If 2025 was the year AI got a vibe check<\/a>, 2026 will be the year the tech gets practical. The focus is already shifting away from building ever-larger language models and towards the harder work of making AI usable. In practice, that involves deploying smaller models where they fit, embedding intelligence into physical devices, and designing systems that integrate cleanly into human workflows.\u00a0<\/p>\n The experts TechCrunch spoke to see 2026 as a year of transition, one that evolves from brute-force scaling to researching new architectures, from flashy demos to targeted deployments, and from agents that promise autonomy to ones that actually augment how people work.\u00a0<\/p>\n The party isn\u2019t over, but the industry is starting to sober up.<\/p>\n In 2012, Alex Krizhevsky, Ilya Sutskever, and Geoffrey Hinton\u2019s AlexNet paper<\/a> showed how AI systems could \u201clearn\u201d to recognize objects in pictures by looking at millions of examples. The approach was computationally expensive, but made possible with GPUs. The result? A decade of hardcore AI research as scientists worked to invent new architectures for different tasks.<\/p>\n That culminated around 2020 when OpenAI launched GPT-3, which showed how simply making the model 100 times bigger unlocks abilities like coding and reasoning without requiring explicit training. This marked the transition into what Kian Katanforoosh, CEO and founder of AI agent platform Workera, calls the \u201cage of scaling\u201d: a period defined by the belief that more compute, more data, and larger transformer models would inevitably drive the next major breakthroughs in AI.<\/p>\n Today, many researchers think the AI industry is beginning to exhaust the limits of scaling laws and will once again transition into an age of research.<\/p>\n Yann LeCun, Meta\u2019s former chief AI scientist<\/a>, has long argued against the over-reliance on scaling, and stressed the need to develop better architectures. And Sutskever said in a recent interview<\/a> that current models are plateauing and pre-training results have flattened, indicating a need for new ideas.\u00a0\u00a0<\/p>\n Techcrunch event<\/p>\n \n\t\t\t\t\t\t\t\t\tSan Francisco<\/span> \u201cI think most likely in the next five years, we are going to find a better architecture that is a significant improvement on transformers,\u201d Katanforoosh said. \u201cAnd if we don\u2019t, we can\u2019t expect much improvement on the models.\u201d<\/p>\n Large language models are great at generalizing knowledge, but many experts say the next wave of enterprise AI adoption will be driven by smaller, more agile language models that can be fine-tuned for domain-specific solutions.\u00a0<\/p>\n \u201cFine-tuned SLMs will be the big trend and become a staple used by mature AI enterprises in 2026, as the cost and performance advantages will drive usage over out-of-the-box LLMs,\u201d Andy Markus, AT&T\u2019s chief data officer, told TechCrunch. \u201cWe\u2019ve already seen businesses increasingly rely on SLMs because, if fine-tuned properly, they match the larger, generalized models in accuracy for enterprise business applications, and are superb in terms of cost and speed.\u201d<\/p>\n We\u2019ve seen this argument before from French open-weight AI startup Mistral: it argues its small models actually perform better<\/a> than larger models on several benchmarks after fine-tuning.\u00a0<\/p>\n \u201cThe efficiency, cost-effectiveness, and adaptability of SLMs make them ideal for tailored applications where precision is paramount,\u201d said Jon Knisley, an AI strategist at ABBYY, an Austin-based enterprise AI company.\u00a0<\/p>\n While Markus thinks SLMs will be key in the agentic era, Knisley says the nature of small models means they\u2019re better for deployment on local devices, \u201ca trend accelerated by advancements in edge computing.\u201d<\/p>\n Humans don\u2019t just learn through language; we learn by experiencing how the world works. But LLMs don\u2019t really understand the world; they just predict the next word or idea. That\u2019s why many researchers believe the next big leap will come from world models: AI systems that learn how things move and interact in 3D spaces so they can make predictions and take actions.\u00a0<\/p>\n Signs that 2026 will be a big year for world models are multiplying. LeCun left Meta to start his own world model lab, and is reportedly seeking a $5 billion valuation<\/a>. Google\u2019s DeepMind has been plugging away at Genie, and in August launched its latest model that builds real-time interactive general-purpose world models. Alongside demos by startups like Decart<\/a> and Odyssey<\/a>, Fei-Fei Li\u2019s World Labs<\/a> has launched its first commercial world model, Marble. Newcomers like General Intuition in October scored a $134 million seed round<\/a> to teach agents spatial reasoning, and video generation startup Runway in December released its first world model, GWM-1<\/a>.\u00a0<\/p>\n While researchers see long-term potential in robotics and autonomy, the near-term impact is likely to be seen first in video games. PitchBook predicts the market for world models in gaming could grow from $1.2 billion between 2022 and 2025 to $276 billion by 2030, driven by the tech\u2019s ability to generate interactive worlds and more life-like non-player characters.\u00a0<\/p>\n Pim de Witte, founder of General Intuition, told TechCrunch virtual environments may not only reshape gaming, but also become critical testing grounds for the next generation of foundation models.<\/p>\n Agents failed to live up to the hype in 2025, but a big reason for that is because it\u2019s hard to connect them to the systems where work actually happens. Without a way to access tools and context, most agents were trapped in pilot workflows.\u00a0<\/p>\n Anthropic\u2019s Model Context Protocol (MCP), a \u201cUSB-C for AI\u201d that lets AI agents talk to the external tools like databases, search engines and APIs, proved the missing connective tissue, and is quickly becoming the standard. OpenAI and Microsoft have publicly embraced MCP, and Anthropic recently donated it to the Linux Foundation\u2019s new Agentic AI Foundation<\/a>, which aims to help standardize open-source agentic tools. Google also has begun standing up its own managed MCP servers<\/a> to connect AI agents to its products and services.\u00a0<\/p>\n With MCP reducing the friction of connecting agents to real systems, 2026 is likely to be the year agentic workflows finally move from demos into day-to-day practice.\u00a0<\/p>\n Rajeev Dham, a partner at Sapphire Ventures, says these advancements will lead to agent-first solutions taking on \u201csystem-of-record roles\u201d across industries.\u00a0<\/p>\n \u201cAs voice agents handle more end-to-end tasks such as intake and customer communication, they\u2019ll also begin to form the underlying core systems,\u201d Dham said. \u201cWe\u2019ll see this in a variety of sectors like home services, proptech, and healthcare, as well as horizontal functions such as sales, IT, and support.\u201d\u00a0<\/p>\n While more agentic workflows might raise worries that layoffs may follow, Katanforoosh of Workera isn\u2019t so sure that\u2019s the message.\u00a0<\/p>\n \u201c2026 will be the year of the humans,\u201d he said.\u00a0<\/p>\n In 2024, every AI company predicted they would automate jobs out of needing humans. But the tech isn\u2019t there yet, and in an unstable economy, that\u2019s not really a popular rhetoric. Katanforoosh says next year, we\u2019ll realize that \u201cAI has not worked as autonomously as we thought,\u201d and the conversation will focus more on how AI is being used to augment human workflows, rather than replace them.\u00a0<\/p>\n \u201cAnd I think a lot of companies are going to start hiring,\u201d he added, noting that he expects there to be new roles in AI governance, transparency, safety, and data management. \u201cI\u2019m pretty bullish on unemployment averaging under 4% next year.\u201d<\/p>\n \u201cPeople want to be above the API, not below it, and I think 2026 is an important year for this,\u201d de Witte added. <\/p>\n Advancements in technologies like small models, world models, and edge computing will enable more physical applications of machine learning, experts say.\u00a0<\/p>\n \u201cPhysical AI will hit the mainstream in 2026 as new categories of AI-powered devices, including robotics, AVs, drones and wearables start to enter the market,\u201d Vikram Taneja, head of AT&T Ventures, told TechCrunch.\u00a0<\/p>\n While autonomous vehicles and robotics are obvious use cases for physical AI that will no doubt continue to grow in 2026, the training and deployment required is still expensive. Wearables, on the other hand, provide a less expensive wedge with consumer buy-in. Smart glasses like Meta\u2019s Ray Bans<\/a> are starting to ship assistants that can answer questions about what you\u2019re looking at, and new form factors like AI-powered health rings<\/a> and smart watches<\/a> are normalizing always-on, on-body inference.<\/p>\n \u201cConnectivity providers will work to optimize their network infrastructure to support this new wave of devices, and those with flexibility in how they can offer connectivity will be best positioned,\u201d Taneja said.<\/p>\n<\/div>\nScaling laws won\u2019t cut it<\/h2>\n
![\"Amazon](\"https:\/\/techcrunch.com\/wp-content\/uploads\/2025\/11\/amazon-data-center.png?w=680\")
\n\t\t\t\t\t\t\t\t\t\t\t\t\t|<\/span>
\n\t\t\t\t\t\t\t\t\t\t\t\t\tOctober 13-15, 2026<\/span>\n\t\t\t\t\t\t\t<\/p>\n<\/p><\/div>\n<\/p><\/div>\n<\/div>\nSometimes less is more<\/h2>\n
Learning through experience<\/h2>\n
![\"Space](\"https:\/\/techcrunch.com\/wp-content\/uploads\/2025\/11\/Untitled-design.png?w=680\")
Agentic nation<\/h2>\n
Augmentation, not automation<\/h2>\n
![\"\"](\"https:\/\/techcrunch.com\/wp-content\/uploads\/2026\/01\/igor-omilaev-gVQLAbGVB6Q-unsplash.jpg?w=680\")
Getting physical<\/h2>\n
![\"\"](\"https:\/\/techcrunch.com\/wp-content\/uploads\/2025\/10\/GettyImages-2235448056.jpg?w=680\")