Two Rival Bets on AGI: Google I/O Highlights

TL;DR

Google I/O revealed two competing AGI paths: Google betting on world models and video generation, while OpenAI bets text-based reasoning alone reaches AGI.

Key Points

• 1.Google's AGI bet centers on world models and video generation. Demis Hassabis argued that simulating the world correctly is a crucial step toward AGI, with Gemini Omni combining video, image, and audio generation as a stepping stone — echoing OpenAI's earlier (now abandoned) claim about Sora.
• 2.OpenAI's rival AGI bet is that text reasoning alone gets there. Greg Brockman stated OpenAI has 'line of sight' to AGI through the GPT reasoning model tree, arguing text intelligence definitively will reach AGI without requiring world models or video generation.
• 3.Gemini 3.5 Flash is fast but not a frontier breakthrough. It matches Gemini 3.1 Pro in intelligence benchmarks, leads in speed (tokens per second), tops Finance Agent V2 and chart-reasoning benchmarks at 84.2%, but trails GPT-4.5 and Claude Opus 4.7 on vibe-coding tasks.
• 4.Google I/O's consumer strategy is integrating 'good enough' AI into search. Sundar Pichai pitched companies on saving billions by switching to cheaper models like Flash, cut the Ultra plan from $250 to $200/month, and admitted agents aren't 'truly helpful' yet in a notable on-stage quote.
• 5.An independent 70-page paper exposed a deep LLM epistemics flaw. Models trained on thousands of documents explicitly prefaced as false — including disclaimers before, after, and within sentences — still learned to fully believe the fabricated claims, affecting GPT-4.1, Quen 3.5, and Kimi K2.5 alike.
• 6.Google DeepMind researcher Mustafa Suleyman called jagged intelligence a structural, unfixable bug. He argued the AI field underestimates how hard jaggedness is to fix and how much it matters, warning that a model brilliant at technical problems but blind to everything else won't drive meaningful scientific progress.
• 7.Andrej Karpathy joining Anthropic to work on recursive self-improvement marks a key fork. His focus will be using Claude to accelerate its own pre-training research — a direct bet that recursive self-improvement can eliminate jaggedness, despite Anthropic once pledging not to advance AI capabilities progress.