Future of AI beyond LLM scaling

The paper “Small Language Models are the Future of Agentic AI” makes a provocative argument: the future of intelligent software agents will not be built on today’s vast and expensive large language models (LLMs), but instead on smaller, more efficient models. The authors, from NVIDIA and Georgia Institute of Technology, define small language models (SLMs) as systems compact enough to run on consumer-grade devices with low latency, typically with fewer than ten billion parameters as of 2025. Unlike their heavyweight LLM cousins, SLMs are lightweight, nimble, and inexpensive to operate. The central claim is that these smaller models are not just “good enough” for most real-world agentic applications; in many cases, they are actually better suited. The reasoning rests on the nature of agentic AI tasks. Most software agents do not spend their time solving grand philosophical puzzles. Instead, they perform repeated, narrowly defined tasks, such as summarising emails, parsing documents, running queries, or automating workflows. For such repetitive, well-scoped activities, an oversized LLM is wasteful, akin to hiring a Nobel laureate to perform simple bookkeeping. Small models, properly tuned, can deliver the same functionality at far lower cost and with faster response times. The paper does not dismiss the role of LLMs altogether. There are still situations, particularly those requiring broad conversational flexibility or complex reasoning, where a large model remains indispensable. But the authors suggest a heterogeneous architecture: a swarm of specialised SLMs handling most of the workload, with larger models reserved for the exceptional cases. This hybrid approach, they argue, combines the best of both worlds: efficiency and adaptability. They are candid about the obstacles to such a shift. The AI industry has invested heavily in the LLM ecosystem—technically, commercially, and culturally. Many companies are now locked into infrastructure and business models optimised for LLMs, making change difficult. To help overcome this, the paper proposes a general conversion framework, a kind of roadmap for developers who want to adapt their existing LLM-based agents into SLM-powered systems without losing functionality. The economic and operational implications are striking. A partial migration from LLMs to SLMs could dramatically reduce costs, improve latency, and make agentic AI viable on devices at the edge, such as laptops, phones, or embedded systems, rather than requiring centralised cloud resources. In other words, smaller models may democratise the field by making intelligent agents cheaper, faster, and more widely accessible. The authors end with an open call to the research community. They do not present their thesis as the last word, but rather as the beginning of a dialogue around the central idea that small, specialised language models could drive the next wave of agentic AI. https://lnkd.in/gecggtJc

It’s time to give LLMs their due and put them in the rear-view mirror so we can watch the road ahead. We can’t drive if we keep staring backward. Under human supervision, LLMs have been useful: fluent interfaces for conversation, retrieval and summarization, coding assistance, and orchestration across tools. They made software feel conversational and pushed AI to the edges of the world with some benefits and, yes, some unintended consequences. But architectural limits don’t bend with scale. Whether it’s GPT-6 or GPT-99, they will still hallucinate, still lack real-time learning & model updates, still fail to adapt autonomously to shifting goals or conditions. Critically, they struggle to deliver what enterprises actually buy: reliability, determinism, auditability, cost predictability, and repeatable outcomes. These aren’t “product gaps” to be patched with more RAG, guardrails, or finetunes; they’re consequences of the architecture itself. “Memory” seems to be the new mantra. Memory stores personal facts, but it doesn’t rewrite the model’s internal understanding. Longer context windows and retrieval make assistants feel more personal, but they don’t create a system that updates beliefs, forms new concepts, or propagates changes through its own model. It makes the assistant chattier, not smarter. The road ahead is adaptation-first. We need systems that can change themselves responsibly in response to new evidence and do so in real time. That means: 1) Learning: incremental, continuous, self-directed, autonomous by reading, by watching, by listening, by thinking 2) Self-revision of internal models: cascading changes across beliefs, behaviors, priorities, policies, understanding, goals & actions. 3) Grounded reasoning and planning: world models that support cause-and-effect, multi-step goals, and tool use. 4) Layered memory with purpose: working, episodic, and semantic memory that improves behavior. 5) Reliability by design: zero hallucinations, verifiable steps, confidence measures, and governance loops that meet enterprise SLAs. An integrated neuro-symbolic architecture, Cognitive AI is built for exactly this: systems that learn continuously, adapt autonomously, and revise their own beliefs with traceability. While LLMs continue to be useful under human supervision, the center of gravity now shifts from scaling tokens to engineering adaptation. LLMs brought us here, Cognitive AI will take us forward.

The rise of AI agents is not a sudden breakthrough, but a steady evolution through multiple layers of capability: LLM Processing Flow – Basic input-to-output transformation. LLM with Document Processing – Expanding scope to handle structured and unstructured documents. LLM with RAGs & Tools – Introducing retrieval, tool-use, and external knowledge integration. Multi-Modal Workflows – Combining text, vision, and audio with emerging memory structures. Advanced Architectures – Incorporating decision-making, orchestration of tools, and multi-level memory (short-term, long-term, episodic). Future AI Agents – Moving beyond capability toward responsibility: safety, ethics, regulation, compliance, interoperability, and human collaboration. This progression highlights a clear trajectory: from narrow assistants to autonomous, enterprise-ready agents that operate within a framework of trust, governance, and accountability. The challenge now is not whether we can reach stage six, but how we ensure safety and control while advancing toward it.

I recently spoke to Gartner about what is next in #AI. Here are my thoughts: We have seen impressive progress in #llm by scaling data and compute. Will this continue to hold? Yes, I believe so, but most of those gains will be in reasoning tasks where we have precise metrics to measure uplift, as well as the ability to have synthetic data to train further, and also the freedom to trade off computation for accuracy at test time. This is seen in the recent o1 model. For reasoning tasks, we will also be able to remove hallucination when we can construct accurate verifiers that can certify every statement that #llm makes. We have been doing this in our Leandojo project for mathematical theorem proving. However, there is one area of reasoning where #llm will never be good enough: understanding the physical world. This is because language is only high-level knowledge, and cannot simulate the complex physical phenomena needed in many applications. For instance, LLMs can talk about playing tennis or look up a weather app, but they cannot internally simulate any of these processes. While images and videos can help improve their knowledge of the physical world, models like Sora learn physics by accident, and hence, still produce physically wrong outputs. How can we overcome this? By teaching AI physics from the ground up. We are building AI models that are trained in a physics-informed manner at multiple scales. They are several orders of magnitude faster than traditional simulations, and can also generate novel designs that are physically valid. You can watch some of those examples in my recent TED talk.

🔧 AI agents are taking off. But we may be building them all wrong. NVIDIA’s latest research suggests we’ve been scaling agents inefficiently: ➡️ It’s not large language models (LLMs) that will scale agentic AI. ➡️ It’s Small Language Models (SLMs) — compact, local, and radically cheaper. That insight forced me to stop and rethink everything. I’ve seen too many teams build agents that call GPT-4… for everything. Even for basic, predictable tasks like: → Formatting JSON → Extracting a few values → Generating API calls Why? Because it’s easy. But it’s also wasteful. We're burning compute — and budgets — on jobs that don’t need a genius to do them. 🔍 NVIDIA’s findings are a wake-up call: ⚡SLMs like Phi-3 and DeepSeek-7B are crushing older LLMs ⚙️ Toolformer (6.7B) outperformed GPT-3 (175B) 🧠 DeepSeek-7B beat GPT-4o on reasoning 📉 40–70% of LLM calls can already be swapped for SLMs And the upside? ✅ 10–30x cheaper inference ✅ No GPUs, no clusters — run on laptops ✅ Fine-tune overnight (LoRA/QLoRA) ✅ Less hallucination, better structure ✅ More modular and scalable system design 🛠 What this means for us: The AI industry has poured billions into LLM infrastructure. But that may soon feel like building a spaceship to cross the street. I’m rethinking my own approach: → Start with SLMs for agent sub-tasks → Only fall back to LLMs when truly necessary → Embrace modular, specialized design Because here’s the truth: Bigger isn’t always better. Smaller is often smarter. Curious to hear your take: Are we finally reaching the post-LLM era for agent design? 🔗 Full paper > https://zurl.co/kCwWU #AI #AgenticAI #SLMs #Automation #FutureOfWork #NVIDIA #LLMs #AIEngineering #CostEfficiency #AIArchitecture

DeepSeek's recent developments have ignited significant discussion in the AI community, and I wanted to take a minute to share some thoughts. If you haven’t heard, the company's latest model, R1, showcases a reasoning capability comparable to OpenAI's o1, but with a notable distinction: DeepSeek claims that their model was trained for much less cost.   It isn’t clear yet if DeepSeek is the real deal or a DeepFake, but regardless of what we learn in the coming days, it’s clear that this is a wake up call -- the path of bigger and bigger LLMs that rely on ever-increasing GPUs and massive amounts of energy is not the only path forward. In fact, it’s clear there is very limited upside to that approach, for a few reasons:   ⭐️ First, pure scaling of LLMs at training time has reached the point of diminishing or maybe near zero returns. Bigger models trained with more data are not resulting in meaningful improvements.   🤔 Further, enterprises don’t need huge, ask-me-anything LLMs for most use cases. Even prior to DeepSeek, there's a noticeable shift towards smaller, more specialized models tailored to specific business needs. As more enterprise AI use cases emerge, it becomes more about inference -- actually running the models to drive value. In many cases, that will happen at the edge of the internet, close to end users. Smaller models that are optimized to run on commodity hardware are going to create more value, long-term, than over-sized LLMs.   💡 Finally, the LLM space is ripe for optimization. The AI models we have seen so far have focused on innovation by scaling at any cost. Efficiency, specialization, and resource optimization are once again taking center stage, a signal that AI’s future lies not in brute force alone, but in how strategically and efficiently that power is deployed.

The future of AI isn't just about bigger models. It's about smarter, smaller, and more private ones. And a new paper from NVIDIA just threw a massive log on that fire. 🔥 For years, I've been championing the power of Small Language Models (SLMs). It’s a cornerstone of the work I led at Google, which resulted in the release of Gemma, and it’s a principle I’ve guided many companies on. The idea is simple but revolutionary: bring AI local. Why does this matter so much? 👉 Privacy by Design: When an AI model runs on your device, your data stays with you. No more sending sensitive information to the cloud. This is a game-changer for both personal and enterprise applications. 👉 Blazing Performance: Forget latency. On-device SLMs offer real-time responses, which are critical for creating seamless and responsive agentic AI systems. 👉 Effortless Fine-Tuning: SLMs can be rapidly and inexpensively adapted to specialized tasks. This agility means you can build highly effective, expert AI agents for specific needs instead of relying on a one-size-fits-all approach. NVIDIA's latest research, "Small Language Models are the Future of Agentic AI," validates this vision entirely. They argue that for the majority of tasks performed by AI agents—which are often repetitive and specialized—SLMs are not just sufficient, they are "inherently more suitable, and necessarily more economical." Link: https://lnkd.in/gVnuZHqG This isn't just a niche opinion anymore. With NVIDIA putting its weight behind this and even OpenAI releasing open-weight models like GPT-OSS, the trend is undeniable. The era of giant, centralized AI is making way for a more distributed, efficient, and private future. This is more than a technical shift; it's a strategic one. Companies that recognize this will have a massive competitive advantage. Want to understand how to leverage this for your business? ➡️ Follow me for more insights into the future of AI. ➡️ DM me to discuss how my advisory services can help you navigate this transition and build a powerful, private AI strategy. And if you want to get hands-on, stay tuned for my upcoming courses on building agentic AI using Gemma for local, private, and powerful agents! #AI #AgenticAI #SLM #Gemma #FutureOfAI

The future of AI won’t be LLMs or world models... it'll be both. The world is captivated by Large Language Models. GPT, Claude, Gemini; trillion-parameter oracles trained on the internet’s collective prose. They can debate philosophy, write code, and pass bar exams. The prevailing belief: scale them big enough, and human-level intelligence will emerge. Yann LeCun isn’t convinced. His argument is blunt: predicting the next token in text differs from understanding how the world works. LLMs rest on the idea that language is the roadmap to intelligence. Model the statistical structure of human communication at scale, and reasoning will follow. It has delivered jaw-dropping results. But language is a map, not the territory. LLMs can describe the world in exquisite detail but have never "seen" it. JEPA, the Joint Embedding Predictive Architecture, flips the script. Instead of generating words or pixels, it predicts latent representations of what comes next in sensory streams like vision or video. It ignores noise, focuses on causal structure, and builds grounded internal models of reality. The goal is comprehension, not chatter. However, JEPA remains largely untested at scale outside of vision and video. This is more than a technical debate. It’s a philosophical one about the minds we are building. LLMs are reactive conversationalists with vast knowledge but no explicit planning loop. JEPAs are predictive explorers with an emerging sense of cause and effect but limited fluency. In the near term, LLMs will dominate language-heavy domains. However, hybrids are coming: LLMs for symbolic reasoning and JEPA-like modules for grounded simulation. In the long term, the two will converge into unified cognitive systems where abstract reasoning and physical understanding are inseparable. This raises the uncomfortable question: if the path to AGI is a fusion of the two, are we pouring too much into the talkers and not enough into the doers? The first team to fuse them into a single mind may well define the next era of AI. The only real question is, who gets there first?

Yann LeCun is not critiquing AI from the sidelines. He helped build the foundations of modern machine learning, from pioneering convolutional neural networks to shaping how vision systems work today, and has served as Meta’s Chief AI Scientist while holding a long academic career at NYU. His perspective is grounded in decades of trying to make machines learn in ways that resemble how humans and animals actually do. That history explains his skepticism. The future of AI will not be won by brute force. Yann LeCun makes a clear and often uncomfortable point: scaling today’s large language models with more data and more compute will not deliver human level intelligence. LLMs are remarkable pattern recognizers, but they remain sample inefficient, prone to hallucination, and disconnected from any real understanding of the physical world. They generate. They do not truly reason, plan, or understand. What comes next requires a shift, not an upgrade. LeCun argues that meaningful progress depends on AI systems that learn predictive world models from perception and interaction, not just text. Intelligence that can plan toward goals, learn continuously from experience, and reason beyond the statistical echoes of training data. In his view, AGI will not “emerge” from scale alone, and apocalyptic runaway superintelligence narratives say more about science fiction than science. Equally important is how this intelligence is positioned. LeCun frames the future as amplifier intelligence: AI as a force multiplier for human capability, not a replacement for human agency. Powerful assistants, firmly under human direction. He also remains a strong advocate for open research and open source models as both a scientific necessity and a societal safeguard as capabilities increase. The takeaway is sobering and optimistic at the same time. The next era of AI will be harder, slower, and more architectural than the last. But it may also be more grounded, more useful, and far more aligned with how intelligence actually works. I’m Colin, General Counsel of Malbek CLM for the Enterprise and author of The Legal Tech Ecosystem. #legaltech #innovation #law #business #learning

Ilya Sutskever has confirmed that LLMs scaling has plateaued. Compute continues to scale, but data is not keeping up. New or synthetic data hasn’t significantly improved performance. What’s next: Like the human brain, which stopped growing in size while humanity advanced through tools and knowledge, progress will come from agents and tools built on top of LLMs. Key areas of focus include: Sequence-to-sequence learning Agentic behavior Teaching self-awareness This shift is similar to the evolution of the iPhone: after hardware improvements plateaued, the focus moved to applications. 2025 will be the year of agents. ↓ Are you technical? Check out https://AlphaSignal.ai to get a daily summary of breakthrough models, repos and papers in AI. Read by 200,000+ devs.