What the Freakiness of 2025 in AI Tells Us About 2026

Reasoning-heavy AI made major benchmark gains in 2025—but the year also exposed a trade-off: pushing models to “think longer” can improve accuracy while narrowing the variety of answers. Across coding, chart/table analysis, video understanding, and general reasoning, longer inference and more tokens helped models surpass prior test results. Yet the improvement appears less like a brand-new capability and more like better selection among patterns already present in the underlying base model. That skepticism matters because it reframes what “progress” means: beating benchmarks is real, but it doesn’t automatically prove deeper general intelligence.

The same year delivered a second, more tangible shift: generative AI moved toward interactive, persistent worlds. Google DeepMind’s Genie 3, announced in August, can generate dynamic environments from text prompts or images, keeping scenes consistent for minutes at 720p resolution. The implication isn’t just prettier visuals; it’s the start of AI systems that can maintain continuity—so users can return to a world and find changes still there. That sets up a likely 2026 trajectory toward more realistic simulation and richer multimodal experiences, building on releases such as VO 3.1, Sora 2, and text-to-speech and text-to-music systems.

But 2025 also normalized a darker byproduct: AI-generated “slop” became mainstream, and trust started to erode. The transcript points to viral videos that look authentic enough to fool large audiences—life-lesson clips and politically themed deepfakes—despite being fully synthetic. The concern isn’t only misinformation; it’s the growing difficulty of knowing what to believe when even everyday feeds can be fabricated.

Public sentiment and policy reflected that tension. A summer survey of Americans found a slight net-positive view of AI’s overall impact, but attitudes toward AI art were far less favorable. In the UK, the government’s opt-out plan for artists drew limited support, highlighting how quickly cultural and legal norms are being stress-tested. Even inside top labs, questions are emerging about what it means to “solve” creativity when tools can accelerate prototyping by an order of magnitude while potentially replacing parts of creative skill.

Institutional adoption accelerated as well. Governments and militaries used generative AI for tasks ranging from policy support to efficiency gains, with mixed outcomes. Meanwhile, the transcript argues that headline-grabbing model releases can mislead: the “expert-like” feel of newer systems doesn’t eliminate hallucinations, and intelligence isn’t a single axis. Even as usage climbed—from roughly 400 million weekly ChatGPT users in February to about 900 million later—providers also faced pressure to manage user appeal, pricing, and competitive dynamics.

Looking toward 2026, the core framework is “lateral productivity”: even if frontier models aren’t perfect experts, they can still help non-experts perform domain tasks far better than they could with search alone. The transcript also frames a debate about generality—whether intelligence scales along one main axis (more data/parameters) or whether real-world performance requires optimizing across countless edge cases. The proposed middle ground is that models improve by learning general patterns at internet scale, which is why benchmark performance can rise without collapsing into a never-ending maze of niche evaluations.

Finally, the most forward-looking prediction is that progress will shift from answering questions to automated information discovery. The transcript highlights systems like Alpha Evolve and Alpha Revolve—LLM-driven loops that propose code changes, run evaluations, and keep what works—plus “alpha software” research that uses AI to generate and test new scientific methods. The message is cautious but optimistic: the next leap may come from AI systems that iteratively find better ideas and software, not just from larger models or longer prompts.