Small models and world models lead AI’s transition to pragmatism

The artificial intelligence industry is entering a transition year, moving away from the age of scaling that began around 2020 when OpenAI launched GPT-3. According to TechCrunch, 2026 will be the year the tech gets practical. Rather than relying solely on making models larger, the industry is shifting toward researching new architectures and deploying targeted applications. Kian Katanforoosh, CEO and founder of AI agent platform Workera, noted that the industry will most likely find a better architecture within the next five years that significantly improves upon transformers. Without such developments, further model improvements may remain limited.

Enterprise adoption is increasingly pivoting toward Small Language Models (SLMs)—which are smaller, more agile language models—due to their cost and performance advantages over out-of-the-box large language models. Andy Markus, chief data officer at AT&T, stated that fine-tuned SLMs will be the big trend and become a staple used by mature AI enterprises in 2026. This shift is supported by French open-weight AI startup Mistral, which argues its small models perform better after fine-tuning. Jon Knisley, an AI strategist at ABBYY, an Austin-based enterprise AI company, added that these models are ideal for tailored applications requiring precision due to their efficiency, cost-effectiveness, and adaptability.

Beyond language, world models—AI systems that learn how things move and interact in 3D spaces—represent the next major technological leap. Startups and research divisions are securing capital to develop these systems; Google’s DeepMind division is working on world models, video generation startup Runway released its GWM-1 world model, and World Labs launched its Marble world model. Meanwhile, Yann LeCun’s world model lab is reportedly seeking a $5 billion valuation, and startup General Intuition raised a $134 million seed round. While the technology has physical applications, its near-term impact is projected to hit the gaming sector first. PitchBook reports that the market for world models in gaming could grow from $1.2 billion between 2022 and 2025 to $276 billion by 2030.

At the same time, agentic workflows are moving from demos into day-to-day practice, enabled by the Model Context Protocol (MCP)—a standard for connecting AI agents to external tools. Backed by Anthropic, which created MCP, and embraced by OpenAI, the protocol reduces the friction of connecting agents to databases and APIs. Rajeev Dham, a partner at Sapphire Ventures, expects these advancements to help agents take on system-of-record roles across various sectors. However, this shift is expected to focus on human augmentation rather than mass automation. Katanforoosh of Workera added that he is pretty bullish on unemployment averaging under 4% next year, asserting that “2026 will be the year of the humans” as the industry realizes AI has not worked as autonomously as previously predicted.

These advancements are also expected to drive physical applications of machine learning. According to TechCrunch, physical AI is anticipated to enter the mainstream in 2026 as new categories of devices, including wearables like the Ray-Ban Meta smart glasses, begin to ship with assistants.

Why it matters

The AI industry is transitioning from a focus on brute-force scaling of large language models to practical, domain-specific applications, world models, and agentic workflows in 2026. This shift lowers deployment costs and makes AI tools highly specialized, altering how enterprises integrate machine learning into daily operations.