Two Rival Bets on AGI: Google I/O Highlights

The real shift from Google I/O isn’t frontier capability; it’s a strategy pivot: pack “good-enough” multimodal AI into the search box, price it aggressively, and push speed. Gemini 3.5 Flash is notably fast and competent across several domains, but Google didn’t claim coding supremacy. This is a bid to win everyday workflows, not to dethrone frontier models in deep reasoning.

Two rival AGI bets emerged. Google’s leadership frames video/world-model simulation as a step toward AGI. In contrast, OpenAI’s leadership signals text-centric reasoning will get there without world simulators. Anthropic adds a third vector: recursive self-improvement (RSI) to accelerate pre-training using the model itself. None of these are proven; they are strategic wagers with different compute, data, and safety tradeoffs.

Operationally, Google’s message to enterprises was blunt: you’re overspending on AI; use faster, cheaper models where you can. Benchmarks suggest Gemini 3.5 Flash is strong on tasks like chart/table reasoning and some finance workflows, but not a leader in agentic coding. Expect domain-specialized strengths over one-model-to-rule-them-all. Pricing isn’t 10x cheaper at parity; actual costs hinge on tokenization and latency needs.

Reliability is the constraint. A new independent paper shows models can “believe” fabricated claims even when repeatedly labeled false—unless negations are placed extremely precisely. DeepMind leadership calls this jaggedness structural, not a patchable bug. Google’s own caveat—agents are still early to be “truly helpful”—and over-restrictive filters on multimodal inputs confirm the gap between demos and production.

Governance is tightening: OpenAI adopting Google’s SynthID watermarking and both companies allowing “lawful” military use signal a shift in provenance and policy posture. For engineering leaders, the signal is to design verifiable workflows, not chase AGI timelines or demo sizzle.