The Future of AI in 2025: Expert Perspectives from Srini Penchikala
In this interview, InfoQ Editor Srini Penchikala offers an insider's view of AI's horizon in 2025. He sees Small Language Models democratizing AI, making powerful technologies more accessible and privacy-friendly and believes that AI Agents will emerge as game-changers, potentially automating complex workflows with unprecedented intelligence. Penchikala also envisions specialized databases and domain-specific solutions transforming how businesses leverage AI, and urges marketers to showcase practical, impactful applications that solve real-world challenges.
Hi Srini, can you tell us a little bit about yourself and your role at InfoQ?
I serve as the Lead Editor for the AI Infrastructure Community on InfoQ. I have the honor and privilege of working with great folks in the AI space who are subject matter experts and practitioners in their field. With so much happening in AI/ML area, especially in Generative AI (GenAI) space, for the last few years, it's a great time to be part of this evolution as we see AI technologies become more and more integrated with every day activities whether it's in the business, at work or at home. I learn something new everyday from the excellent team at InfoQ.
With the advent of ChatGPT, we’ve been hearing a lot about Large Language Models (LLMs). What are Small Language Models and why are they important?
The AI space which was accelerated in terms of growth and adoption when GPT and ChatGPT were released back in 2022, GenAI and LLMs have pretty much taken over the AI landscape.
Large Language Models (LM) are the foundation of AI applications and play a critical role in developing the solution. LLMs at a high level, are language models with many parameters, and are pre-trained on a large amount of data. They typically contain billions of parameters and are capable of understanding and generating content (text, image or audio) based on input prompt statements. These models are very expensive to create, but they are very powerful in the capabilities they offer to solve AI/ML problems in the business domains.
We are witnessing LLMs being dominantly used in various business and technology use cases to create exciting new opportunities that we have never even thought about before. Not only for the end users but also for software engineers, the DevOps engineers and other groups to be more productive. LLMs have empowered pretty much the full life cycle of the process.
LLMs are powerful but they require significant computing resources to operate and you cannot just run it on a small machine or a small cluster of machines. There's also the privacy and data security factors where you may not be able to send the data to the cloud to leverage these LLMs.
This is where Small Language Models or SLMs can help. They offer many of the same benefits as LLMs, but they're smaller in size, they're trained using smaller data sets and they don't require a lot of computing resources. These models are obviously not for every use case out there. They are valuable for use cases where you have a constraint on the resources or you want to localize the model execution.
Another important technology to mention in the context of LLMs is the Retrieval Augmented Generation or RAG. With RAG you can train the base model with your company's private information and then, you can use the new self-hosted model to ask questions and prompts that are specific to your business domain. We'll be seeing a lot more applications using RAG in 2025. Already a majority of production scale AI applications are using RAG.
AI agents have come a long way over the last 3-5 years. How do you feel they have matured and what’s their potential for automating software development processes?
There are a lot of different and diverse definitions of AI agents. With all the power of LLMs, the AI process execution can be vastly enhanced by a program that "acts" on the output of an LLM. These programs are called AI agents which can call, as part of a multi-step workflow, a certain task in order to implement a step in the workflow (https://huggingface.co/blog/smolagents).
Agent based applications can automate the end-to-end AI pipelines with multiple agents working together to accomplish the goals of the overall process. Another advantage of muti-agent based workflow applications is that each agent can talk to the agents in a feedback loop that enriches the agents to iteratively perform the steps better and better in every round and more importantly minimize the hallucinations.
Companies like OpenAI are going to have AI Agents as a big part of their strategy this year.
“The concept of AI agents is a big leap from current capabilities that require step-by-step prompting and a significant step towards superintelligence or AGI (Artificial General Intelligence) which is human-level or beyond human-level intelligence.” (Source)
I personally think 2025 will be the year of AI Agents as they can take the LLM applications to the next level by automatically taking actions based on the LLM output, instead of human users manually providing the input prompts in each and every step of the process.
With distributed databases underpinning many AI solutions, how do you see database vendors evolving their offerings to support AI-driven use cases?
Databases and Data Engineering processes are still important components of AI solutions. All AI applications require some type of data storage and data processing, so the database space has also been going through major transformation and innovations during the explosion of AI adoption in diverse applications.
Gen AI solutions have led to the growth and adoption of purpose driven data stores like Vector Databases. A vector database deals with types of data called Vectors which are the output of machine learning models and the way Generative AI models represent any type of data. These databases are highly specialized to work with vectors with specialized queries that search and find content by relevance, by similarity, by alignment in the numerical representation of vector embeddings.
A typical LLM based RAG application pipeline includes steps like Tokenization, Embedded Model, Language Model, Data Chunking, Orchestration, Retrieval, and Generation. So, there is a lot of areas that databases supporting LLM applications can innovate and improve in the next couple of years.
Also, speaking of RAG applications the data thats stored and processed in the real world is not always textual data, it can be audio, images, and video type of content that will require multi-model RAG solutions to solve the business problems. This is where the database vendors can focus on helping the developer community to efficiently store and process different types of data.
In addition to specialized databases like Vector databases, other data stores such as graph databases are seeing some renewed interest in AI solutions using RAG techniques. Since Graph NoSQL databases store the data in a graph format, these databases are a perfect complement to support LLM applications. We are going to see more and more Graph RAG based applications, to evolve from isolated nodes in a graph database to graph of knowledge (GoK) and finally Knowledge Graph RAG based applications.
I consider last year 2024 as the year of RAG in terms of AI adoption. In 2025, I envision the data storage solutions will see a significant growth in supporting the real world business application data to keep up with the ever growing "Omni-model RAG" solutions.
How can marketers effectively communicate the benefits of AI-driven tools to non-technical decision-makers, ensuring they understand both the potential and the practical applications?
Even with all the hype out there, AI technologies do offer significant advantages and productivity benefits in a plethora of applications and use cases. AI solutions can help solve many business and technical use cases. This is what marketers can focus on when communicating with non-technical decision makers and customers, highlighting the business problems and opportunities the AI applications can solve without getting into technical details.
With growing concerns about data privacy and responsible AI, what steps should vendors take to build trust with their audience, and how can marketers highlight these efforts effectively?
AI hallucinations, where AI models output inaccurate, non-relevant or misleading results, are a big concern when using AI applications. This can be minimized with practices like using high-quality training data, specifically defining the context on how the AI models will be used, and continuous testing and training of the model.
But a more comprehensive and strategic approach is to use AI Agents to address the AI hallucinations. Multi-agent workflows can be used to iteratively and incrementally improve the accuracy of the models and reduce the biases.
Also, AI governance is more important than ever to ensure the quality and fairness of AI solutions.
And, finally Explainable AI with its core principles of transparency, interpretability, causality, and fairness, can help the users trust the results and output created by AI applications.
Marketers and advocates can highlight to their audience these techniques to ensure that, if properly trained, executed, and governed, AI programs can be very powerful and accountable.
What do you see as the biggest opportunities for AI/ML software vendors in 2025?
In 2025, AI/ML software vendors will need to align their strategies and products with the development and adoption trends of AI technologies. Some of these focus areas are listed below:
More Business Domain Specific AI solutions (e.g. we may start to see domain specific LLM’s like a foundation model for healthcare industry to start taking shape)
Full SDLC Lifecycle
Multi-model RAG solutions
AI Agents and Agent based workflow solutions
Edge AI with AI powered devices
More Self-Hosted LLM models to deploy locally inside the organizations’ data centers or in private cloud w/o having to depend on cloud based commercial solutions.
AI Infrastructure will see significant changes to accommodate the developments in language models. The infrastructure will also start to embrace sustainable and green computing strategies to bring down the cost of training models.
As AI continues to evolve, how should marketers approach balancing the hype around new technologies with realistic, actionable benefits for their target audience?
The latest developments in AI technologies offer significant benefits to consumers, end users and IT professionals in all domains and sectors. But no doubt there is a lot of hype about AI. Marketers need to communicate what their audience should be excited about without getting distracted and disrupted by all the hype out there.
This can be done by focusing on the true value of AI technologies especially with trends like Generative AI and Languages Models (Large, Small and Micro) can bring to day-to-day applications.
Edge computing is another area where language models, especially small language models will shine. Edge AI programs will be hosted and executed on devices like smartphones, tablets, laptops etc., bringing the power of AI down to the devices at the hands of everyday users.
Marketers should highlight the business and technology use cases where AI technologies outperform the traditional approaches whether it’s healthcare, manufacturing, financial services or any other business domain.