GenAItechLab.com’s Post

New Book: Building Disruptive AI & LLM Technology from Scratch https://lnkd.in/eT4wrhA4 This book features new advances in game-changing AI and LLM technologies built by GenAItechLab.com. Written in simple English, it is best suited for engineers, developers, data scientists, analysts, consultants and anyone with an analytic background interested in starting a career in AI. The emphasis is on scalable enterprise solutions, easy to implement, yet outperforming vendors both in term of speed and quality, by several orders of magnitude. Each topic comes with GitHub links, full Python code, datasets, illustrations, and real-life case studies, including from Fortune 100 company. Some of the material is presented as enterprise projects with solution, to help you build robust applications and boost your career. You don’t need expensive GPU and cloud bandwidth to implement them: a standard laptop works. Part 1: Hallucination-Free LLM with Real-Time Fine-Tuning Part 2: Outperforming Neural Nets and Classic AI Part 3: Innovations in Statistical AI About the author Vincent Granville is a pioneering GenAI scientist and machine learning expert, co-founder of Data Science Central (acquired by a publicly traded company in 2020), Chief AI Scientist at ML Techniques and GenAI Techlab, former VC-funded executive, author (Elsevier) and patent owner — one related to LLM. Vincent’s past corporate experience includes Visa, Wells Fargo, eBay, NBC, Microsoft, and CNET. ➡️ See content and get your copy, at https://lnkd.in/eT4wrhA4

I am really excited about the o3 model launch - particularly because the model does amazing in the ARC benchmark, which if you have heard Francois Chollet speak before, is notoriously difficult (gpt-4 achieving 5%, whereas o3 in low-compute mode achieved 75%). While the specifics of the o3 model architecture isn’t public, this generation of models essentially has a secondary level of abstraction where an evaluator model evaluates and combines a set of “ways of solving a problem” (based on training on expert data) when tackling a novel problem. If you consider the gpt-4 gen models as system 1 thinking, searching over “ways of solving problems” and combining them in novel ways is more akin to system 2. Here’s a snippet from Francois’ commentary on o3 doing system 2 thinking: “o3 represents a form of deep learning-guided program search. The model does test-time search over a space of "programs" (in this case, natural language programs – the space of CoTs that describe the steps to solve the task at hand), guided by a deep learning prior (the base LLM).” Research is a very nonlinear trajectory, and it’s great to see a new inflection point on that curve to intelligence emerge. Very exciting! Read more on the ARC benchmark accomplishment. https://lnkd.in/gdKkqYB7

OpenAI's "Strawberry" Project: Advancing AI Reasoning and Research Capabilities OpenAI is reportedly developing a new project codenamed "Strawberry" to enhance the reasoning abilities of AI models. This project, previously known as Q\* or Q-Star, focuses on advanced reasoning technology similar to Stanford's "Self-Taught Reasoner" (STaR) method. With the aim of enabling AI models to perform autonomous web searches and conduct "deep research," Strawberry is expected to bring about a new generation of AI systems capable of complex planning and execution. Internally, OpenAI has tested this new model that scored over 90 percent on the MATH benchmark, a collection of high-level math problems. This performance surpasses previous models like GPT-4 and GPT-4o, indicating significant advancements in mathematical and reasoning skills. The MATH benchmark, utilized to measure AI performance in solving complex mathematical problems typically found in high school and college competitions, serves as a testament to the AI's mathematical prowess. The Strawberry project involves a special form of "post-training," adapting pre-trained models for specific tasks using a "deep research" dataset. This approach is part of OpenAI's broader vision to create AI agents that can reason logically before taking action, heralding a significant leap in machine understanding. The development of Strawberry, alongside projects like Quiet-STaR, aims to equip the next generation of AI systems with enhanced understanding and reasoning capabilities, potentially revolutionizing fields such as software engineering and machine learning. Microsoft CTO Kevin Scott has also highlighted the potential of next-generation AI models, suggesting they could achieve significant advances in reasoning. # Thank you Maia Silva for your submission!

OpenAI’s newly announced o3 model has sparked excitement about AI’s future despite concerns that AI progress was slowing. Achieving 75.7% on the ARC benchmark (87.5% with high compute)—far surpassing prior results—the model demonstrates significant advances in reasoning and adaptability. Key innovations in o3 include: Program Synthesis - Combines learned patterns and methods to tackle novel tasks. Natural Language Program Search - Uses Chains of Thought (CoTs) to explore multiple solutions and select the best fit. Evaluator Model - Judges its own reasoning steps to refine answers. Executing CoTs - Reuses logical frameworks to solve problems, achieving “Grandmaster” programming performance. Deep Learning-Guided Search - Tests and refines solutions during inference but raises scalability concerns. Challenges: High computational costs and reliance on reinforcement learning raise doubts about economic feasibility and scalability. Enterprise Impact: OpenAI plans a scaled-down “o3-mini” version by January 2025 for businesses, balancing innovation with affordability. Companies are advised to test o3-mini while leveraging existing AI models like OpenAI’s o4 for practical applications. The AI community remains divided—some praise o3’s technical strides, while others question its long-term viability. Nonetheless, AI development continues to reshape industries, promising further advancements in 2025 and beyond.

📖 Generate financial industry-specific insights using generative AI and in-context fine-tuning In this blog post, we demonstrate prompt engineering techniques to generate accurate and relevant analysis of tabular data using industry-specific language. This is done by providing large language models (LLMs) in-context sample data with features and labels in the prompt. The results are similar to fine-tuning LLMs without the complexities of fine-tuning models.

Researchers from Snowflake and CMU Introduce SuffixDecoding: A Novel Model-Free Approach to Accelerating Large Language Model (LLM) Inference through Speculative Decoding Researchers from Snowflake AI Research and Carnegie Mellon University introduce SuffixDecoding, a robust model-free approach that avoids the need for draft models or additional decoding heads. Instead of relying on separate models, SuffixDecoding uitlizes efficient suffix tree indices built upon previous output generations and the current ongoing inference request. The process begins by tokenizing each prompt-response pair using the LLM’s vocabulary, extracting all possible suffixes (subsequences from any position to the end) to construct the suffix tree structure. Each node in the tree represents a token, and the path from the root to any node corresponds to a subsequence that appeared in the training data. This model-free approach eliminates the complications and GPU overhead associated with integrating draft models or additional decoding heads, presenting a more efficient alternative for accelerating LLM inference. For each new inference request, SuffixDecoding constructs a separate per-request suffix tree from the current prompt tokens. This design is crucial for tasks where the LLM output is expected to reference or reuse content from the input prompt, such as document summarization, question-answering, multi-turn chat conversations, and code editing. The suffix tree maintains frequency counts at each node to track how often different token sequences occur, enabling efficient pattern matching. Given any sequence of recent tokens from the current generation, SuffixDecoding can quickly traverse the tree to find all possible continuations that appeared in the prompt or previous outputs. At each inference step, SuffixDecoding selects the best subtree(s) of continuation tokens based on frequency statistics and empirical probability. These speculated tokens are then passed to the LLM for verification, which is carried out in a single forward pass thanks to a tree attention operator with a topology-aware causal mask.... Read the full article here: https://lnkd.in/g-BdegpE Paper: https://lnkd.in/gUEvFJ4F Snowflake Snowflake Developers Machine Learning Department at CMU Gabriele Oliaro Zhihao Jia Daniel Campos Aurick Qiao

🔍 Exploring the top 12 Artificial Intelligence tools & frameworks was a mind-boggling journey, revealing the vast horizons AI has to offer. Dive into the key highlights in this article! 🌐 🧠 Imagine AI tools as architects of neural networks, taking on the heavy-duty tasks while we focus on the big picture decisions. It's like painting a masterpiece without the hassle of mixing colors one by one. 📊 Numerical data and findings in AI tools showcase the monumental impact of innovating with Scikit Learn, TensorFlow, Keras, PyTorch, and more. Numbers, more than words, paint the real picture of AI's significance. 🤖 Auto ML emerges as a game-changer, ushering in a new era where software builds software. It's like having an assistant fine-tune your masterpiece with precision, saving time and effort. 🚀 The potential of OpenNN and H20: Open Source AI Platform lies in their ability to elevate data analysis and decision-making to a whole new level, akin to giving wings to innovation. ⚡️ Google ML Kit sparks imagination, enabling developers to weave personalized features with machine learning magic, revolutionizing app experiences on Android and iOS. Stay tuned for more insights and delve deeper into the realm of Artificial Intelligence tools & frameworks! 🔍🧠🚀 #AI #MachineLearning #Innovation https://lnkd.in/eNZcKFCP

Interested in information extraction with generative AI? Check out this article where I explore an application of LangChain with OpenAI APIs to read from images of receipts! https://lnkd.in/evb4Z9qx

🚀 When Plateau is Actually a Fork Recently, the AI world buzzed about scaling laws. OpenAI’s Orion showed modest improvements over GPT-4, and rumors suggest Google’s Gemini fell short. This sparked talk of an AI plateau. 🧠 Marc Andreessen noted that models are “hitting the same ceiling,” while Ilya Sutskever reflected that “the 2010s were the age of scaling; now we’re back in the age of wonder and discovery.” But is this a plateau—or something else? In science, a plateau means a steady state. With generative AI, we’re far from steady. It’s a fork in the road, where new strategies can break through. 🔗 Compound AI: A New Approach Instead of chasing bigger models, Compound AI combines multiple models, tools, and systems to solve tasks more efficiently. 📜 Evolution of Compound AI: • 1990s: 🌳 Ensemble learning (e.g., Random Forests). • 2010s: 🛠️ Pipeline systems like IBM Watson. • 2020s: 🤖 Tool-integrated models like Codex & AlphaCode. 💡 BAIR’s paper, “The Shift from Models to Compound AI Systems” (Feb 2024), framed this as the future. Models like F1 now excel in coding and reasoning by leveraging these principles. https://lnkd.in/d2qxy4n8. 📈 Scaling Smarter, Not Bigger Test-time compute is reshaping scaling. ⚡ OpenAI’s o1 showed that letting models “think longer” during inference improves performance without massive resource increases. 🔍 Must-Read Papers: 1️⃣ “Scaling LLM Test-Time Compute Optimally” – Google DeepMind. https://lnkd.in/dqy9UZtC 2️⃣ “Training Verifiers to Solve Math Problems” – OpenAI. https://lnkd.in/dZqjNYaX 🔧 Reasoning Meets Action We’re moving from scaling reasoning to integration—where models act using external tools and workflows. ✨ Example: “The Dawn of GUI Agent” paper tested Claude 3.5 on 20 real-world desktop tasks, combining reasoning with dynamic action. https://lnkd.in/dWJf2c2S 🛤️ The Real Transition We’re not at a plateau—we’re transitioning: 1️⃣ Scaling smarter: Resource-efficient test-time compute. 2️⃣ Integration: Building systems that reason and act. 💬 What’s your take? Are we at a plateau—or a breakthrough moment? Let’s discuss!

New OpenAI’s o3 model represents a significant leap in adaptability and problem-solving. Following François Chollet article, with the ARC-AGI-1 Public Training set, o3 scores 75.7% on the Semi-Private Eval at a high-efficiency setting costing around $2,012 for 33 million tokens and about $20 per task. At a low-efficiency setting with far more compute, o3 hits 87.5%. On the Public Eval set, o3 achieves 82.8% at a high-efficiency setting that costs $6,677, and a steep 91.5% at low-efficiency. Compared to previous GPT-based models that barely moved from 0% to 5% over years, this improvement feels like a sudden jump in capability. The uniqueness of o3 lies in its ability to adapt to entirely new tasks by searching and executing “programs” represented as Chains of Thought, rather than just retrieving memorized content. It relies on massive token exploration and backtracking through huge solution spaces, showing that architecture and the method of combining known elements into new programs are what truly drive progress. Although the current cost remains high compared to human task-solving rates, these costs will likely decrease over time, making test-time program synthesis more feasible and eventually competitive with human effort. Future benchmarks such as ARC-AGI-2 are expected to challenge o3 and reveal its current limitations. Even at high compute settings, performance might drop below 30% while a human can easily score over 95%. The promise of tougher evaluations pushes the field forward, helping researchers refine system architectures and reduce reliance on brute-force compute. The breakthroughs achieved with o3 suggest a move toward more general, flexible, and human-like intelligence. More info here: https://lnkd.in/dczqc9fp #HumanCrafted #AIAssisted #AGI #o3 #AIResearch #Benchmarking #IntelligentSystems #NLP #LLM