Scaling "Thinking": Gemini 2.5 Tech Lead Jack Rae on Reasoning, Long Context, & the Path to AGI

Episode Summary

Executive Summary: Jack Ray of Google DeepMind argues Gemini 2.5 Pro’s leap came from steady, scalable progress in RL-based reasoning, long-context training, and multimodal integration—not a sudden breakthrough. He says frontier labs are converging on similar reasoning systems because the path is obvious and low-hanging fruit remains, while AGI still needs stronger agents, memory, and interpretability.

Main Topics: Why RL for reasoning suddenly looks effective (Priority: 5/5): Ray says reinforcement learning on verifiable correctness signals has been improving reasoning for over a year at DeepMind, and the public just recently noticed when it crossed a capability threshold. He frames the change as smooth, cumulative progress rather than a discrete scientific breakthrough. Convergence on reasoning/think models across frontier labs (Priority: 5/5): He explains that multiple labs independently arrived at chain-of-thought-style reasoning models because test-time compute was an obvious next frontier. Once the field saw early results and talent concentrated there, progress accelerated rapidly across the industry. Chain of thought, safety, and what should be shown to users (Priority: 4/5): The conversation examines whether to expose raw thinking tokens, whether to RL-shape them, and whether doing so risks deceit or reward-hacking. Ray supports showing raw thoughts in experimental releases while stressing faithfulness, safety, and usefulness. Human data, process data, and how models learn to reason (Priority: 4/5): Ray says human chain-of-thought is hard to elicit faithfully, but process traces from real work are valuable. He emphasizes that much reasoning behavior already exists in pretraining data, and that RL and synthetic data are used selectively to improve outcomes. Latent-space reasoning and interpretability (Priority: 4/5): They discuss reasoning without explicit tokens, via latent vectors or continuous internal computation. Ray is open to exploring it if it improves capability and remains interpretable, drawing an analogy to MuZero and stressing that interpretability must keep pace. Roadmap from current systems to AGI (Priority: 5/5): Ray argues that AGI will likely come from combining better reasoning, agents, memory, long context, and multimodal world models. He says memory is not solved, long context is already very powerful, and the remaining path is mostly about compounding known bottlenecks.

Key Arguments: RL on correctness/verifiable signals has been helping Gemini reasoning for a long time; the public perception of a sudden jump reflects threshold effects, not a sudden invention. The frontier labs are converging on similar reasoning architectures because test-time compute and reasoning were obvious, high-value directions once the field saw initial evidence. Smaller models and early RL attempts often fail because many implementation details must align; one broken component can make the whole recipe appear ineffective. The best training recipe is usually the simplest one that works: minimize hand-coded priors, let the model learn behavior end-to-end, and use human or synthetic data when it improves generalization. Raw chain-of-thought can be useful as scratch space, but it should not be over-optimized for readability if that would reduce capability or introduce deceitful behavior. Human inner monologue is difficult to transcribe faithfully; process data is more valuable when it captures actual task execution rather than forced verbalization of thoughts. Reasoning and agency are tightly related but still distinct research areas; building effective agents requires additional work on environments, actions, and tool use. Latent-space reasoning is promising if it remains interpretable and safe; capability gains should not be tabooed preemptively. Pretraining mainly learns the world model via compression, but post-training/RL must still create new skills and behaviors needed for AGI. Long context is a major practical unlock, but memory is not solved; more advanced memory systems will likely be needed for lifelong, open-ended agents. Multimodal integration is crucial because jointly trained text-image-video-audio systems create deeper world models than tool-calling architectures.

Data Points: Gemini 2.5 Pro context window: 400,000 tokens - Host describes dumping a full research codebase into Gemini 2.5 Pro and being impressed by the model’s command of the context. Thinking group formation at Google: September-October - Ray says the reasoning/test-time compute group formed around this period before the December experimental launch. First thinking model launch: December - He cites an experimental Flash-with-Thinking release as an early milestone in Google’s reasoning push. Memory/context scale: 1 million to 10 million tokens - Ray notes that modern context lengths are beginning to feel like lifelong memory scales, depending on representation. Training compute split discussion: Post-training vs pretraining is moving from 'two orders of magnitude less' toward a larger share - He suggests the old dichotomy is becoming more of a spectrum as RL scale rises. External feedback on 2.5 Pro: 128k-context leaderboard using Gemini 2.5 Pro effectively - Ray mentions an external leaderboard shared on X showing strong long-context utilization by 2.5 Pro. Anthropic interpretability limitation: 50% of behaviors explained - The host references Anthropic’s tracing work and notes its replacement models only explain about half the behaviors. Safety testing: 'unprecedented level' - Ray says Google performs extensive safety testing before release, though experimental launches have a different evaluation tier.

Pivotal Quotes: "There hasn't been one key thing which has discreetly made it working. It's just kind of crossed the capability threshold where people have really taken notice." — Jack Ray: On why reinforcement learning for reasoning appears to have 'suddenly' started working across the industry. "The simplest recipe that leads to the model, choose that one." — Jack Ray: On Google’s preference for minimal priors and end-to-end learning when shaping cognitive behaviors. "I don't feel like right now, you can even see it on the ground, the models are quite different already." — Jack Ray: On why frontier labs have not converged to a single AGI path, even if they share similar reasoning trends.

Implications: The episode suggests frontier AI progress is still mostly about scaling obvious ideas well: RL reasoning, long context, agents, memory, and multimodal training. For users, the near-term gains are practical; for industry, the race now hinges on execution, safety, and interpretability.