
I think it’s pretty clear by now: world models are the key piece DeepMind believes will get them to AGI, just as math is for OpenAI and coding is for Anthropic.