Guessing what's going to be new in AI for 2025

 AI has been the buzz technology for the last two or three years in the main stream with LLM's like chatgpt and a host of variants that flooded our market but as I like to guess/foresee where this buzz and real progress is going to take its kinda of fun based on reading to collate some thoughts here.  I will categorize this into three broad strokes. Large language models that have some kind of reasoning capability, ability to create virtual playgrounds with context and AI applications in scientific research. 

Large Language models Becoming more intelligent (able to reason ...kind of) 

Here’s where things get even more interesting. When OpenAI dropped o1 in September, it didn’t just introduce a new language model; it sparked a paradigm shift. Two months later, o3 took things further, pushing boundaries we didn’t even know existed. Unlike traditional models like GPT-4, which spit out answers as they come to mind—sometimes right, sometimes wrong—these new models are designed to think through their responses. They break down complex problems into manageable steps, trying one approach after another until they get it right. This so-called “reasoning” capability (we know, the term is loaded) is game-changing for accuracy, especially in math, physics, and logic. And let’s be honest, it’s a crucial leap for AI agents. What I like about this in practical applications is when chatbots to whatever agentic AI models we come up with are able to build on context and add that context to some sort of learning model and then keep building that knowledge-base in the background. 

Take for example  Sal Khan in his TED talk demonstrated this with the next evolution of Khan academy

Scenario: A student struggling with calculus uses an AI tutor for help. Older models provide step-by-step solutions to specific problems but don't adapt well to different learning styles or gaps in foundational knowledge.

With Reasoning AI: The AI identifies the student’s weak points, breaking complex problems into smaller, manageable steps while adjusting explanations to match their understanding. If one explanation doesn’t resonate, the AI tries another approach, like visualizing the problem with graphs or interactive models. Sal suggests that Khan Academy will have the ability like a personal tutor to carry a conversation with a student and make suggestions that help in thinking about problems and reason through it with AI chat augumenting it. 

Outcome: More effective learning tailored to individual needs.

AI Is Booming or a Boon to Boost Science

 AI is accelerating scientific discovery. Last October Nobel Prize winners in Chemistry to Demis Hassabis and John M. Jumper of Google DeepMind for their work on AlphaFold,  cracked the protein-folding problem, and to David Baker for tools that design entirely new proteins. This wasn’t just a win for AI; it was a massive step forward for humanity. So based on this trend in natual sciences so expect 2025 to bring a surge of data sets and models targeted at unlocking the mysteries of the natural world. Proteins were just the beginning because they had the perfect data sets for training AI. The hunt is on for the next big breakthrough—and it’s anyone’s guess what that will be.

Virtual Playgorunds 

If 2023 was all about generative images and 2024 brought us generative video, what’s the next frontier? You guessed it—generative virtual worlds, aka video games. And this isn’t some far-off dream; the groundwork is already being laid. Back in February, Google DeepMind gave us a sneak peek of what’s possible with their generative model, Genie, which turned a single still image into an interactive 2D side-scrolling game. Then in December, they upped the ante with Genie 2—a model capable of transforming a starter image into an entire interactive virtual world. And it’s not just Google. Other companies are racing to develop similar technology. 2025 could be the year where the line between gaming and AI artistry blurs completely.

Last February Genie 1 showed us the ability to transform a single still image into an interactive 2D side-scrolling game, similar to classic platformers like Mario Bros. This innovation allowed users to upload any image and watch it evolve into a playable environment, complete with logic and physics rules generated by AI. In December: Genie 2 expanded this capability by creating fully immersive virtual worlds from a starter image, enabling more complex environments, characters, and storylines. Imagine sketching a tree, and the AI generates an entire forest biome with interactive flora, fauna, and unique weather patterns. 

Another good example will  be Sony AI and Procedural Storytelling

• The Vision: Imagine a PlayStation game where players make decisions, and AI generates the story dynamically. Instead of pre-written narratives, every player experiences a unique storyline, shaped by their choices and the AI’s ability to adapt in real time.
• Practical Example: A fantasy RPG where your interactions with NPCs generate unique quests, alliances, and outcomes that vary dramatically from other players’ experiences.

Mojang Studios brings generative AI into Minecraft, and suddenly, building your dream world is as simple as describing it. I could just say, “Create a volcanic island with a hidden temple and treasure,” and—like magic—the game would bring it to life in real-time. Picture a fully realized environment, complete with fiery lava flows, intricate challenges, and treasures waiting to be uncovered. No painstaking block-by-block construction, just pure creativity unleashed.

2025 isn’t just another year for AI; it’s shaping up to be a year of extraordinary leaps. So, buckle up—the future’s unfolding faster than we ever imagined.