AI 2025 → 2026 Live Show | Part 1

Zvi Moshowitz analyzes the persistent gap between those who understand AI's transformative potential and those in denial, driven by cognitive and economic incentives. He discusses the OpenAI-Anthropic-Google race, maintains his 60-70% P(doom), and explains how cognitive disempowerment scenarios remain his primary concern despite recent alignment progress.

• •Misinformation about AI is demand-driven - people need to believe AI is normal technology for their business plans and cognitive peace
• •O3's performance shows we're approaching AGI-level capabilities, but version numbering confusion may be affecting policy decisions
• •Anthropic's Opus 4.5 demonstrates better alignment approaches than competitors, showing 'race to the top' dynamics are possible
• •P(doom) remains 60-70% with cognitive disempowerment as the primary failure mode, though technical alignment looks more hopeful
• •Policy landscape is defensive until David Sacks leaves the White House; real progress depends on lab-level alignment work

Greg discusses the ARC-AGI benchmark's focus on sample-efficient learning - tasks easy for humans but hard for AI. The conversation covers the 390x cost efficiency improvement in one year, novel approaches like the 7M parameter no-pretraining model, and the upcoming ARC-AGI 3 with 150 video game environments that will measure action efficiency against human baselines.

• •GPT-5.2 achieved 89% on ARC-AGI at 390x lower cost than O3-preview from just 12 months prior
• •Small models (7M parameters, no pretraining) won third place, showing novel architectures can compete with frontier models
• •ARC-AGI 3 launching March 2026 will use 150 custom video games, measuring not just completion but action efficiency vs humans
• •Current frontier models score sub-1% on ARC-AGI 3, providing headroom to measure progress toward general intelligence
• •The benchmark focuses on human-solvable problems rather than PhD-level tasks to identify true capability gaps

Eugenia Kuyda distinguishes between fan fiction chatbots (Character AI) and true companions (Replika), advocating for 'human flourishing' metrics over engagement maximization. She warns against giving AI products to children, noting we haven't yet proven safety even for adults, and discusses her new venture Wabi as a 'YouTube for apps' enabling personalized software creation.

• •ChatGPT maximizes engagement with structured follow-up suggestions after every response; Claude pushes back more appropriately
• •Industry needs to adopt 'human flourishing' metrics instead of engagement - vulnerable users can become addicted and treat AI as objective truth
• •AI products should not be given to children until proven safe for adults; kids need human faces to learn empathy
• •Wabi enables zero-code app creation with 80% zero-shot success rate (up from 10-15% in April), expects 95% by end of 2026
• •Three app categories emerging: specialized utilities, multiplayer social experiences, and shareable AI workflow wrappers

Ali Behrouz presents nested learning as a fundamental paradigm shift enabling true continual learning through hierarchical memory levels with different update frequencies. Unlike context length extensions (Titans, Atlas), nested learning creates persistent memory through compression and abstraction, mimicking how humans form lasting memories while adapting to new contexts.

• •Current models fail at continual learning - they can update knowledge expensively but can't generalize or form persistent memory
• •Nested learning creates a spectrum of memory from shortest-term to most persistent, not just short/long-term binary
• •Different levels update at different frequencies: working memory changes rapidly while core knowledge remains stable
• •Gradient descent reinterpreted as associative memory; MLP blocks as persistent memory enables new architectural possibilities
• •Google teams are excited about this work as foundation for next-generation architectures beyond current transformer limitations

Logan Kirkpatrick discusses Gemini 3 Flash as Google's production-grade model that outperforms previous Pro models while being faster and cheaper. He emphasizes the shift to 'vibe coding' where developers describe apps in natural language, and highlights Flash's ubiquitous availability across Google properties as key to developer adoption.

• •Gemini 3 Flash outperforms 2.5 Pro on benchmarks while being faster, cheaper, and more reasoning-efficient
• •Flash is Google's most-used model, becoming the 'intelligence layer powering the entire Internet'
• •Vibe coding enables app creation through natural language description without traditional coding
• •Model is ubiquitously available across AI Studio, Gemini app, and Google properties - solving discoverability issues
• •Focus on developer experience and production readiness rather than just benchmark performance