How AI powers Web3, DeFi, and crypto investment at 2crypto.io

Excited to share our latest work!🎉 We propose an efficient test-time scaling framework for small vision-language models via Test-Time Augmentation and Test-Time Adaptation. 🌐 Project Page: https://lnkd.in/dwrHQ5i5 📄 Paper: https://lnkd.in/dXa-f-3y 💻 Code: https://lnkd.in/dD7ZA-Wa ✨ Test-Time Augmentation: token-level aggregation of augmented inputs ⚡ Test-Time Adaptation: lightweight adaptation via pseudolabels ✅ Consistent gains on 9 benchmarks ✅ Runs on consumer GPUs ➡️ Resource-efficient, practical, and generalizable Efficient Test-Time Scaling for Small Vision-Language Models #TestTimeScaling #VLM #Multimodal #LLM #DeepLearning #AI

I trained a 2.9M-parameter AI to solve complex mazes using recursive reasoning on a MacBook After watching Matthew Berman's excellent video on the subject, I was inspired to have a go at creating my first machine learning model. What I Built - A Tiny Reasoning Model (TRM) I implemented the TRM architecture (based on Samsung SAIT Montreal's research) and adapted it specifically for maze navigation. Here's what makes this unique: 2.9 million parameters — smaller than most mobile apps A Single 2-layer neural network that gets called 18 times per prediction Each recursion refines the previous answer, like repeatedly asking "can I make this path better?" Runs entirely on Apple Silicon using native MPS acceleration - Solves randomly generated 30×30 mazes - Input: 3-channel spatial encoding (walls, start position, goal position) - Output: Binary path prediction for all 900 cells - Ground truth from A* algorithm The Results: ✅ 95%+ success on EASY mazes (achieved quickly) ✅ 90%+ on MEDIUM difficulty ✅ Progressively competitive on harder levels ✅ Model learns recursive refinement patterns ✅ Total training time: Minutes, not days Question for the AI community: If a 2.9M parameter model can learn complex spatial reasoning through recursion, what does that mean for the "bigger is always better" scaling paradigm? Are we investing too heavily in model size (and GPUs) when architecture innovation and training methodology might yield better returns? I'd love to hear your thoughts, especially from folks working on parameter-efficient models, curriculum learning, or spatial reasoning tasks. Matthew Berman's video https://lnkd.in/e5xM4i5p https://lnkd.in/eJaJB5JC Original author of this amazing work: Alexia Jolicoeur-Martineau Samsung SAIL Montréal #MachineLearning  #AI  #DeepLearning  #PyTorch  #RecursiveReasoning  #SpatialIntelligence #ParameterEfficiency  #CurriculumLearning #AppleSilicon  #AIResearch  #NeuralNetworks  #ComputerVision #PathfindingAI

I spent some time reading the Thinking Machines Lab post that clearly outlines how LoRA in reinforcement learning can match full fine-tuning performance all while using just 2/3 of the computational resources. Blog and paper: https://lnkd.in/ehsCi8UE This fundamentally changes the accessibility of advanced AI development. The misconception we've been living with: The prevailing wisdom said you needed lots of GPUs to achieve competitive thinking models through full fine-tuning. Turns out, with properly configured LoRA, you can achieve the same results on a single GPU. The belief that "LoRA is worse" wasn't about the method - it was about implementation. These findings prove that parameter-efficient fine-tuning is highly effective for most post-training use cases. Key technical insights: • Apply LoRA across every layer, including MLP/MoE blocks - not just attention layers • Use learning rates approximately 10x higher than traditional full fine-tuning • Even rank=1 configurations deliver strong RL performance • Achieve these results with 33% less compute than FFT What this means for innovation: Anyone can now train sophisticated RL models using algorithms like GRPO or GSPO - even on for free platforms like Google Colab / Kaggle with a T4 chip using the awesome optimization projects from the amazing folks at Unsloth AI. So the barrier is no longer hardware access; it's all about the hyperparameters, approach and strategy. This development will make advanced RL training accessible to a much broader range of researchers and developers, potentially accelerating innovation across the entire field. As I always say, 'necessity is the mother of all innovations' #MachineLearning #AI #ReinforcementLearning #DeepLearning #OpenSource

🧠 If your AI forgets 99% of what it reads... can it ever truly reason? Bigger models won't win the future. The AI that remembers will. For years, AI progress has focused on more parameters, bigger models, and higher compute. Real progress in AI won't come from scaling parameters - it will come from remembering more. ✨I’m excited to share that: Galaxia can now process and transforms over 200 million characters into a fully explainable knowledge hypergraph in a single pass. That’s 60,000+ pages, or 200 copies of The Lord of the Rings - all in-memory, on CPUs, with full transparency and traceability. This is not “long context.” This is long-term, persistent memory. 💡What This Means LLMs don’t think - they respond within the limits of their short-term memory window. Once that window fills, earlier information disappears - logic breaks, hallucinations begin, explainability ends. But when AI gains persistent long-term memory: 🔹 It stops hallucinating - because context isn’t lost 🔹 Reasoning scales across thousands of documents - not just a few pages 🔹 Every answer is explainable - with full provenance of where it came from 💡Why This Is a Paradigm Shift Galaxia introduces semantic hypergraph memory - a persistent reasoning layer that transforms raw data into structured, interconnected knowledge at scale. This directly solves the core architectural limitations of transformer-based LLMs: 🔹 No approximation - knowledge is explicitly modeled, not guessed 🔹 No context ceiling - reasoning spans entire document corpora 🔹 No black box - every inference path is transparent and traceable This is not an incremental improvement to LLMs - it is a structural advancement that complements and completes them, marking the shift from probabilistic text to explainable intelligence. More here: https://lnkd.in/edVSxGhg

If your AI forgets 99% of what it reads… can it ever truly think? Forget “bigger models.” The future belongs to AI that remembers. Meet Galaxia — the first AI platform that transforms 200M+ characters (60,000 pages!) into a fully explainable knowledge hypergraph — all in-memory, on CPUs. No hallucinations. No black box. Just reasoning that lasts. 🧠 Persistent memory = ✔ Context that never gets lost ✔ Explainability that’s built-in ✔ Reasoning that scales infinitely This isn’t long context. This is long-term intelligence. #AI #DeepTech #ExplainableAI #MemoryIsTheFuture #Galaxia #smabbler

Check out how building video analytics AI agents just got a major upgrade. The new NVIDIA AI Blueprint for Video Search and Summarization (VSS 2.4) has four key upgrades that make it easier than ever to turn video streams into actionable insights. #nvidia #ai #artificialintelligence #computervision #genai #generativeai

AI That Learns Your Way - and Thinking Machines That Think With You Fine-tuning AI used to be rocket science: huge datasets, massive GPUs, and tons of coding. But in 2025, that’s changing fast. AI is becoming adaptable, efficient, and personal. Tools like OpenAI Fine-Tuning Studio and Mistral Fine-Tuner now let anyone shape big models (GPT-4o, Mistral 7B, LLaMA 3) to fit their own needs. You don’t need to be a deep-tech engineer - it’s more about guiding the model than rebuilding it. And visionaries like Mira Murati, former OpenAI CTO and now founder of Thinking Machines Lab, are taking it further - creating tools that make model customization simpler, safer, and more human-friendly. Better data beats bigger data. Clean, curated datasets make your AI smarter. Fine-tuning AI is no longer just for researchers. It’s a practical, safe, and fast way to make AI work for your needs. #AI #MachineLearning #FineTuning #DataScience #TechTrends #MiraMurati #ThinkingMachinesLab #Innovation #FutureOfAI

The 3 Layers of the AI Stack — Explained in 60 Seconds AI may appear magical, but underneath it lies a clear structure — a three-layer pyramid that defines how everything works. 1️⃣ Infrastructure Layer (Foundation) This is the base of the pyramid — the hardware and cloud backbone that powers AI. It includes GPUs, data centers, and computing resources that make model training and deployment possible. Without this layer, nothing stands. 2️⃣ Model Development Layer (Core Intelligence) This is where foundation models such as GPT, Claude, and Gemini are built and trained. Data scientists and engineers feed massive datasets, fine-tune parameters, and develop architectures that enable reasoning, language, and perception. This layer gives AI its mind. 3️⃣ Application Layer (User Experience) This is the visible tip of the pyramid — the tools we use every day. Chatbots, copilots, image generators, and enterprise solutions all live here. It’s where AI translates intelligence into impact. The future lies in mastering how these three layers connect. Those who understand the full AI stack, from infrastructure to applications, will shape how AI transforms industries. I’ve simplified the entire concept into a one-page infographic: “The 3 Layers of the AI Stack — Explained Simply.” Here’s a simple one-page infographic that visualizes it all. #AI #ArtificialIntelligence #MachineLearning #GenerativeAI #AIStack #LearningEveryday #TechSimplified #MCT #IAinAI #AIinIA #auditdudeCN Brij kishore Pandey Sudharshana Bharathi

⚡ What if every computer could power the world’s imagination? That’s what we’re building with Promptus — a distributed compute network where thousands of GPUs come together to render, generate, and build creative AI. We’re transforming unused compute into a living global network that fuels art, 3D, and video for creators everywhere. Because the future of creation isn’t centralized — it’s connected. 💡 Want to be part of it? Join the mission, power the network, and help us build the world’s creative engine. 👉 Learn more at https://lnkd.in/daYUmw5Y #Promptus #GPUcompute #distributed #computenetwork #AI

AI success isn’t limited by models or GPU’s, it’s limited by data. 🤖   While organizations invest heavily in AI infrastructure, many still face critical bottlenecks: slow data access, stale datasets, and idle GPUs. The real challenge? Making distributed data instantly actionable, wherever it lives.    Our latest blog explores why accessible data is the missing key to AI success, and how the right tools unlock its full potential. 🗝️   Read here: https://lnkd.in/eDFYcseD #AI #DataMovement #RemoteDataAccess #AIInnovation