Insights From Recent AI Research

A lot has changed since my #LLM inference article last January—it’s hard to believe a year has passed! The AI industry has pivoted from focusing solely on scaling model sizes to enhancing reasoning abilities during inference. This shift is driven by the recognition that simply increasing model parameters yields diminishing returns and that improving inference capabilities can lead to more efficient and intelligent AI systems. OpenAI's o1 and Google's Gemini 2.0 are examples of models that employ #InferenceTimeCompute. Some techniques include best-of-N sampling, which generates multiple outputs and selects the best one; iterative refinement, which allows the model to improve its initial answers; and speculative decoding. Self-verification lets the model check its own output, while adaptive inference-time computation dynamically allocates extra #GPU resources for challenging prompts. These methods represent a significant step toward more reasoning-driven inference. Another exciting trend is #AgenticWorkflows, where an AI agent, a SW program running on an inference server, breaks the queried task into multiple small tasks without requiring complex user prompts (prompt engineering may see end of life this year!). It then autonomously plans, executes, and monitors these tasks. In this process, it may run inference multiple times on the model while maintaining context across the runs. #TestTimeTraining takes things further by adapting models on the fly. This technique fine-tunes the model for new inputs, enhancing its performance. These advancements can complement each other. For example, an AI system may use agentic workflow to break down a task, apply inference-time computing to generate high-quality outputs at each step and employ test-time training to learn unexpected challenges. The result? Systems that are faster, smarter, and more adaptable. What does this mean for inference hardware and networking gear? Previously, most open-source models barely needed one GPU server, and inference was often done in front-end networks or by reusing the training networks. However, as the computational complexity of inference increases, more focus will be on building scale-up systems with hundreds of tightly interconnected GPUs or accelerators for inference flows. While Nvidia GPUs continue to dominate, other accelerators, especially from hyperscalers, would likely gain traction. Networking remains a critical piece of the puzzle. Can #Ethernet, with enhancements like compressed headers, link retries, and reduced latencies, rise to meet the demands of these scale-up systems? Or will we see a fragmented ecosystem of switches for non-Nvdia scale-up systems? My bet is on Ethernet. Its ubiquity makes it a strong contender for the job... Reflecting on the past year, it’s clear that AI progress isn’t just about making things bigger but smarter. The future looks more exciting as we rethink models, hardware, and networking. Here’s to what the 2025 will bring!

Top 10 research trends from the State of AI 2024 report: ✨Convergence in Model Performance: The gap between leading frontier AI models, such as OpenAI's o1 and competitors like Claude 3.5 Sonnet, Gemini 1.5, and Grok 2, is closing. While models are becoming similarly capable, especially in coding and factual recall, subtle differences remain in reasoning and open-ended problem-solving. ✨Planning and Reasoning: LLMs are evolving to incorporate more advanced reasoning techniques, such as chain-of-thought reasoning. OpenAI's o1, for instance, uses RL to improve reasoning in complex tasks like multi-layered math, coding, and scientific problems, positioning it as a standout in logical tasks. ✨Multimodal Research: Foundation models are breaking out of the language-only realm to integrate with multimodal domains like biology, genomics, mathematics, and neuroscience. Models like Llama 3.2, equipped with multimodal capabilities, are able to handle increasingly complex tasks in various scientific fields. ✨Model Shrinking: Research shows that it's possible to prune large AI models (removing layers or neurons) without significant performance losses, enabling more efficient models for on-device deployment. This is crucial for edge AI applications on devices like smartphones. ✨Rise of Distilled Models: Distillation, a process where smaller models are trained to replicate the behavior of larger models, has become a key technique. Companies like Google have embraced this for their Gemini models, reducing computational requirements without sacrificing performance. ✨Synthetic Data Adoption: Synthetic data, previously met with skepticism, is now widely used for training large models, especially when real data is limited. It plays a crucial role in training smaller, on-device models and has proven effective in generating high-quality instruction datasets. ✨Benchmarking Challenges: A significant trend is the scrutiny and improvement of benchmarks used to evaluate AI models. Concerns about data contamination, particularly in well-used benchmarks like GSM8K, have led to re-evaluations and new, more robust testing methods. ✨RL and Open-Ended Learning: RL continues to gain traction, with applications in improving LLM-based agents. Models are increasingly being designed to exhibit open-ended learning, allowing them to evolve and adapt to new tasks and environments. ✨Chinese Competition: Despite US sanctions, Chinese AI labs are making significant strides in model development, showing strong results in areas like coding and math, gaining traction on international leaderboards. ✨Advances in Protein and Drug Design: AI models are being successfully applied to biological domains, particularly in protein folding and drug discovery. AlphaFold 3 and its competitors are pushing the boundaries of biological interaction modeling, helping researchers understand complex molecular structures and interactions. #StateofAIReport2024 #AITrends #AI

Stanford HAI just released the 2025 AI Index Report — and it’s a compelling snapshot of where AI is headed. If you're building, investing in, or regulating AI, this report is a must-read. It captures both mainstream momentum and emerging outliers that will shape the next wave of innovation. Here are the highlights that stood out to me — along with a few surprises: Model development is accelerating: The U.S. led with 40 notable models in 2024, while China developed 15. But what’s notable is that the performance gap is narrowing fast — Chinese models are now scoring near-parity with U.S. counterparts on benchmarks like MMLU and HumanEval. Private AI investment soared: U.S. – $67.2B China – $7.8B U.K. – $4.5B The capital flow shows no signs of slowing, and the geopolitical implications are hard to ignore. AI adoption surged: A full 78% of organizations reported using AI in 2024 — up from 55% the year before. AI has officially gone mainstream in enterprise. Massive efficiency gains: 40% improvement in AI hardware energy efficiency 280x drop in inference cost for GPT-3.5–level models (Nov 2022 to Oct 2024) This is reshaping the economics of AI at scale. The regulation wave is building: The U.S. issued 59 AI-related federal regulations in 2024 — double the previous year. AI legislative mentions rose 21.3% across 75 countries — a sign of how urgently governments are responding. Now for the outliers and trends that deserve your attention: DeepSeek’s R1 model in China hit near state-of-the-art performance using a fraction of the compute. This is especially striking given U.S. export restrictions — and challenges our assumptions about scale and access. AI is becoming a global movement. Nations in Southeast Asia, the Middle East, and Latin America are now building serious AI capabilities. This decentralization of innovation is just getting started. Open-weight models are surging. Llama (Meta), DeepSeek, and others are driving the shift toward open access — fueling grassroots experimentation and enterprise adoption alike. But risks are rising, too. The report documents a growing number of AI-related incidents and model failures — underscoring the urgency of safety, governance, and responsible deployment. Reasoning remains a challenge. Even the most advanced models still struggle with complex logic and contextual decision-making — making it clear that true autonomy is still a frontier, not a given. TL;DR? AI is scaling, spreading, and getting smarter — but the risks and responsibilities are scaling with it. And the next big breakthrough might not come from where we expect. Here’s the full report: https://lnkd.in/gUeYMWAv Which of these trends do you think will shape 2025 the most? Curious to hear your take.

I recently delved into some intriguing research about the often-overlooked potential of Small Language Models (SLMs). While LLMs usually grab the headlines with their impressive capabilities, studies on SLMs fascinate me because they challenge the “bigger is better” mindset. They highlight scenarios where smaller, specialized models not only hold their own but actually outperform their larger counterparts. Here are some key insights from the research: 𝟏. 𝐑𝐞𝐚𝐥-𝐓𝐢𝐦𝐞, 𝐏𝐫𝐢𝐯𝐚𝐜𝐲-𝐅𝐨𝐜𝐮𝐬𝐞𝐝 𝐀𝐩𝐩𝐥𝐢𝐜𝐚𝐭𝐢𝐨𝐧𝐬: SLMs excel in situations where data privacy and low latency are critical. Imagine mobile apps that need to process personal data locally or customer support bots requiring instant, accurate responses. SLMs can deliver high-quality results without sending sensitive information to the cloud, thus enhancing data security and reducing response times. 𝟐. 𝐒𝐩𝐞𝐜𝐢𝐚𝐥𝐢𝐳𝐞𝐝, 𝐃𝐨𝐦𝐚𝐢𝐧-𝐒𝐩𝐞𝐜𝐢𝐟𝐢𝐜 𝐓𝐚𝐬𝐤𝐬: In industries like healthcare, finance, and law, accuracy and relevance are paramount. SLMs can be fine-tuned on targeted datasets, often outperforming general LLMs for specific tasks while using a fraction of the computational resources. For example, an SLM trained on medical terminology can provide precise and actionable insights without the overhead of a massive model. 𝟑. 𝐀𝐝𝐯𝐚𝐧𝐜𝐞𝐝 𝐓𝐞𝐜𝐡𝐧𝐢𝐪𝐮𝐞𝐬 𝐟𝐨𝐫 𝐋𝐢𝐠𝐡𝐭𝐰𝐞𝐢𝐠𝐡𝐭 𝐀𝐈: SLMs leverage sophisticated methods to maintain high performance despite their smaller size: • Pruning: Eliminates redundant parameters to streamline the model. • Knowledge Distillation: Transfers essential knowledge from larger models to smaller ones, capturing the “best of both worlds.” • Quantization: Reduces memory usage by lowering the precision of non-critical parameters without sacrificing accuracy. These techniques enable SLMs to run efficiently on edge devices where memory and processing power are limited. Despite these advantages, the industry often defaults to LLMs due to a few prevalent mindsets: • “Bigger is Better” Mentality: There’s a common belief that larger models are inherently superior, even when an SLM could perform just as well or better for specific tasks. • Familiarity Bias: Teams accustomed to working with LLMs may overlook the advanced techniques that make SLMs so effective. • One-Size-Fits-All Approach: The allure of a universal solution often overshadows the benefits of a tailored model. Perhaps it’s time to rethink our approach and adopt a “right model for the right task” mindset. By making AI faster, more accessible, and more resource-efficient, SLMs open doors across industries that previously found LLMs too costly or impractical. What are your thoughts on the role of SLMs in the future of AI? Have you encountered situations where a smaller model outperformed a larger one? I’d love to hear your experiences and insights.

Folks interested in AI / AI PM, I recommend watching this recent session by the awesome Aishwarya Naresh Reganti talking about Gen AI Trends. ANR is a "Top Voice" that I follow regularly, leverage her awesome GitHub repository, consume her Instagram shorts like candy and looking forward to her upcoming Maven Course on AI Engineering. https://lnkd.in/g4DiZXBU Aishwarya highlights the growing importance of prompt engineering, particularly goal engineering, where AI agents break down complex tasks into smaller steps and self-prompt to achieve higher-order goals. This trend reduces the need for users to have extensive prompt engineering skills. In the model layer, she discusses the rise of small language models (SLMs) that achieve impressive performance with less computational power, often through knowledge distillation from larger models. Multimodal foundation models are also gaining traction, with research focusing on integrating text, images, videos, and audio seamlessly. Aishwarya emphasizes Retrieval Augmented Generation (RAG) as a successful application of LLMs in the enterprise. She notes ongoing research to improve RAG's efficiency and accuracy, including better retrieval methods and noise handling. AI agents are discussed in detail, with a focus on their potential and current limitations in real-world deployments. Finally, Aishwarya provides advice for staying updated on AI research, recommending focusing on reliable sources like Hugging Face and prioritizing papers relevant to one's specific interests. She also touches upon the evolving concept of "trust scores" for AI models and the importance of actionable evaluation metrics. Key Takeaways: Goal Engineering: AI agents are learning to break down complex tasks into smaller steps, reducing the need for users to have extensive prompt engineering skills. Small Language Models (SLMs): SLMs are achieving impressive performance with less computational power, often by learning from larger models. Multimodal Foundation Models: These models are integrating text, images, videos, and audio seamlessly. Retrieval Augmented Generation (RAG): RAG is a key application of LLMs in the enterprise, with ongoing research to improve its efficiency and accuracy. AI Agents: AI agents have great potential but face limitations in real-world deployments due to challenges like novelty and evolution. Staying Updated: Focus on reliable sources like Hugging Face and prioritize papers relevant to your interests. 🤔 Trust Scores: The concept of "trust scores" for AI models is evolving, emphasizing the importance of actionable evaluation metrics. 📏 Context Length: Models can now handle much larger amounts of input text, enabling more complex tasks. 💰 Cost: The cost of using AI models is decreasing, making fine-tuning more accessible. 📚 Modularity: The trend is moving towards using multiple smaller AI models working together instead of one large model.

Recommended report: World Intellectual Property Organization (2024). #Generative #Artificial #Intelligence. Patent Landscape Report. Geneva. 💡 This report provides an analysis of patenting activity and scientific publications in the field of Generative Artificial Intelligence (#GenAI). It aims to shed light on the current technology development, key players, and potential applications of GenAI technologies. 👍🏼 #Goals: - Examine the global development and trends in GenAI patenting and research - Analyze patent trends for different GenAI models, modes (data types), and application areas >> The number of GenAI patent families has grown from just 733 in 2014 to over 14,000 in 2023, an increase of over 800% since the introduction of the transformer architecture for large language models in 2017. >> The growth in scientific publications has been even more dramatic, increasing from only 116 publications in 2014 to more than 34,000 in 2023. >> Over 25% of all GenAI patents and over 45% of all GenAI scientific publications were published in 2023 alone. 🧠 The remarkable surge in GenAI patents and #publications, especially in recent years, underscores the disruptive potential of these technologies and the intense race among companies and research institutions to secure intellectual property rights and drive innovation in this rapidly evolving domain. 👁️ 5 Key Ideas: 1. GenAI patent families and scientific publications have increased significantly since 2017, driven by advancements in deep learning, availability of large datasets, and improved AI algorithms. 2. Tencent, Ping An Insurance Group, and Baidu are the top patent owners in GenAI, with Chinese organizations dominating the top rankings. 3. China, the United States, and South Korea are the leading locations for GenAI invention based on patent data. 4. Generative Adversarial Networks (GANs), Variational Autoencoders (VAEs), and decoder-based Large Language Models (#LLMs) are the main GenAI models with the most patents. 5. Key application areas for GenAI patents include software, life sciences, document management, business solutions, industry and manufacturing, transportation, security, and telecommunications. 🎯 3 Conclusions: 1. The release of OpenAI's ChatGPT in 2022 has been a pivotal moment for GenAI, driving public enthusiasm and further research and development efforts. 2. While Chinese organizations lead in terms of GenAI #patenting, companies like Alphabet/Google, IBM, and Microsoft are also major players, particularly in scientific publications and impactful research. 3. GenAI is expected to have a significant impact across various industries, enabling applications such as drug development, content creation, customer service, product design, and autonomous driving. Source: https://lnkd.in/eAZy-Pvc

At 456 pages, Stanford’s AI Index Report 2025 is an absolute data powerhouse—an annual benchmark of how AI is reshaping our world. I haven’t read the entire report (who has?), but I dove deep into a few chapters that aligned with my interests—and let me tell you, they are mind-blowing! Here are 5 high-level insights that really stood out: 1 - GenAI is now mainstream. Business adoption surged from 55% in 2023 to 78% in 2024. AI has become a core driver of productivity and growth. 2 - AI is going global—but still unevenly. The U.S. leads in foundational models, but China dominates in publications and patents. Model performance gaps are shrinking fast. 3- Inference costs have plummeted. Running GPT-3.5-class models went from $20 to $0.07 per million tokens in just 18 months—a 280x drop. 4 - AI video and agents are leveling up. OpenAI's SORA, Meta's MovieGen, and Google’s VEO 2 are pushing the limits of multimodal generation. 5 - Responsible AI is evolving—but slowly. New benchmarks exist, but adoption lags. Governments are stepping in to fill the governance gap. I speed-read Research & Development and Technical Performance chapters. Here are the key takeaways from those chapters: Chapter 1: Research & Development 1.1 - AI R&D is industry-led. 90% of notable models came from industry in 2024, but academia still produces the most-cited research. 1.2 - AI models are getting bigger—and greener. Hardware is 40% more energy efficient year-over-year, but training emissions are still skyrocketing (GPT-4: 5,184 tons of CO₂). 1.3 - China leads in volume, the U.S. in impact. China publishes the most AI research, but the U.S. still dominates in influential citations. 1.4 - Inference is getting cheaper. Querying GPT-3.5-level models now costs just $0.07 per million tokens. 1.5 - Patent boom. AI patents grew 30% in one year, with China holding 70% of all global AI patents. Chapter 2: Technical Performance 2.1- Benchmark mastery is accelerating. AI systems improved by 48.9 points on GPQA and 67.3 points on SWE-bench in just one year. 2.2 - Open models are catching up. The performance gap between closed vs. open-weight models dropped from 8% to 1.7%. 2.3 - Complex reasoning is still tough. Even with new methods like "test-time compute," models struggle with logic-heavy tasks. 2.4 - AI video generation is leaping forward. Models like SORA and Veo 2 set new standards for quality and realism. This report is long, dense, and complex—but immensely rewarding. You don’t need to read every page. Pick a section that aligns with your interests. That’s what I did—and I walked away with insights worth sharing. #GenAI #StanfordAIIndex #FutureOfWork #ResponsibleAI #OpenSourceAI #Innovation https://lnkd.in/exP6BRr4

You know how they say "if you do what you've always done, you'll get what you've always gotten"? The opposite is true — and increasingly your secret weapon: uncover insights no one else is paying attention to, and you're at an immediate advantage. We live a world where virtually any type of information, answer, or research is at our fingertips. All that's required is a little time, effort, and ideally knowledge of Boolean or prompt engineering / plug-ins to find those answers. In the words of Seth Godin: "if there’s something I don’t know, it’s almost certainly because I haven’t cared enough to find out." Between your day job, side projects, admin tasks, endless meetings, and fire drills, one source you probably don’t have time (or the desire…) to tap is 𝗮𝗰𝗮𝗱𝗲𝗺𝗶𝗰 𝗽𝗮𝗽𝗲𝗿𝘀. But especially re: how AI might transform the future of strategy & innovation work, academia is an important beacon of what’s to come. Grab my summaries of 4 recent papers on developments that could impact you soon👇 1️⃣ 𝗦𝘁𝗿𝗮𝘁𝗲𝗴𝗶𝗰 𝗱𝗲𝗰𝗶𝘀𝗶𝗼𝗻-𝗺𝗮𝗸𝗶𝗻𝗴 𝗰𝗼𝘂𝗹𝗱 𝘀𝗼𝗼𝗻 𝗴𝗲𝘁 𝗮 𝗯𝗼𝗼𝘀𝘁 𝗳𝗿𝗼𝗺 𝗔𝗜 Researchers recently investigated the potential role of GenAI in evaluating strategic alternatives. 💡 𝗞𝗲𝘆 𝗶𝗻𝘀𝗶𝗴𝗵𝘁: the strategic decision making process may no longer be exclusively human-centric; it will increasingly incorporate AI as a co-contributor, offering insights based on its “enhanced computational capabilities and sophisticated information analysis” that add to human judgment and expertise. 2️⃣ 𝗢𝗵 𝘁𝗵𝗲 𝗽𝗹𝗮𝗰𝗲𝘀 𝘄𝗲 𝗰𝗼𝘂𝗹𝗱 𝗴𝗼 𝘄𝗶𝘁𝗵 𝗮𝗻 “𝗶𝗻𝘀𝗽𝗶𝗿𝗮𝘁𝗶𝗼𝗻 𝗿𝗲𝘁𝗿𝗶𝗲𝘃𝗮𝗹” 𝗲𝗻𝗴𝗶𝗻𝗲 A new model generates new ideas by retrieving “inspirations” from past scientific papers, optimizing for novelty by iteratively comparing idea suggestions to prior papers and updating them until sufficient novelty is achieved. 💡 𝗞𝗲𝘆 𝗶𝗻𝘀𝗶𝗴𝗵𝘁: This represents a step toward evaluating and developing language models that generate new ideas – whether in science, or our orgs. 3️⃣ 𝗜𝘀 𝗔𝗜 “𝗺𝗼𝗿𝗲 𝗵𝘂𝗺𝗮𝗻 𝘁𝗵𝗮𝗻 𝗵𝘂𝗺𝗮𝗻𝘀”? A Turing test of whether AI chatbots are behaviorally similar to humans found that AI and human behavior are remarkably similar. 💡 𝗞𝗲𝘆 𝗶𝗻𝘀𝗶𝗴𝗵𝘁: This makes AI well-suited for roles involving conflict resolution or customer service, where negotiation and dispute resolution are valuable skills. 4️⃣ 𝗙𝗿𝗼𝗺 𝗼𝗻𝗲-𝘁𝗿𝗶𝗰𝗸 𝗽𝗼𝗻𝗶𝗲𝘀 𝘁𝗼 𝗽𝗼𝘄𝗲𝗿𝗳𝘂𝗹 𝗔𝗜-𝗮𝗰𝗰𝗲𝗹𝗲𝗿𝗮𝘁𝗲𝗱 𝗲𝗰𝗼𝘀𝘆𝘀𝘁𝗲𝗺𝘀 The best AI results are increasingly coming from compound systems with multiple components, not just “monolithic models” (one-trick ponies like LLMs predicting the next phrase in a sentence based on what commonly flows together). 💡 𝗞𝗲𝘆 𝗶𝗻𝘀𝗶𝗴𝗵𝘁, 𝗲𝘀𝗽𝗲𝗰𝗶𝗮𝗹𝗹𝘆 𝗳𝗼𝗿 𝘂𝘀 𝗵𝘂𝗺𝗮𝗻𝘀: The various components in these “intelligent” systems will almost certainly include humans, for our unique intellect, if nothing else. #ai #innovation

🤖 The Future of AI is Beyond Language: Introducing "World Models" Top AI researchers like Fei-Fei Li and Yann LeCun are revolutionizing artificial intelligence by moving beyond traditional language models. Here's what makes their approach groundbreaking: 🌐 World Models: Not just processing words, but understanding spatial intelligence 📐 3D Reasoning: Training AI to comprehend and interact with complex environments 🧠 Mental Constructs: Mimicking how humans actually perceive and predict the world Key Insights: - Language is limited - the world is fundamentally three-dimensional - AI needs to understand context, not just statistical word relationships - Spatial intelligence is the next frontier of machine learning Li's World Labs has already raised $230M to develop these advanced models, focusing on creating AI that can: - Generate infinite virtual worlds - Enhance robotics - Improve perception in complex scenarios The challenge? Gathering sophisticated spatial data is incredibly complex. But the potential is transformative. What do you think? Are we witnessing the next quantum leap in AI technology? #ArtificialIntelligence #FutureOfTech #WorldModels #AIInnovation

In the Spotlight: How does AI transform the earliest stage of innovation – ideation? That is the guiding question explored in a new article by Christian Pescher and Gerard Tellis, recently published in the Journal of Product Innovation Management (JPIM). Their paper, titled “The Role of Artificial Intelligence in the Ideation Process,” offers a comprehensive review of how AI reshapes the front end of innovation – from identifying opportunities to generating and evaluating ideas. 🔍 Key takeaways: 1. Firm culture will become an even more critical driver of radical (vs. incremental) innovation in the age of AI. As AI increasingly mediates this relationship, managers should actively foster a culture that supports innovation. 2. AI enhances the speed, efficiency, and cost-effectiveness of ideation. Managers should leverage AI tools to accelerate and scale the ideation process – and stay adaptive as the technology evolves rapidly. 3. AI improves the average creativity of generated ideas, but research is conflicting on whether it enhances the creativity of top ideas. Until conclusive evidence shows otherwise, managers should continue to invest in exceptional human talent. 4. AI performs well in idea screening but still falls short in idea selection. To avoid overlooking high-potential concepts, firms should combine AI-driven insights with human judgment. 🚀 Although research on AI in ideation is still in its early stages, a clear and fast-growing research agenda is taking shape – signaling a transformative shift ahead. cc: Minu Kumar, Gerda Gemser, Ruby Lee, Luigi M. De Luca