From reasoning scaling to humanoid robots, from universal multimodality to Science AI—since 2025, the AI field has seen changes across six concurrent tracks that may well be recorded in textbooks. This is a panoramic overview map, not a deep dive—several of these tracks have already been thoroughly dissected in other DeepDives, so here we only provide positioning and cross-references.
Core logic: not just training bigger, but letting the model "think more" during inference. OpenAI o3 lets the model go through a hidden chain of thought before answering; DeepSeek-R1 proved reasoning capabilities can emerge through pure reinforcement learning, without supervised fine-tuning as a prerequisite; Claude's Extended Thinking provides a budget_tokens parameter, letting developers directly control the computational budget for "thinking." The UK AI Safety Institute's Frontier AI Trends Report lists this as the most significant capability leap for 2025–2026. Inference is no longer a fixed post-training capability—it has become a tunable resource investment.
Core logic: a single model simultaneously understands and generates text, images, audio, and video. In May 2026, Google released Gemini Omni, which DeepMind defines as a "world model"—not generating video by predicting pixels frame by frame, but understanding objects' physical properties, continuity, and temporal relationships; OpenAI's Sora 2 and ByteDance's Seedance 2.0 simultaneously upgraded their world simulation capabilities. Multimodality is no longer "stitching several models together," but a single model's unified understanding of the physical world.
Core logic: from code completion to independently completing software engineering tasks. SWE-bench Verified surged from ~10% in 2023 to 80.8% for Claude Opus 4.6 in early 2026 (GPT-5.2 80.0%, Gemini 3.1 Pro 80.6%); ~85% of developers now regularly use AI coding assistance tools, and Claude Code reached $1 billion annualized revenue within 6 months of launch. The division of labor in software engineering is being rewritten—developers increasingly play the role of "architecture decision-maker" rather than "code implementer."
This track has been more deeply dissected in our "Machine Economy" and "Agent Engineering Evolution" DeepDives, so we won't repeat the expansion here → agent-economy · agent-engineering-evolution
Core logic: language model capabilities are finally beginning to transfer to the physical world. NVIDIA Isaac GR00T is a foundation model designed specifically for humanoid robots; Figure AI has been continuously deployed at BMW factories for over 11 months, building a new factory with annual capacity of 12,000 units; robotics funding reached €38.6 billion in 2025, accounting for 9% of global venture capital, with the "world model" track financing soaring from €1.3 billion to €6.5 billion. Sony AI's Project Ace is the first known autonomous system that can beat humans at a professional level—competitive table tennis, and it's a real deployment, not a demo.
Core logic: AI is starting to truly accelerate scientific discovery, not just assist with literature retrieval. AlphaFold 3 no longer only predicts protein structures—it can simultaneously model protein, DNA, RNA, and small molecule interactions, achieving 76.4% accuracy on protein-ligand docking, a 1.8× improvement over prior methods. Drug discovery is evolving from "AI predicts candidate molecules" to "AI designs and validates the entire molecule-target-mechanism chain," and target discovery cycles that previously took ten years are being compressed to months. The same paradigm is being replicated in climate science, materials discovery, mathematical proofs, and other fields.
Core logic: instead of activating all parameters, only call on the "necessary experts" for each token. Since early 2025, nearly all frontier models have adopted the MoE architecture—DeepSeek-V3/R1, Llama 4, Mistral Large 3, the entire Gemini family. Mixtral 8x7B has 46.7 billion parameters, but its inference speed is equivalent to a 13-billion-parameter model; Gemma 4's 26 billion parameters actually only activate about 4 billion. MoE is not just an efficiency improvement—it reshapes the relationship between model capability and cost, serving as the infrastructure prerequisite that makes the previous five technologies economically deployable.
These six technologies appear independent, but a hidden thread runs through them all: AI is shifting from a 'point tool' to 'general infrastructure'—sufficiently general, sufficiently efficient, sufficiently cheap. The next question is no longer "Can AI do it?" but "Who decides what it does."
First published 2026-07-24