AI has been a hot topic in 2024, so how is it evolving? What are we seeing in AI today and what do we expect to see in the next 12-18 months? We asked Andrew BrustChester Comfort, Chris Ray, Dana Hernandez, Howard Holton, Ivan McPhee, Seth ByrnesWhit Walters and William McKnight weigh.
First of all, what else is popular? When is AI successful?
Chester: I see people using AI beyond experimentation. People have had the opportunity to experiment, and now we're getting to the point where real, vertical-specific use cases are being developed. I've been tracking healthcare closely and seeing more finely tuned models targeting specific use cases, such as using artificial intelligence to help doctors be more present during conversations with patients using auditory listening and note-taking tools.
I believe that “small is the new big” is a key trend, such as hematology vs. pathology vs. pulmonology. Artificial intelligence in imaging technology is nothing new, but it is now coming to the fore with new models being used to speed up cancer detection. This should be supported by the medical professional: AI cannot be the only source of diagnoses. The radiologist needs to check, verify and confirm the results.
Dana: In my reporting, I see the effective use of AI from a specific industry perspective. For example, providers specializing in finance and insurance use AI for tasks such as financial crime prevention and process automation, often through specialized, small-scale language models. These industry AI models are an important trend that I see continuing into next year.
William: We are seeing shorter cycle times in areas such as pipeline development and master data management becoming more autonomous. An area that is gaining momentum is data surveillance, and 2025 could be its year.
Andrey: Generative AI works well in the area of code generation—generating SQL queries and creating natural language interfaces to query data. This has proven to be effective, although it is now a bit commercialized.
More interesting are the advances in the data layer and architecture. For example, Postgres has a vector database add-on that is useful for advanced data generation (RAG) queries. I see a shift from the “wow” factor of demonstrations to practical use, using the right models and data to reduce hallucinations and make data more accessible. Over the next two to three years, vendors will move beyond basic query analysis to more sophisticated tools.
How can we see the development of larger language models?
Whit: Around the world, we will see AI models based on cultural and political values. It's less about technical developments and more about what we want from our AIs. Consider Elon Musk's xAI, based on Twitter/X. It's uncensored and very different from Google Gemini, which tends to lecture you if you ask the wrong question.
Different vendors, geographies, and governments will either seek greater freedom of speech or will seek to control the results of AI. The difference is noticeable. Next year we will see an increase in the number of models without guardrails, which will provide more direct answers.
Ivan: There is also a strong emphasis on structured prompts. A slight change in wording, such as using the words “detailed” instead of “comprehensive,” can lead to completely different answers. Users need to learn how to use these tools effectively.
Whit: Indeed, operational design is critical. Depending on how the words are built into the model, you can get completely different answers. If you ask the AI to explain what it wrote and why, it will make it think deeper. We'll soon see specialized hint tools – agent-based models – that can help optimize hints for better results.
How is AI developing and improving the use of data through analytics and business intelligence (BI)?
Andrey: Data is the foundation of AI. We have seen how generative artificial intelligence on large volumes of unstructured data can lead to hallucinations, and projects are curtailed. We're seeing a lot of frustration in the enterprise space, but progress is coming: we're starting to see a marriage between AI and BI that goes beyond natural language queries.
Semantic models exist in BI to make data more understandable and can be extended to structured data. In combination, we can use these models to create useful chatbot-like functionality by getting answers from structured and unstructured data sources. This approach creates business-beneficial results while reducing hallucinations through contextual improvements. This is where AI will become more grounded and data democratization will be more effective.
Howard: Agreed. BI has not yet worked perfectly over the past decade. Those involved in BI often do not understand the business, and the business does not fully understand the data, leading to friction. However, this problem cannot be solved by one generation of AI, it requires mutual understanding between both groups. Imposing data-driven approaches without this will not take organizations very far.
What other issues do you think could hinder the progress of AI?
Andrey: The euphoria over AI has diverted attention and budgets away from data science projects, which is unfortunate. Businesses need to see them as the same.
Whit: There is also an AI startup bubble—too many startups, too much funding, burning money without generating revenue. This looks like an unsustainable situation and we'll see it break down a little bit next year. The outflow is so large that it has become simply ridiculous to maintain it.
Chris: Because of this, I see vendors creating solutions to “secure” GenAI/LLM. Penetration Testing as a Service (PTaaS) providers offer LLM-focused testing, while Cloud Application Protection Providers (CNAPP) offer controls for LLM deployed to customers' cloud accounts. I don't think buyers have even begun to understand how to effectively use LLM in an enterprise, yet vendors are pushing new products/services to “secure” them. This is ripe for emergence, although some “LLM” security products/services will be widespread.
Set: When it comes to supply chain security, vendors are beginning to offer AI model analysis to identify patterns used in environments. This seems a little advanced, but it's starting to happen.
William: Another looming factor for 2025 is the EU Data Law, which will require AI systems to be able to switch off at the touch of a button. This could have a big impact on the ongoing development of artificial intelligence.
The million dollar question: How close are we to artificial general intelligence (AGI)?
Whit: AGI remains a pipe dream. We don't understand consciousness well enough to recreate it, and simply using computing power to solve a problem won't make something conscious—it'll just be a simulation.
Andrey: We may be moving towards AGI, but we need to stop thinking that predicting the next word is intelligence. This is just a statistical prediction—an impressive application, but not a truly intelligent one.
Whit: Exactly. Even when AI simulates “mind,” it is not true reasoning or creativity. They are simply recombining what they have been taught. It's about how far you can push combinatorics on a given set of data.
Thanks everyone!
Fast The Evolving Revolution: AI in 2025 first appeared on Gigaom.
