AI Accelerates Math & Science Discovery with Gemini Deep Think

Did AI just gave us a "Cheat Code" for the Physical World? While researching for something I stumbled upon, GitHub repository of GNoME project by Google Deepmind (https://lnkd.in/dmMtSvSU). For decades, discovering a single new material—like the stuff inside your phone battery or a solar panel—was a game of "scientific hide and seek." It took years of trial, error, and luck. Google DeepMind just changed the rules of the game with GNoME. They didn't just find one new material. They found 2.2 million. Here is the "4W + 1H" breakdown of why this is the most important AI story you might have missed: 🌍 WHO: Google DeepMind (the AI architects) and the A-Lab at Berkeley (the robotic builders). 🧪 WHAT: An AI model that predicted 381,000 stable new crystal structures. The Scale: Humanity had only found 48,000 in all of history. Google just 10x-ed our "periodic table" of possibilities overnight. 📍 WHERE: In the "Digital Twin" of physics. Instead of mixing chemicals in a jar, they simulated the laws of nature inside a supercomputer. 🎯 WHY: Because "Software is eating the world," but Hardware runs the software. To get to the next level of EV batteries, superconductors, and carbon capture, we need materials that don't exist yet. ⚙️ HOW (The Scaling Secret): They used Graph Neural Networks (GNNs). Think of it like a "Data Flywheel": AI predicts a material structure. A physics simulation checks if it’s stable. The result goes back into the AI to make it smarter. 💡 What’s In It For You? (The WIIFY) You don't need to be a Data Scientist to feel the impact of this: 800 Years Saved: DeepMind estimates this leap would have taken 8 centuries of human lab work. We just fast-forwarded to the future. The End of "Battery Anxiety": Among the discoveries are 528 potential new lithium-ion conductors. This is the foundation for batteries that charge in seconds and last for days. From Scarcity to Abundance: We are moving from a world where we "find" materials to a world where we "design" them for specific problems. As someone who works at the intersection of Product and Machine Learning, I find this fascinating because it proves that the true power of Scaling Deep Learning isn't just in generating text or images, it’s in solving the bottlenecks of the physical world. The "Lab of the Future" is no longer just tubes and beakers; it’s code and robots. references : Official Nature Article: https://lnkd.in/dQAthBbs Google Deepmind Blog : https://lnkd.in/dgfCPP94 #AI #DeepLearning #GNoME #Innovation #ProductManagement #MaterialsScience #TechTrends #GoogleDeepMind

When AI beat Kasparov, we learned machines could calculate. When AlphaGo beat Lee Sedol, we learned machines could surprise us. Now mathematicians are asking a sharper question: Can AI produce a proof no human has ever seen? This week, Scientific American covered a new initiative called “First Proof.” (see comments for link) Ten unpublished research lemmas. Solutions encrypted. No training leakage. No recycled competition tricks. This isn’t a benchmark. It’s a boundary test. But here’s the part that fascinates me. When computers first entered mathematics, they didn’t replace mathematicians. They expanded the adjacent possible. • Computational experimentation helped open chaos theory into a serious field. • Complexity theory matured because we could simulate limits. • Entire branches of discrete math, cryptography, and algorithmic number theory accelerated because machines made exploration cheap. The computer didn’t prove most of those theorems. It changed what was thinkable. Now we’re at another inflection. For many mathematicians, the proof is the proof. Rigor. Formalism. Closure. But historically speaking, formal proof as the dominant lens of mathematics is relatively recent. For most of human history, math was geometric intuition, mechanical construction, pattern play. The Bourbaki-style formal edifice is a blink in civilizational time. So AI doing proofs? Yes, that’s interesting. But I’m more interested in something else: Can AI handle the combinatorial grind of proof so humans can return to imagination? Can it hold 200 symbolic branches in working memory while we explore new abstractions? Can it stress-test conjectures at machine scale while we design new conceptual frames? That’s a Centaurian question and what my cofounder, Scott Wolfson and I are most interested in. The real opportunity isn’t: “Did AI pass the exam?” It’s: “What happens to mathematics when proof becomes augmented infrastructure?” History suggests a pattern: New tools don’t shrink human genius. They relocate it. If AI can assist with proof, we may see: • Faster conjecture cycles • New hybrid fields (symbolic + neural mathematics) • A renaissance of mathematical imagination Not machine mathematicians. Augmented mathematicians. And if the “First Proof” challenge exposes limits? Good. Boundaries clarify where collaboration matters. Either way, the frontier moves. The future of math may not be AI proving theorems. It may be humans dreaming bigger ones. #CentaurianAI #AugmentedIntelligence #Mathematics #FutureOfWork #AdjacentPossible

To Achieve Human-Level AI, Computer Science Should Be Enhanced By Knowledge Computing Science Artificial intelligence has been driven by computer science since its advent. As computer science is fundamentally data science, we are always getting data-driven artificial intelligence, such as machine intelligence and generative artificial intelligence. These forms of AI are based on a black box development model and therefore incapable of providing explanations. Also, the study of artificial intelligence has lasted decades, while little has been done on the study of intelligence itself. This, however, contradicts how computer science is applied in other fields, such as business applications. In these fields, first, the business process is studied, then business decision making is examined. After these vital steps, computer science takes over to develop models, databases, algorithms, and programs. What should be the role of business processes in studying human-level AI? Obviously, it is human intelligence. The comprehensive development of human-level-based AI therefore requires first a theory of human intelligence associated with different intelligence modalities like language intelligence, visual intelligence, semantic reasoning intelligence, etc. Development of these modalities should follow the general concept of intelligence, which is defined by the Oxford Dictionary as the ability to acquire and apply knowledge and skills. Since knowledge is fundamental to human intelligence, a theory of knowledge must be developed next. For each intelligence modality, it must contain knowledge structures for representing entities and the relationships between them in the real world, as well as knowledge processing techniques for developing models for acquiring and applying knowledge. Finally, knowledge mathematics should be developed and applied to create knowledge computing science. After that computer science will become a data- and knowledge-driven science as well as an engine for producing human-level artificial intelligence. Creating knowledge computing science and its application to development of human-level artificial intelligence are studied in my paper https://lnkd.in/ecbdJm2Z. The paper introduces a framework for researching and developing knowledge computing AI (KCAI). Foundations of KCAI are: semantic knowledge calculus, axiomatic computational relation ontologies representing the meaning of relations between entities in semantic knowledge graphs; algebraic models of semantic reasoning in semantic triple chains; algebraic models of intelligence layers in semantic knowledge graphs providing completion, compression, and question answering. The paper describes a prototype for a knowledge computing AI system based on a white box development model that implements the functionalities described above and provides self-explanatory decision-making.

As we have entered the second half of this decade (2026-2030), I thought maybe it is time to summarize what we have done in the first half of this decade and highlight our goals for the next few years. In the first half of this decade (2021-2025) we did lots of good work.  We introduced the Golden Era of Mathematics and highlighted the foundations of Modern Mathematics [1]. We also wrote about the Rise of Mathematical Spaces [2] and using LLMs and GenAI for Exploring Mathematical Spaces [3]! For this second half of the decade, we are going to move on the same direction, by focusing on a few specific important goals: 1- To Explore Mathematical spaces using Generative AI and LLMs. 2- To define, formalize, and tackle important real-world problems using mathematical techniques (including using LLMs and GenAI tools to solve open problems in AI) Regarding the first goal, we may consider studying mathematical spaces more in depth, moving to new mathematical spaces, or trying to explore spaces that have not been explored that much. Some rich mathematical spaces include vector spaces and polynomial spaces! Among the new mathematical spaces are perfectoid spaces and Ruliad space [2]. These spaces might potentially be rich in terms of mathematical properties and possibly full of unsolved problems. It would be an interesting line of research to use modern AI technologies and LLMs for exploring these mathematical spaces more in depth. Interested readers might also consider: "A Long March through Galois Theory" by Grothendic! It is indeed a long march (1600 pages) in Mathematical Spaces (i.e., moduli spaces)! With regard to the second goal, we would like to focus on solving the real-world problems, particularly in mathematics and AI with applications in other domains such as robotics, optimization, trustworthy AI, software automation, etc. Some of the important open problems in AI: inaccuracy (e.g., hallucination), environmental impacts (high energy consumption), trust, safety and security, data privacy, value alignment, reasoning, memory (memorization) and long context windows in LLMs, optimization, etc. Mathematical techniques can potentially be used for addressing such open problems. The Golden Era of Mathematics: From Computer Science to Data Science [1] https://lnkd.in/etR2uvjC On the Rise of new Mathematical Spaces and towards AI-Driven Scientific Discovery [2] https://lnkd.in/eGZgy8xZ Exploring Mathematical Spaces using Generative AI and Large Language Models [3] https://lnkd.in/ekCeHddE #AI #LLMs #LLM #OpenProblemsInAI #Math #AI4Math #Math4AI, #LLMs4Math #Math4LLMs #MathematicalReasoning #Reasoning #GoldenEraOfMathematics #GoldenMath #Reasoning #Planning #WorldModels #Lean #FormalLanguage4Math #VectorSpace #MathematicalSpaces #SoftwareAutomation #TrustworthyAI #MathematicalChallengesInAI

A good overview of what a "stack" looks like taking frontier research. I’ve condensed the research into a 14-paper must read curriculum . If you want to understand where 2026 is going, read these in order. Part I: The Silicon Brain (Architecture & Efficiency) How we are rewriting the Transformer for scale and structure. 1. Attention is All You Need (2017): The prerequisite. Read it again. 2. Deepseek mHc (Manifold-Constrained Hyper-Connections): The evolution of the residual stream. How to make connections denser and less noisy. 3. The Free Transformer (Meta): Stripping away attention mechanisms to solve the compute bottleneck. 4. Recursive Language Models (MIT): Moving beyond "next token" prediction to hierarchical sentence planning. 5. Representations Learned In-Context (Deepmind): The mechanics of how models learn without weight updates. Part II: The Reasoning Engine Moving from pattern matching to logical proofs. 6. Chain of Thought (2023): The prompting strategy that unlocked reasoning. 7. Deepseek Math V2: The stress test. Showing the upper limits of current reasoning on hard math problems. Part III: The AI Scientist (Agents) The shift from Chatbot to Lab Partner. 8. Towards an AI-Coscientist (Deepmind):The blueprint for autonomous scientific discovery agents. 9. What Takes to be a Good Research Agent (Meta):Defining the "personality" traits (curiosity, diligence) required for science. 10. Human-AI Complementarity (Deepmind): Why AI won't replace scientists, but fill the "cognitive gaps" we can't handle. Part IV: The Biological Mirror (The "Levin" Pivot) Why Silicon needs to learn from Cells to solve fragility. 11. Bio-Inspired AI (Levin): Why we should copy the body (robustness), not just the brain (neural nets). 12. Cognition All the Way Down 2.0: Intelligence is a property of life, not just brains. 13. Ingressing Minds: How intelligence enters a substrate whether it's a biological embryo or a silicon chip. 14. Evolution by Natural Induction I & II: A proposal for a new evolutionary mechanism to explain rapid adaptation. #AI #AIResearch #CognitiveStack

A DeepSeek "memory moment" for AI architecture? DeepSeek just dropped research that questions something fundamental about how LLMs work. The problem: Transformers don't actually have memory. Every time you type "Diana, Princess of Wales", the model wastes compute reconstructing this common phrase from scratch. It's like a mathematician recounting 1-10 before solving every equation. Language modeling has two very different jobs: → Compositional reasoning (connecting ideas, logic) → Knowledge retrieval (recognizing static patterns) Current Transformers have no native lookup mechanism. They're forced to simulate memory through expensive neural operations. DeepSeek's solution is called Engram. A "conditional memory" module using modernized N-gram embeddings that retrieves static patterns in O(1) constant time. The key insight: there's an optimal balance between compute (MoE experts) and memory (Engram). They discovered a U-shaped scaling law. Pure MoE? Suboptimal. Reallocating 20-25% of sparse parameters to memory yields the best results. The benchmark gains surprised even the researchers. → Knowledge tasks improved as expected (MMLU +3.4, CMMLU +4.0) → But reasoning improved MORE (BBH +5.0, ARC-Challenge +3.7) → Code and math jumped too (HumanEval +3.0, MATH +2.4) → Long-context retrieval: 84.2% → 97.0% on multi-query NIAH Why does memory help reasoning? Mechanistic analysis shows: by offloading pattern reconstruction, Engram effectively deepens the network. Shallow Engram layers produce representations equivalent to much deeper MoE layers. The infrastructure implications are massive. Deterministic addressing allows offloading 100B-parameter tables to host CPU memory with less than 3% throughput overhead. This decouples storage from expensive GPU HBM entirely. The takeaway: the next generation of sparse models might not just be about smarter routing. Some knowledge simply doesn't need to be computed at all. Paper co-authored by DeepSeek founder Wenfeng Liang. Paper: https://lnkd.in/gRwsCapf Code: https://lnkd.in/gPhhxeP9 ↓ 𝐖𝐚𝐧𝐭 𝐦𝐨𝐫𝐞 𝐀𝐈 𝐮𝐩𝐝𝐚𝐭𝐞𝐬 𝐥𝐢𝐤𝐞 𝐭𝐡𝐢𝐬? Follow Pavan Kumar Dubasi for daily AI research breakdowns. ♻️ Repost to help your network stay ahead. 💾 Save for reference. 🔔 Turn on notifications so you don't miss the next one. #DeepSeek #AI #MachineLearning #LLM #AIResearch #Engram #TransformerArchitecture

𝑾𝒉𝒚 𝒅𝒐𝒆𝒔 𝒂𝒔𝒌𝒊𝒏𝒈 𝒂𝒏 𝑳𝑳𝑴 𝒕𝒐 "𝒕𝒉𝒊𝒏𝒌 𝒔𝒕𝒆𝒑-𝒃𝒚-𝒔𝒕𝒆𝒑" 𝒂𝒄𝒕𝒖𝒂𝒍𝒍𝒚 𝒊𝒎𝒑𝒓𝒐𝒗𝒆 𝒎𝒂𝒕𝒉𝒆𝒎𝒂𝒕𝒊𝒄𝒂𝒍 𝒂𝒄𝒄𝒖𝒓𝒂𝒄𝒚? Asking an LLM to "think step-by-step" to a prompt miraculously boosts mathematical accuracy, not by chance, but through fundamental changes in neural network computation. Welcome to another edition of #BehindTheBasics (Issue 8) ! Training a larger model is not always necessary to solve harder problems; instead, intelligence can be dynamically scaled during inference. Large Language Models are autoregressive, utilizing a single "forward pass" through internal layers to generate one token at a time. Demanding an immediate answer to a complex mathematical operation forces a massive computational leap within a fixed, limited computational depth, often causing the model to hallucinate or fail Standard prompting seeks the direct probability of the final answer given only the initial question. Mathematically, the standard prediction requires computing P(y | x) (Given a question ‘x’, calculate the probability of the answer ‘y’). Because the gap between question and answer remains vast, the resulting probability distribution across all possible next tokens appears flat and noisy. The network essentially guesses in the dark. Chain of Thought alters the underlying mathematics by breaking the problem into intermediate reasoning steps (z₁, z₂,…., zₙ). The model now computes a sequence of conditional probabilities: P(z₁, z₂,…., zₙ, y | x) = P(z₁ | x).P(z₂ | x)….P(y | x, z₁, z₂,…., zₙ) Generating intermediate tokens forces a new forward pass for every single word. Outputting n-tokens of intermediate thought secures n-sequential forward passes, linearly expanding compute time during inference. Attention mechanisms process the raw text of the intermediate steps. Chain of Thought serializes multidimensional problems into linear, high-probability predictions. The approach builds a stable bridge across the context window toward the correct final token. Generating multiple distinct reasoning paths multiplies test-time compute. Sampling these multiple paths marginalizes out noise, ensuring robust mathematical accuracy by trading test-time compute. Applying a majority vote extracts the final, definitive answer. The brute-force method approximates the marginal probability of the answer over all possible reasoning paths: Standard LLMs attempt to leap across a wide river in a single bound often results in failure. Chain of Thought method places stepping stones creates a bridge of short, manageable strides, reaching the opposite bank safely. #BehindTheBasics #MachineLearning #DataScience #LargeLanguageModels #ChainOfThought #DeepLearning 

I built "CrystalDiff", an end-to-end 3D diffusion model that generates stable Perovskite crystals out of pure noise. Instead of importing high-level graph wrappers, I wanted to truly understand the math. So, I implemented the continuous-time diffusion loop and an E(n)-Equivariant Graph Neural Network entirely from scratch in PyTorch. This ensures the neural net respects physical rules like 3D rotational symmetry. By analyzing the Radial Distribution Function (RDF) of the results, the model actually learned the Pauli Exclusion Principle and correct covalent bond lengths on its own! I wrote a full breakdown of the math, the code, the app and my journey learning this below: 📖 Read the blog: https://lnkd.in/dXy4YVHS 💻 GitHub Repo: https://lnkd.in/dH8wTGBF 🌍 App: https://lnkd.in/dvE2aqaG Towards AI, Inc. #MachineLearning #GenerativeAI #MaterialScience #DeepLearning

How a 19th-century mathematical curve is shaping modern AI image generation Insights from “HilbertA: Hilbert Attention for Image Generation with Diffusion Models” Most machine learning models, particularly transformers and diffusion models, operate on sequences, while images are inherently spatial. When an image is flattened row by row, spatially neighboring pixels can become distant in the sequence, weakening the model’s ability to capture local structure. This mismatch is subtle but critical for attention mechanisms, which depend on meaningful neighborhood relationships. The Hilbert curve offers a simple yet powerful solution. By traversing an image along a Hilbert path, a two-dimensional image can be mapped into a one-dimensional sequence while largely preserving spatial locality. Pixels that are neighbors in the image remain neighbors in the sequence, allowing the model to respect image geometry without explicitly relying on convolutions. This idea is explored in “HilbertA: Hilbert Attention for Image Generation with Diffusion Models.” Instead of attending over a naïvely flattened image, the model orders pixels using a Hilbert curve and applies attention along this locality-preserving sequence. The result is improved spatial coherence and structural consistency in generated images, achieved through a smarter mathematical representation rather than a more complex model. More broadly, this highlights a recurring pattern in AI: many advances are not entirely new inventions, but rediscoveries of deep mathematical ideas placed in a modern computational context. AI progress is not only about scaling models or data, but about recognizing that old mathematics often already contains the structures we need. Source of information: https://lnkd.in/g_x7ZiJu Source of animation: https://lnkd.in/gbRxv3UB

AI agents can now do AI research. But they’re doing it in ways we didn’t expect. New benchmark (AIRS-BENCH) from Meta, Oxford, and UCL tested whether AI agents can independently complete real ML research tasks, like coding, time series forecasting, protein modeling, and text classification. The answer? Yes. But not like humans do. When humans solve a text classification problem, they fine-tune one model on the training set and test it. Clean. Simple. Elegant. However when GPT-OSS-120B solved the same problem, It built a two-level stacked ensemble combining multiple transformer models with a meta-learner, ran 5-fold stratified cross-validation, and concatenated logits from RoBERTa-large AND DeBERTa-v3-large into feature vectors. It over-engineered the solution entirely. Surprisingly, only mid-tier models have been tested. No Claude Opus. No GPT-4.5. No Gemini Ultra. The best models tested still could not match elite human researchers. Consider what that implies for where we are headed. I believe we are about to watch AI agents learn to make research reasoning shorter. More powerful models will not just perform better. They will find simpler solutions. That may be the real scaling law no one is discussing. Most companies are waiting for AI to become “ready.” Meanwhile, their competitors are deploying intelligent automation right now. Not in 6 months. Not after a lengthy consulting engagement. In 5 to 10 days. My approach is built on a principle I call the Automation Preparation Paradox. You cannot automate chaos, only order. So before I deploy a single agent, I build the data foundation and process architecture that allows AI to perform like a senior operator, not a junior engineer brute-forcing its way to an answer. The difference between an AI agent that over-engineers a stacked ensemble and one that finds the elegant solution? It is not just model size. It is the quality of the system around it. The data. The workflows. The guardrails. That is where the real leverage lives. If mid-tier models can independently tackle ML research tasks today, imagine what properly architected AI agents can do for your revenue pipeline, your lead qualification, or your operational workflows tomorrow. The gap between “AI as a tool” and “AI as a strategic advantage” is closing fast. The question is whether you are building the foundation now or scrambling to catch up later. Read More: https://lnkd.in/eiSuzWNz #ArtificialIntelligence #AIAgents #MachineLearning #Automation #AIResearch #FutureOfWork #GenerativeAI #BusinessAutomation #AIStrategy #DeepLearning #TechLeadership #DigitalTransformation #AIBenchmark #LLMs #Innovation