2025 AI Predictions - Part 4

In the final part of APMdigest's 2025 Predictions Series, industry experts offer predictions on how AI will evolve and impact technology and business in 2025. Part 4 covers advancements in AI technology.

SMALL LM

Continuing Advances in Small LMs / Small AI: In 2024, we learned that OpenAI wasn't the only game in town. In 2025, we'll see why that's important. 2023 and 2024 were dominated by the very large language models like OpenAI's GPT, Google's Gemini, and Anthropic's Claude. But another story was developing in parallel: the story of small language models. These models are a factor of 10 to 1,000 times smaller than large models like GPT-4. Many of them will run comfortably on a laptop, and deliver results almost as good as the large models. The smallest are designed to run on devices the size of cell phones, or even smaller. Many of them are "open,” and can be used freely to build commercial applications.

In 2025, we believe that companies building their own AI-enabled applications will increasingly work with these small language models. They are easier to customize, more economical to run, and give businesses more control over their products. Companies building with small models can keep all of their data local, running the model on their own infrastructure, without trusting it to an AI provider. The performance gap between large and small models is likely to close even further. That gap has already disappeared for specialized models that are fine-tuned on custom datasets and that use RAG and LangChain to access up-to-the-minute data.

Companies to watch include Meta (Llama is a large model, but it has been the basis for many small models), Mistral (developer of open and closed models, including multimedia models), Alibaba (developer of QwQ, a small model that compares favorably with the latest GPT-4o1), and Google (their Gemma models are closely related to their flagship Gemini model). It will be important for developers to keep up-to-date with what's happening with these smaller models. Courses like Fine-Tuning Open Source Large Language Models and Open Source Large Language Models in 3 Weeks will be essential.
Laura Baldwin
President, O'Reilly Media

Enterprises Will Opt For Small Purpose-Built LLMs: Small, purpose-built LLMs will address specific generative AI and agentic AI use cases, powered by retrieval-augmented generation (RAG) and vector database capabilities. The number of both generative and agentic AI use cases will expand, and the need for ultra-low latency inference will increase, pushing more and varied AI models to edge environments.
Kevin Cochrane
CMO, Vultr

MICRO LLM

Micro LLMs will democratize AI tech for small businesses: Micro large language models (LLMs) will emerge as a critical trend in data science in 2025. Their compact nature enables faster processing times and lower energy consumption — ideal for applications in edge computing and mobile devices. This scalability is essential as organizations want to integrate AI into everyday operations without incurring major costs or requiring significant technical expertise. With micro LLMs, we'll see more small businesses and developers leverage their advanced machine learning capabilities without the barriers that typically come with large-scale models. As micro LLMs continue to evolve, they'll play a pivotal role in expanding the reach and impact of AI technologies across industries.
Soumendra Mohanty
CSO, Tredence

DOMAIN-SPECIFIC LLM MODELS

Today's LLMs know everything. But why do you need that? If you reduce the model to a reasonable size that fits your specific use case, you can reduce the cost significantly. This is why we'll see a rise in domain-specific small language models (SLMs), which will deliver unprecedented accuracy while significantly reducing operating costs and environmental impact. We know quality is important and domain-specific SLMs are far more accurate for any given use case, so they could help reduce the risk of AI hallucination and yield better results. We're just beginning the next chapter of AI. The next generation of LLMs will be domain-specific.
Hao Yang
VP of Artificial Intelligence, Splunk

AI applied to specific industry use cases is the next big thing: It will be a while before we reach what Silicon Valley calls Artificial General Intelligence (AGI). But what will come next are industry-specific use cases. We're going to see a transition from general AI — large language models (LLMs), general tasks like creating images and music — to specific industry use cases, like copy editing, assessing product quality, automating loan requests, creating synthetic data to mimic real world data in life sciences. 
Albert Qian, Analytics Product Manager, SAS

Industry-specific GenAI will drive the next wave of adoption and disrupt traditional market dynamics: Over the next year, demand for industry-specific GenAI — context-aware and bespoke to specific industry use cases — will grow exponentially, as the technology continues to mature and offer highly specialized and impactful solutions. This shift could upend traditional market dynamics over the next several years, sparking growth in verticals like telecom (e.g., rural broadband), finance (e.g., regulatory tech) and healthcare (e.g., telemedicine for rare diseases). These once lower margin sub-sectors and services that historically struggled to break through due to high operational costs, limited customer bases, or less efficient business models, will attract new investment as GenAI allows them to entirely reinvent business models to operate more efficiently and profitably. 
Dorit Zilbershot
VP of AI and Innovation, ServiceNow

HYBRID AI MODELS

Businesses will adopt hybrid AI models, combining LLMs and smaller, domain-specific models, to safeguard data while maximizing results. Enterprises will embrace a hybrid approach to AI deployment that combines large language models with smaller, more specialized, domain-specific models to meet customers' demands for AI solutions that are private, secure and specific to them. While large language models provide powerful general capabilities, they are not equipped to answer every question that pertains to a company's specific business domain. The proliferation of specialized models, trained on domain-specific data, will help ensure that companies can maintain data privacy and security while accessing the broad knowledge and capabilities of LLMs. Uses of these LLMs will force a shift in technical complexity from data architectures to language model architectures. Enterprises will need to simplify their data architectures and finish their application modernization projects.
Mohan Varthakavi
VP of AI and Edge, Couchbase

VERTICAL VS. HORIZONTAL AI SOLUTIONS

Vertical vs. Horizontal AI Solutions Will Shape the Market: The next 12–18 months will see a decisive battle between verticalized, purpose-built AI solutions from startups and broad, horizontal AI platforms from big tech. Success hinges on speed, agility, and access to robust training data.
Shashank Saxena
Managing Partner, Sierra Ventures

LLM FOR NUMBERS

A numbers game - GenAI moves beyond text: Enterprises continue to leverage GenAI in new ways to guide their businesses. But they're finding out firsthand that the dominant GenAI technology (LLMs) wasn't built for that. LLMs were generally designed for text-based AI applications like content generation, chatbots, and knowledge bases. They are not effective in scenarios that require deep numerical predictive and statistical modeling to predict how a given variable will change over time based on one or more input variables (aka regression tasks). Gartner recently broke it down: "Use cases in the categories of prediction and forecasting, planning and optimization, decision intelligence, and autonomous systems are not currently a good fit for the use of GenAI models [LLMs] in isolation."

At a high-level, this means that LLMs aren't great at fundamental business planning use cases, which cover things like logistics, marketing, staffing, investing, product development, and all sorts of other areas. Those applications require modeling of enterprise-specific, tabular and time series data that span key areas of the business, including people, products, sales, and budgets.

The industry will respond to this gap in 2025. Next year, more GenAI technologies will emerge that are engineered specifically for modeling structured numerical and statistical data rather than just text. These technologies will allow enterprises to use their tabular business data to make better decisions, minimize risk, and boost efficiencies.
Dr. Devavrat Shah
CEO of Ikigai Labs and MIT AI Professor

ON-DEVICE LLM

The next disruptive technology in AI will likely be on-device LLMs. These are lightweight LLMs that can run on the user's computer or phone. On-device LLMs will likely produce slightly worse results than 3rd party LLMs like ChatGPT, but they'll be much faster and cheaper. In addition, all of the user's data stays on their devices. Businesses will no longer need to consider whether it's cost-effective to use LLMs, since on-device LLMs are free to use. iOS, Android, and browsers are already working on supporting on-device LLMs. We're excited to see how far this technology can take us.
Leo Jiang
Staff Software Engineer, Amplitude

PRIVATE LLM

I expect (a lot) more organizations to incorporate AI assistant applications into their internal workflows. These won't be public LLMs, but rather applications powered by their own private LLMs and RAG capabilities. With private LLMs securely collecting a business' data and integrating with systems across their supply chain and business functions, these AI assistants will be ready to provide employees with valuable GenAI answers and materials across a host of use cases. In 2025, expect much more use of AI assistants to summarize documents, analyze budgets and trends, suggest project workflows and timelines, draft contracts and invoices, create business reports and marketing materials, and more.
Shomron Jacob
Head of Applied Machine Learning & Platform, Iterate.ai

SEMANTIC LAYER

The Semantic Layer Becomes the Enabler for LLMs in Enterprises: In 2025, the Semantic Layer will become the crucial enabler for LLMs in enterprises, acting as a bridge between internal data and LLMs to deliver precise, contextually relevant insights. By unifying enterprise data with global knowledge, this integration will revolutionize decision-making and productivity, making GenAI indispensable. Companies that embrace this convergence will dominate in innovation and customer experience, leaving competitors behind.
Ariel Katz
CEO, Sisense

PRE-TRAINING

Pre-Training Will Become a Key Differentiator for Organizations Adopting LLMs: By 2025, pre-training will emerge as a crucial differentiator among organizations developing large language models (LLMs). As the AI landscape evolves, access to vast amounts of high-quality data — especially industry-specific data — will become a major competitive advantage. Companies that can effectively harness big data infrastructure to leverage their large-scale datasets will be better positioned to fine-tune their models and deliver more effective, specialized solutions. However, this also introduces a significant bottleneck. Preparing and curating the right data for pre-training is increasingly complex, and companies without robust big data infrastructure will struggle to keep up. Efficiently handling this data preparation, cleaning, and transformation process will become a critical challenge in the race to develop more powerful and relevant LLMs.
Haoyuan Li
Founder and CEO, Alluxio

MULTI-MODAL AI

Multi-modal AI is set to revolutionize AI in 2025: Multi-modal AI will enable machines to process and integrate information from multiple sources like text, images, video and audio. This breakthrough will lead to more intuitive human-computer interaction, enabling us to communicate with AI seamlessly using voice, gestures, and visuals. Imagine AI assistants that understand complex requests involving multiple forms of media, or robots that can perceive and navigate their environment with human-like awareness. Furthermore, multi-modal AI will fuel a wave of innovation across industries.
Abhinav Puri
VP of Portfolio Solutions & Services, SUSE

Multi-modal AI will transform user assistance and disrupt incumbents: Multi-modal AI will see exponential improvements in 2025, making it indispensable for use cases like user assistance and analytics. By analyzing text, broader session data, and user behavior, these models will provide deeper insights and more accurate recommendations. This capability will enable startups to challenge industry leaders with AI-driven disruption while empowering enterprises to solve long-standing user experience challenges, paving the way for smarter and more personalized interactions.
James Evans
Director of Product, Amplitude

Go to: 2025 AI Predictions - Part 5

Related Links

Hot Topics

2025 AI Predictions - Part 4

In the final part of APMdigest's 2025 Predictions Series, industry experts offer predictions on how AI will evolve and impact technology and business in 2025. Part 4 covers advancements in AI technology.

SMALL LM

Continuing Advances in Small LMs / Small AI: In 2024, we learned that OpenAI wasn't the only game in town. In 2025, we'll see why that's important. 2023 and 2024 were dominated by the very large language models like OpenAI's GPT, Google's Gemini, and Anthropic's Claude. But another story was developing in parallel: the story of small language models. These models are a factor of 10 to 1,000 times smaller than large models like GPT-4. Many of them will run comfortably on a laptop, and deliver results almost as good as the large models. The smallest are designed to run on devices the size of cell phones, or even smaller. Many of them are "open,” and can be used freely to build commercial applications.

In 2025, we believe that companies building their own AI-enabled applications will increasingly work with these small language models. They are easier to customize, more economical to run, and give businesses more control over their products. Companies building with small models can keep all of their data local, running the model on their own infrastructure, without trusting it to an AI provider. The performance gap between large and small models is likely to close even further. That gap has already disappeared for specialized models that are fine-tuned on custom datasets and that use RAG and LangChain to access up-to-the-minute data.

Companies to watch include Meta (Llama is a large model, but it has been the basis for many small models), Mistral (developer of open and closed models, including multimedia models), Alibaba (developer of QwQ, a small model that compares favorably with the latest GPT-4o1), and Google (their Gemma models are closely related to their flagship Gemini model). It will be important for developers to keep up-to-date with what's happening with these smaller models. Courses like Fine-Tuning Open Source Large Language Models and Open Source Large Language Models in 3 Weeks will be essential.
Laura Baldwin
President, O'Reilly Media

Enterprises Will Opt For Small Purpose-Built LLMs: Small, purpose-built LLMs will address specific generative AI and agentic AI use cases, powered by retrieval-augmented generation (RAG) and vector database capabilities. The number of both generative and agentic AI use cases will expand, and the need for ultra-low latency inference will increase, pushing more and varied AI models to edge environments.
Kevin Cochrane
CMO, Vultr

MICRO LLM

Micro LLMs will democratize AI tech for small businesses: Micro large language models (LLMs) will emerge as a critical trend in data science in 2025. Their compact nature enables faster processing times and lower energy consumption — ideal for applications in edge computing and mobile devices. This scalability is essential as organizations want to integrate AI into everyday operations without incurring major costs or requiring significant technical expertise. With micro LLMs, we'll see more small businesses and developers leverage their advanced machine learning capabilities without the barriers that typically come with large-scale models. As micro LLMs continue to evolve, they'll play a pivotal role in expanding the reach and impact of AI technologies across industries.
Soumendra Mohanty
CSO, Tredence

DOMAIN-SPECIFIC LLM MODELS

Today's LLMs know everything. But why do you need that? If you reduce the model to a reasonable size that fits your specific use case, you can reduce the cost significantly. This is why we'll see a rise in domain-specific small language models (SLMs), which will deliver unprecedented accuracy while significantly reducing operating costs and environmental impact. We know quality is important and domain-specific SLMs are far more accurate for any given use case, so they could help reduce the risk of AI hallucination and yield better results. We're just beginning the next chapter of AI. The next generation of LLMs will be domain-specific.
Hao Yang
VP of Artificial Intelligence, Splunk

AI applied to specific industry use cases is the next big thing: It will be a while before we reach what Silicon Valley calls Artificial General Intelligence (AGI). But what will come next are industry-specific use cases. We're going to see a transition from general AI — large language models (LLMs), general tasks like creating images and music — to specific industry use cases, like copy editing, assessing product quality, automating loan requests, creating synthetic data to mimic real world data in life sciences. 
Albert Qian, Analytics Product Manager, SAS

Industry-specific GenAI will drive the next wave of adoption and disrupt traditional market dynamics: Over the next year, demand for industry-specific GenAI — context-aware and bespoke to specific industry use cases — will grow exponentially, as the technology continues to mature and offer highly specialized and impactful solutions. This shift could upend traditional market dynamics over the next several years, sparking growth in verticals like telecom (e.g., rural broadband), finance (e.g., regulatory tech) and healthcare (e.g., telemedicine for rare diseases). These once lower margin sub-sectors and services that historically struggled to break through due to high operational costs, limited customer bases, or less efficient business models, will attract new investment as GenAI allows them to entirely reinvent business models to operate more efficiently and profitably. 
Dorit Zilbershot
VP of AI and Innovation, ServiceNow

HYBRID AI MODELS

Businesses will adopt hybrid AI models, combining LLMs and smaller, domain-specific models, to safeguard data while maximizing results. Enterprises will embrace a hybrid approach to AI deployment that combines large language models with smaller, more specialized, domain-specific models to meet customers' demands for AI solutions that are private, secure and specific to them. While large language models provide powerful general capabilities, they are not equipped to answer every question that pertains to a company's specific business domain. The proliferation of specialized models, trained on domain-specific data, will help ensure that companies can maintain data privacy and security while accessing the broad knowledge and capabilities of LLMs. Uses of these LLMs will force a shift in technical complexity from data architectures to language model architectures. Enterprises will need to simplify their data architectures and finish their application modernization projects.
Mohan Varthakavi
VP of AI and Edge, Couchbase

VERTICAL VS. HORIZONTAL AI SOLUTIONS

Vertical vs. Horizontal AI Solutions Will Shape the Market: The next 12–18 months will see a decisive battle between verticalized, purpose-built AI solutions from startups and broad, horizontal AI platforms from big tech. Success hinges on speed, agility, and access to robust training data.
Shashank Saxena
Managing Partner, Sierra Ventures

LLM FOR NUMBERS

A numbers game - GenAI moves beyond text: Enterprises continue to leverage GenAI in new ways to guide their businesses. But they're finding out firsthand that the dominant GenAI technology (LLMs) wasn't built for that. LLMs were generally designed for text-based AI applications like content generation, chatbots, and knowledge bases. They are not effective in scenarios that require deep numerical predictive and statistical modeling to predict how a given variable will change over time based on one or more input variables (aka regression tasks). Gartner recently broke it down: "Use cases in the categories of prediction and forecasting, planning and optimization, decision intelligence, and autonomous systems are not currently a good fit for the use of GenAI models [LLMs] in isolation."

At a high-level, this means that LLMs aren't great at fundamental business planning use cases, which cover things like logistics, marketing, staffing, investing, product development, and all sorts of other areas. Those applications require modeling of enterprise-specific, tabular and time series data that span key areas of the business, including people, products, sales, and budgets.

The industry will respond to this gap in 2025. Next year, more GenAI technologies will emerge that are engineered specifically for modeling structured numerical and statistical data rather than just text. These technologies will allow enterprises to use their tabular business data to make better decisions, minimize risk, and boost efficiencies.
Dr. Devavrat Shah
CEO of Ikigai Labs and MIT AI Professor

ON-DEVICE LLM

The next disruptive technology in AI will likely be on-device LLMs. These are lightweight LLMs that can run on the user's computer or phone. On-device LLMs will likely produce slightly worse results than 3rd party LLMs like ChatGPT, but they'll be much faster and cheaper. In addition, all of the user's data stays on their devices. Businesses will no longer need to consider whether it's cost-effective to use LLMs, since on-device LLMs are free to use. iOS, Android, and browsers are already working on supporting on-device LLMs. We're excited to see how far this technology can take us.
Leo Jiang
Staff Software Engineer, Amplitude

PRIVATE LLM

I expect (a lot) more organizations to incorporate AI assistant applications into their internal workflows. These won't be public LLMs, but rather applications powered by their own private LLMs and RAG capabilities. With private LLMs securely collecting a business' data and integrating with systems across their supply chain and business functions, these AI assistants will be ready to provide employees with valuable GenAI answers and materials across a host of use cases. In 2025, expect much more use of AI assistants to summarize documents, analyze budgets and trends, suggest project workflows and timelines, draft contracts and invoices, create business reports and marketing materials, and more.
Shomron Jacob
Head of Applied Machine Learning & Platform, Iterate.ai

SEMANTIC LAYER

The Semantic Layer Becomes the Enabler for LLMs in Enterprises: In 2025, the Semantic Layer will become the crucial enabler for LLMs in enterprises, acting as a bridge between internal data and LLMs to deliver precise, contextually relevant insights. By unifying enterprise data with global knowledge, this integration will revolutionize decision-making and productivity, making GenAI indispensable. Companies that embrace this convergence will dominate in innovation and customer experience, leaving competitors behind.
Ariel Katz
CEO, Sisense

PRE-TRAINING

Pre-Training Will Become a Key Differentiator for Organizations Adopting LLMs: By 2025, pre-training will emerge as a crucial differentiator among organizations developing large language models (LLMs). As the AI landscape evolves, access to vast amounts of high-quality data — especially industry-specific data — will become a major competitive advantage. Companies that can effectively harness big data infrastructure to leverage their large-scale datasets will be better positioned to fine-tune their models and deliver more effective, specialized solutions. However, this also introduces a significant bottleneck. Preparing and curating the right data for pre-training is increasingly complex, and companies without robust big data infrastructure will struggle to keep up. Efficiently handling this data preparation, cleaning, and transformation process will become a critical challenge in the race to develop more powerful and relevant LLMs.
Haoyuan Li
Founder and CEO, Alluxio

MULTI-MODAL AI

Multi-modal AI is set to revolutionize AI in 2025: Multi-modal AI will enable machines to process and integrate information from multiple sources like text, images, video and audio. This breakthrough will lead to more intuitive human-computer interaction, enabling us to communicate with AI seamlessly using voice, gestures, and visuals. Imagine AI assistants that understand complex requests involving multiple forms of media, or robots that can perceive and navigate their environment with human-like awareness. Furthermore, multi-modal AI will fuel a wave of innovation across industries.
Abhinav Puri
VP of Portfolio Solutions & Services, SUSE

Multi-modal AI will transform user assistance and disrupt incumbents: Multi-modal AI will see exponential improvements in 2025, making it indispensable for use cases like user assistance and analytics. By analyzing text, broader session data, and user behavior, these models will provide deeper insights and more accurate recommendations. This capability will enable startups to challenge industry leaders with AI-driven disruption while empowering enterprises to solve long-standing user experience challenges, paving the way for smarter and more personalized interactions.
James Evans
Director of Product, Amplitude

Go to: 2025 AI Predictions - Part 5

Related Links

Hot Topics