LLMs to LRMs

Amazon has laucnehd their AI Olympus for ads. it is currently in Beta format at the moment. Olympus reportedly features 2 trillion parameters, allowing it to analyze large visual content in images and videos and search for specific scenes using text prompts.

In a move to counter some of the use cases for NotebookLM Eleven Labs has aunched GenFM which can create an AI audio version your pdf docs.

Google Gemini will receive a significant update for its code analyzing functionalitites.

Just in time for Christmas testing Perplexity has released Perplexity Shopping, for searching and shopping.

Intel has written their own guide to surviving in a world of AI.

The Allen Institute has released an open surce model that beats LLAMA 3.1. I personally cannot wait to try this tool out. I have been working alot with Gemini Flash recently and I have been surprised by the quality of the output I am getting.

The open source AI assistant Khoj now integrates with Obsidian and Whatsapp.

Facing a changing climate, our focus shifts from prevention to adaptation. Trees are crucial, offering carbon sequestration and vital shade in a warming world. But effective urban forestry requires sophisticated planning, necessitating accurate models.

A groundbreaking collaboration between MIT, Purdue University, and Google has delivered Tree-D Fusion: the first large-scale database (over 600,000) of 3D-modeled North American trees, ready for environmental simulations.

Leveraging generative AI, this system not only identifies trees but predicts their growth and environmental impact. This allows city planners to proactively strategize tree placement, transforming urban forestry from reactive maintenance to anticipatory planning.

The researchers envision expanding Tree-D Fusion globally, using AI to support biodiversity, promote sustainability, and ultimately, improve planetary health.

From LLMs to LRMs

The AI industry is buzzing with a paradigm shift. The term "LLM" (Large Language Model) that once dominated conversations is being replaced by "LRM," or Large Reasoner Model. These new models emphasize post-training enhancements and are transforming the field. Even Microsoft’s CEO recently endorsed the concept of "inference-time compute," a nod to the limitations of pre-training and the rise of this new approach.

The past few weeks have seen a flurry of activity, with companies like OpenAI, Alibaba, MIT, and others unveiling their own LRM frameworks. These models signal the dawn of a new era, one focused not on simply building larger models but on making smarter ones that can adapt and reason in real-time.

Why LLMs Are Losing Their Shine

Scaling laws, once the driving force behind AI improvements, have plateaued. The traditional approach of expanding model sizes and training them on massive datasets no longer yields significant gains. Today’s LLMs face fundamental limitations:

• They can't learn on the fly: Imagine living your whole life based on what you learned in school, unable to acquire new knowledge as the world changes. That’s the reality for current models.

• They are bound by their training data: LLMs excel only when encountering data or tasks similar to what they've seen before. Outside of this familiar territory, they falter.

These challenges have pushed researchers to explore post-training methods that can overcome these bottlenecks.

Enter the LRM: The Next Evolution

LRMs represent a significant leap forward. Instead of stopping at pre-training, these models undergo advanced techniques that refine their reasoning capabilities. Here’s what sets LRMs apart:

1. Memory Updates:
   Traditional LLMs rely on static knowledge baked into their training. LRMs, on the other hand, can dynamically update their memory. For example, Writer's LRM approach allows real-time updates, enabling the model to adapt and improve its performance when it encounters new information.

2. Graph-Based Reasoning:
   By presenting data in graph form, LRMs can extract and connect key ideas more effectively. This technique taps into common patterns, or isomorphisms, across seemingly unrelated topics, enhancing the model's ability to generalize. For instance, an LRM exposed to graph-structured knowledge can reason about unfamiliar data by recognizing underlying similarities.

3. Active Learning During Inference:
   Unlike traditional models, LRMs can learn and refine their knowledge as they operate. This process, called test-time training, allows the model to adapt to new tasks or facts dynamically. Imagine practicing a basketball shot a few times before taking the real shot—LRMs essentially do this during their reasoning process.

The Role of "Inference-Time Compute"

A defining feature of LRMs is their reliance on "inference-time compute." This means they allocate more computational resources during the reasoning phase rather than just the training phase. By giving the model "more time to think," LRMs can:

• Update their memory in real-time to incorporate new information.

• Explore multiple solutions before delivering the best answer.

• Actively refine their understanding to learn new skills on the go.

This shift marks a fundamental change in how AI operates, moving away from static knowledge to dynamic reasoning.

Cutting-Edge Techniques Driving LRMs

Beyond the foundational principles of memory updates and active learning, several advanced techniques are setting the stage for the next generation of AI:

• Reinforcement Learning from Verifiable Rewards (RLVR):
  Developed by the Allen Institute, this method trains models using tasks with clearly verifiable outcomes. For example, in areas like math or coding, an automatic verifier provides feedback that helps the model improve over time.

• GFlowNets:
  Considered a groundbreaking learning algorithm, GFlowNets offer a new way for AI to reason and sample information. This technique has garnered praise from leading researchers like Yoshua Bengio, who see it as a path to unlocking genuine reasoning in AI.

The Road Ahead: What LRMs Mean for the Future

By 2025, AI development will look very different. The focus will shift from building ever-larger models to refining their ability to think and adapt. LRMs are the key to this transformation, enabling models to overcome current limitations and unlocking new capabilities.

These advancements will redefine how we interact with AI. Models will no longer be static repositories of knowledge but dynamic systems capable of reasoning, learning, and evolving in real-time. Whether it’s updating memory, solving complex problems, or adapting to new environments, LRMs promise to reshape the AI landscape.

As we dive deeper into this new era, the question is not whether LRMs will replace LLMs, but how quickly they will become the new standard. The age of static AI is over, and the future belongs to models that can truly think.