Pune AI Community

Prompt engineering is a practical way to spread AI literacy. Its for all of us. Specifying Role, Context, Constraints, Examples, Output format etc and doing iterations to get the desired output. (My Sketchnote below) Shared this in a session at “मंथन” NEP (National Education Policy) 2020 workshop for junior college teachers from DES (Deccan Education Society), hosted at Fergusson College, Pune. Around 180 teachers from multiple cities #Mumbai, #Sangli, and #Pune joined. Teaching teachers matters. They amplify AI literacy further to a wider audience. Also participated in a panel discussion where we covered: ➡️ What AI tools fit different subjects and how should teachers discover them. Maths. Science. Arts. Commerce. ➡️ How can AI improve learning inside the classroom and outside it. ➡️ What happens if teachers decide not to use AI. ➡️ How should evaluation work in a post AI world. If assignments are easy to generate, what remains meaningful. ➡️ Most students and teachers are consumers of AI. How do they become smart consumers and eventually producers. ➡️ Can AI hurt cognitive skills. If yes, how do we avoid that. Thanks to Amit Kulkarni for the opportunity to be part of this 🎓 Manoj Deshmukh, Manoj Deshpande Dr Apoorva Shastri Dr. Rajesh Ingle Fergusson College, Brihan Maharashtra College Of Commerce, Willingdon College Of Arts and Sciences, Sangli, Chintamanrao College Of Commerce - India D E Societys Kirti M Doongursee College of Arts Science & Commerce, Kashinath Dhuru Road, Off. Veer Savarkar Marg, Near Agar Bazar Dadar West Mumbai 400 028 . #AILiteracy #PromptEngineering #NEP2020 #TeacherDevelopment #Education #DESPune #AIinEducation

When we talk about Deep Learning, we usually talk about layers, weights, and data. But underneath all that is a surface. The invisible shape/geometry where the data lives on. For years, we assumed that it was flat. Euclidean. Think about drawing the globe on a flat sheet. You get distortion. Greenland inflates. Distances lie. That same mapping error is quietly happening inside mixed curvature product manifolds. By making it flat, we gain computational efficiency at the cost of expressivity. The real world is not flat. Social networks bend differently (hyperbolic). The surface of the earth curves. So researchers built Product Manifolds. Mathematical LEGO that snaps together different curvatures to model messy reality. So, data 'surface' is combination of a bit of Euclidean, some hyperbolic and a touch of spherical. You get efficiency and something easy to optimize. The solution is obvious. Closed form geodesic distances. Optimization runs smoothly. Backprop behaves. One takeaway. Efficient geometry is not always faithful geometry. #GeometricDeepLearning #DeepLearning #AIResearch #RepresentationLearning #Manifolds #MachineLearningTheory #GenerativeAI #MathForAI (Ref: pic as is and info above is broadly based on podcast by Byte Goose AI)

AI is shaping the future - be part of it 🚀 We’re pleased to welcome Pune AI Community as a Community Partner for GDGOC Pune. Together with Pune AI Community and leadership from Yogesh Kulkarni, we’re bringing opportunities to learn, network, and explore real-world AI applications. 📅 25 April 2026 | 📍 Pune 🔗 www.wowpune.in 👉 Register and stay ahead in AI. #ArtificialIntelligence #MachineLearning #AICommunity #GDGOC #PuneTech #FutureTech #Leadership #Innovation

AI is everywhere. But are there areas that will not just survive, but actually thrive in an AI first world? Being Anti-Fragile per Nassim Taleb 🤔 We are seeing a collapse of the traditional build layer. Tools like Lovable, Replit and others can turn prompts into apps in seconds. When everyone can build, building itself stops being the moat. Your UI can be replicated in a week. Your feature can be cloned overnight. So what actually lasts? The future of the web seems to organize around five durable concepts that AI cannot easily replace. First is Trust. 🌐 When millions of AI generated apps flood the web, verification becomes gold. Signals like payment reliability, platform credibility and identity assurance matter more. Trust becomes the routing layer for both humans and agents. Second is Context. 📊 AI without your data is just a chatbot. AI with your company data becomes a teammate. Platforms that own structured context and permission layers become critical choke points. Third is Distribution. 📣 Building is easy. Getting discovered is hard. When supply becomes infinite, curation becomes scarce. Gatekeepers and niche creators both gain power in the attention economy. Fourth is Taste. 🎨 When production is free, choosing what to build becomes the real skill. Human judgment, product sense and orchestration quality define winners. AI assists, but taste directs. Fifth is Liability. ⚖️ When AI makes decisions, someone must be accountable. Healthcare, finance, legal and governance layers will grow because responsibility cannot be automated away. The key question for builders today is simple: What do you own that becomes more valuable when AI gets better? If the answer is nothing, rethink your positioning. If the answer is trust, context, distribution, taste or liability, you are building on durable ground. AI is not removing these fundamentals, the domain expertise. It is amplifying them. 🚀 (Ref: broadly based on the analysis by Nate B. Jones) #AI #FutureOfWork #Startups #ProductStrategy #AgenticAI #Innovation #TechTrends #Entrepreneurship #ArtificialIntelligence Google Google Developers Group Google Developer Experts

👀 Just created a GitHub repository dedicated for Frugal (Frugal Runtime for Underserved GPU-less AI Learners) Engineering, a full AI stack that runs on a dusty i3 laptop with 8GB RAM. It's called AIOOLL: Artifical Intelligence On Old Linux Laptop. No GPU. No cloud. No credit card. 💰 I have an old Intel i3 laptop sitting on my desk. 8GB of RAM. An SSD. Runs Ubuntu. Most people would say: "You can't do real AI work on that." AIOOLL says: 👀 Watch. Here's what it runs, fully offline, 100% locally: ✅ Classical ML (scikit-learn): spam detection, regression, cross-validation... ✅ Deep Learning (PyTorch CPU): bidirectional LSTM, tabular MLP ... ✅ RAG Pipeline (LangChain + ChromaDB): Q&A with local embeddings ... ✅ AI Agents (LangGraph): multi-step research + code review agents ... ✅ Computer Vision (OpenCV): face & motion detection at 15+ FPS ... ✅ Local LLM inference (Ollama): Gemma2, Qwen2, Phi-3 at 8 tok/sec ... Every module has: → Academic-grade implementation → Pytest test suite → Streamlit UI with visualizations → Real benchmarks on i3 hardware The name AIOOLL has two O's in the middle. 👀 Those are eyes. Wide open at what old hardware can actually do. 🔗 github.com/yogeshhk/AIOOLL Star it if you believe AI should be accessible to everyone, not just people with cloud budgets or gaming rigs.  And if you're sitting on an old laptop right now wondering if you can do real AI work, Yes. You can. 👀 #AI #MachineLearning #OpenSource #Python #LocalAI #Ollama #LangChain #PyTorch #Linux #BuildInPublic

Gemma 4 is here 🚀 and the most interesting part is the new terminology around “effective” models. • “Effective” means smarter parameter usage using per layer embeddings (PLE) • The E2B model has 2.3 billion effective parameters, 5.1 billion total parameters.  • The PLE technique gives each decoder layer its own small embedding for every token. • More intelligence per parameter without just scaling model size 🧠 • Small models now capable of strong reasoning and agentic workflows • Designed to run locally on consumer GPUs and even edge devices • E2B and E4B target mobile and offline multimodal use cases 📱 • 128K context for edge models, 256K for larger ones • Apache 2.0 license and available across HuggingFace, LM Studio, Ollama, and more Big takeaway: Extended “effective” models push serious AI capability into small, local, efficient deployments. Edge AI just got more practical. ⚡ We are trying it locally and will keep you posted!! #AI #Gemma #OpenSourceAI #LocalLLM #EdgeAI #AgenticAI #GenerativeAI #AIEngineering Google Rajat Bhatia Google Developer Experts Google Developers Group GDG Pune GDG Cloud Pune

Did you hear? 😲 Claude Code’s code got leaked!!! Looks like there are 11 layers of architecture. Thousands of lines. An agent orchestration platform disguised as a terminal chat. 🤯 No, we have not looked at it, but here are some findings collected from different sources and tweets 👇 1️⃣ CLAUDE.md is loaded not just at session start but every single turn. Every. Single. Turn. So it is very critical to have it well made. Not nothing and not the whole manual. Put your architecture decisions in there. Your file conventions. Your testing patterns. Your “never do this” rules. The model reads them every single turn. 📄 2️⃣ Subagents share the prompt cache. Parallelism is basically free. Most people use Claude Code like a single worker. One task at a time. Have say five agents. Same cost as one. Because they all hit the prompt cache. ⚡ 3️⃣ There are five compaction strategies. Context pressure is a real problem. Meaning use `/compact` proactively. Do not wait for the system to auto compact and lose context you care about. 🧠 4️⃣ Sessions are persistent and resumable. Stop starting fresh. Every conversation is saved as JSONL and the source code supports:--continue to resume your last session and --resume to pick a specific past session 🔁 Did you know all these features? Let me know anything new that you have gathered from the leak. 👀 Btw, this is the reality check: You are one npm away from a disaster!! 💥 #AI #ClaudeCode #LLM #DeveloperTools #AIEngineering #Coding #TechInsights #PromptEngineering #BuildInPublic #SoftwareEngineering 🚀 Anthropic The Hacker News #claude #anthropic #leak

From the Founder's desk ... ✍

🎓 Guiding the Next Generation into AI 🚀 Had the opportunity to conduct a mentorship session for students under the Educational Assistance Scheme (EAS), an initiative by Seva Sahayog Foundation. The session was delivered in Marathi (with English slides) to make it more relatable and effective💡 Covered: 🔹 What is AI?: breaking down the fundamentals 🔹 How to build a career in AI: practical pathways & guidance It was nice to interact with 120+ students from engineering, computer science, and commerce backgrounds, all eager to learn and grow 🙌 EAS, now in its 7th year, has supported 1000+ students in pursuing higher education through financial assistance and mentorship. Truly impactful work 👏 Grateful for the opportunity to contribute and support such bright minds in their journey 💙 🎥 Watch the full session here: https://lnkd.in/df-3E9_S 💬 What advice would you give to students starting their AI journey? #AI #Mentorship #Education #AIForStudents #CareerGuidance #Pune #SevaSahayog #AICommunity Dr. Anand Kolharkar Atul Nagras Shailesh Ghatpande