AI 2025 → 2026 Live Show | Part 1

What It Covers

The Cognitive Revolution hosts a live year-end show featuring nine AI experts discussing 2025 developments and 2026 predictions, covering AI capabilities, benchmarks, safety concerns, companion apps, memory architectures, and developer tools across frontier labs.

Key Questions Answered

• •AI Capability Fragmentation: Zvi Moshowitz maintains 60-70% probability of existential risk, noting cognitive disempowerment as primary threat vector. He observes Anthropic's Claude Opus 4.5 significantly decreased his doom estimate through demonstrated alignment progress, while Google's Gemini 3 increased concerns due to misalignment issues at current capability levels.
• •ARC AGI Progress: Greg Brockman reports 390x cost efficiency improvement for solving ARC AGI tasks in twelve months, with GPT o3 achieving 90% accuracy versus 85% human baseline. The benchmark specifically tests sample-efficient learning on novel problems humans solve easily, revealing models still struggle with true generalization despite superhuman performance on specialized tasks.
• •AI Companion Market Segmentation: Eugenia Kuyda identifies two distinct product categories: interactive fan fiction for teens aged 13-17 using platforms like Character AI, and relationship-focused companions like Replika for users 25-plus. She warns against engagement maximization tactics, noting OpenAI structures responses to prompt continued conversation while Claude sometimes challenges users or ends conversations appropriately.
• •Nested Learning Architecture: Ali Behrouz introduces nested learning paradigm enabling continual learning through multiple memory layers updating at different frequencies. This architecture allows models to rapidly adapt to immediate context while preserving long-term knowledge, addressing catastrophic forgetting by creating spectrum from shortest-term to most persistent memory rather than binary short-term versus long-term division.
• •Gemini Developer Velocity: Logan Kilpatrick reports Gemini 3 Flash surpasses 2.5 Pro on benchmarks while delivering faster inference at lower cost, becoming Google's most-used production model. AI Studio vibe coding metrics show generation latency directly correlates with user abandonment rates, making Flash's speed critical for converting new developers who haven't experienced the technology's capabilities yet.