If 2023 was a year for questioning AI's fundamental direction, 2024 will shift the focus toward practical application. The industry's attention is moving beyond simply constructing ever-larger language models to the more challenging task of making AI genuinely useful. This practical turn means deploying compact models where they are most effective, integrating intelligence directly into hardware, and creating systems that seamlessly fit into human processes. The initial excitement remains, but a more measured, implementation-focused mindset is taking hold. The strategy of merely scaling up models has reached its limit.

The modern AI wave began in 2012 with the AlexNet paper by Alex Krizhevsky, Ilya Sutskever, and Geoffrey Hinton, which demonstrated how systems could learn to recognize objects by processing millions of images. This GPU-enabled approach, though computationally intensive, sparked a decade of intense research into new architectures for various tasks. This period culminated around 2020 with OpenAI’s GPT-3, which revealed that dramatically increasing a model's size could unlock capabilities like coding and reasoning without task-specific training. This ushered in what Kian Katanforoosh, CEO of AI platform Workera, terms the "age of scaling," dominated by the belief that more computing power, data, and larger transformers would inevitably produce the next breakthroughs.

Now, a growing consensus among researchers suggests the industry is hitting the diminishing returns of scaling. Many believe a new phase of fundamental research is beginning. Yann LeCun, formerly Meta’s chief AI scientist, has consistently criticized over-reliance on scaling, advocating for better architectural designs. Sutskever recently observed that current models are showing signs of plateauing, with pre-training improvements flattening, signaling a need for fresh ideas. Katanforoosh echoes this, stating, “I think most likely in the next five years, we are going to find a better architecture that is a significant improvement on transformers. And if we don’t, we can’t expect much improvement on the models.”

**The Power of Compact Models**

While large language models excel at broad knowledge generalization, the next wave of business adoption will likely be propelled by smaller, more agile models fine-tuned for specific domains. “We’ve 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,” noted one expert.

This perspective is championed by companies like Mistral, which argues its compact models can outperform larger ones on certain benchmarks after fine-tuning. Jon Knisley, an AI strategist at enterprise AI company ABBYY, explained, “The efficiency, cost-effectiveness, and adaptability of SLMs make them ideal for tailored applications where precision is paramount.” He added that their smaller size makes them particularly suitable for deployment on local devices, a trend accelerated by progress in edge computing.

**Learning by Interacting with the World**

Human learning extends beyond language to include interaction with and observation of the physical world. Current LLMs, which primarily predict the next word, lack this understanding. Consequently, many researchers point to "world models" as the next major frontier. These are AI systems that learn how objects move and interact in three-dimensional spaces, enabling them to make predictions and take actions.

Momentum behind world models is building rapidly for 2024. LeCun departed Meta to establish his own lab focused on this area, reportedly seeking a substantial valuation. Google DeepMind continues developing its Genie project, recently launching a new model for building interactive world models. Alongside demonstrations from startups like Decart and Odyssey, Fei-Fei Li’s World Labs released its first commercial world model, Marble. New entrants are also gaining traction; General Intuition secured a $134 million seed round to teach agents spatial reasoning, and Runway released its first world model, GWM-1.

While the long-term potential lies in robotics and autonomous systems, the immediate impact is expected in video gaming. Analysts project the market for world models in gaming could explode from $1.2 billion between 2022 and 2025 to $276 billion by 2030, fueled by the technology's ability to generate dynamic worlds and realistic non-player characters.

**The Rise of Connected Agents**

AI agents fell short of lofty expectations in 2023, largely because integrating them with real-world tools and data systems proved difficult. Most were confined to limited pilot programs. A key breakthrough came with Anthropic’s Model Context Protocol (MCP), a connective framework likened to a "USB-C for AI" that allows agents to interact with databases, search engines, and APIs.

MCP is rapidly becoming a standard, with public backing from OpenAI and Microsoft. Anthropic has since contributed the protocol to the Linux Foundation’s new Agentic AI Foundation to foster open-source standardization. Google is also establishing its own managed MCP servers. By reducing the friction of connecting agents to live systems, 2024 is poised to be the year agentic workflows transition from demonstrations to daily use.

Rajeev Dham, a partner at Sapphire Ventures, believes these advancements will allow agent-first solutions to assume critical "system-of-record roles" across various industries. “As voice agents handle more end-to-end tasks such as intake and customer communication, they’ll also begin to form the underlying core systems,” Dham said. “We’ll see this in a variety of sectors like home services, proptech, and healthcare, as well as horizontal functions such as sales, IT, and support.”

**Augmentation Over Replacement**

The increase in agentic workflows might spark fears of job displacement, but Katanforoosh offers a different outlook. “2024 will be the year of the humans,” he stated. He notes that while many AI companies previously emphasized full automation, the technology isn't yet capable of that, and such rhetoric is less popular in an uncertain economic climate. The coming year, he predicts, will bring a realization that “AI has not worked as autonomously as we thought,” shifting the conversation toward how AI augments human work rather than replaces it.

“And I think a lot of companies are going to start hiring,” he added, anticipating new roles in AI governance, transparency, safety, and data management. “I’m pretty bullish on unemployment averaging under 4% next year.” This sentiment aligns with the idea that people want to leverage AI as a tool, not be supplanted by it, making 2024 a pivotal year for defining this human-centric relationship.

**AI in the Physical World**

Progress in compact models, world models, and edge computing is paving the way for more machine learning applications in the physical realm. While autonomous vehicles and robotics will continue to advance, their development and deployment remain costly. Wearable technology presents a more accessible entry point with existing consumer acceptance.

Products like Meta’s Ray-Ban smart glasses are beginning to offer assistants that can answer questions about the user's surroundings. New form factors, such as AI-powered health rings and advanced smartwatches, are normalizing the concept of continuous, on-body AI inference. This new wave of devices will push connectivity providers to optimize their network infrastructure, with those offering the most flexible connectivity solutions being best positioned for success.