Tech with Mak

There are 2 career paths in AI right now: The API Caller: Knows how to use an API. (Low leverage, first to be automated, $150k salary). The Architect: Knows how to build the API. (High leverage, builds the tools, $500k+ salary). Bootcamps train you to be an API Caller. This free 17-video Stanford course trains you to be an Architect. It's CS336: Language Modeling from Scratch. The syllabus is pure signal, no noise:  ➡️ Data Collection & Curation (Lec 13-14) ➡️ Building Transformers & MoE (Lec 3-4) ➡️ Making it fast (Lec 5-8: GPUs, Kernels, Parallelism) ➡️ Making it work (Lec 10: Inference) ➡️ Making it smart (Lec 15-17: Alignment & RL) Choose your path. (I will put the playlist in the comments.) ♻️ Repost to save someone $$$ and a lot of confusion. ✔️ You can follow @techNmak, for more insights.

@Umesh__digital That's so true!

@techNmak Planning is imp for building a product

The fastest way to build the wrong thing is to skip the planning. We've been speedrunning that for 2 years now. Not anymore.

@Arindam_1729 Indeed!

@techNmak This is cool

@AntonMartyniuk Cool stuff man!

AI agent forgot the implementation details we discussed yesterday And that's the real problem Every team building AI agents hits the same wall. A user talks to your agent on Monday. Gives it context, code examples, PR history. On Tuesday, the agent has zero memory. The user starts from scratch. Most teams try to fix this in 3 ways. All fall short. ❌ Expanding context windows Bigger context windows are just expensive sticky notes. Your conversation fits, but disappears the moment you refresh. No learning, no memory, no persistence. ❌ RAG systems can't connect the dots They retrieve documents but can't remember that YOU prefer specific architectures, or that your team decided against microservices last month. ❌ Most companies patch together 3+ databases Vector DB for embeddings, graph DB for relationships, SQL for metadata. Result: fragile architecture, security nightmares, and zero shared transactions. I spent weeks researching how to solve persistent agent memory without building a Frankenstein stack. That's when I found Oracle's AI Database approach. Instead of using multiple systems, ONE database handles everything: → Vectors for semantic understanding → Graphs for relationship mapping → Relational data for business context → All with ACID compliance across data types This means when your agent stores a new memory, vectors, graphs, and relational data update together. No partial writes. No inconsistent states. 📌 What stood out to me the most: Row-level isolation ensures your conversations stay private. Full EU AI Act and GDPR compliance, including "right to be forgotten" while maintaining 10-year audit trails. If you are building AI agents that need to remember users across sessions, this is worth exploring. 👉 Get started with the Oracle AI Database free resources: fandf.co/4rbGwb8 —— ♻️ Repost to help others fix AI agent memory ➕ Follow me ( @AntonMartyniuk ) to improve your .NET and Architecture Skills Many thanks to @oracle for sponsoring this post

@coderabbitai "Move fast and break things" - Every AI agent writing code without a plan "Move fast and ship the right things" - What this actually enables Great job @coderabbitai !!!!

Introducing CodeRabbit Plan. Hand those prompts to whatever coding agent you use and start building!

@ghumare64 Looks solid! 👌

🚨BREAKING: AgentOS just launched Open-source infrastructure for running AI agents in production. • Rust-first runtime • TS/Python/Rust workers • Native triggers, state, streams, channels • Built-in eval + feedback loops 100% open source.

Read here: nvidia.com/en-us/ai/nemoc…

Here's the GitHub: github.com/NVIDIA/NemoClaw

🚨 NVIDIA just made OpenClaw safe to run 24/7. Here's how they did it without breaking what makes OpenClaw useful. The problem with AI agent security: Most solutions restrict what the agent can do. → Can't access files → Can't make network requests → Can't call APIs But that defeats the purpose. OpenClaw's value IS its access. NemoClaw takes a different approach: Four protection layers, two enforcement modes: Locked at creation (immutable): → Filesystem: Only /sandbox and /tmp are writable → Process: Privilege escalation blocked via Landlock + seccomp Hot-reloadable at runtime (flexible): → Network: Add/remove allowed hosts without restart → Inference: Reroute model calls to different backends Why this split matters: > Filesystem and process boundaries should never change. That's your security foundation. > But network and inference policies need flexibility. New integrations. New models. New use cases. NemoClaw gives you both: Immutable security boundaries + flexible operational policies. When OpenClaw tries to reach an unlisted host: 1. OpenShell blocks the request 2. Surfaces it in the TUI for your approval 3. You decide: Allow or deny 4. If allowed, policy updates without restart That's "trust within boundaries, intervene on exceptions." Image credit - The New Stack

GitHub Repo: github.com/abhigyanpatwar…

🚨 BREAKING: Someone just built a tool that lets you talk to your codebase like it's a database. It's called GitNexus. One command turns your repo into a queryable knowledge graph: npx gitnexus analyze Now you can ask questions no search tool could answer before: "What breaks if I change this function?" "Which execution flows touch authentication?" "What's the blast radius of my uncommitted changes?" Here's why this is different from code search: Code search finds text matches. GitNexus finds relationships. It indexes: → Every function, class, method, interface → Every import, call, inheritance chain → Every execution flow from entry point to completion → Every functional cluster with cohesion scores Then exposes 7 MCP tools so your AI agent can query it: > impact - "47 functions depend on this, here's the risk by depth" > context - "This function is called by 8 things, calls 3 things, participates in 2 execution flows" > detect_changes - "Your uncommitted changes affect LoginFlow and RegistrationFlow" > rename - "Renaming this touches 5 files, here's a dry run" > cypher - Raw graph queries for anything else The insight: Your codebase is already a graph. Functions call functions. Classes inherit classes. Modules import modules. GitNexus makes that implicit structure explicit and queryable. 16.8K stars. I'll put the GitHub in the comments.

The uncomfortable truth about AI coding agents in 2026: They can fix your bugs. They can write your features. But they cannot stop themselves from breaking things that already work. SWE-CI measured the "zero-regression rate" - the percentage of tasks where a model completes the entire maintenance process without breaking a single previously passing test. The results: → Most models: below 25%  → Only 2 models (both Claude Opus) exceeded 50% → Every other model breaks existing functionality in 3 out of 4 scenarios This means in a real production codebase, an autonomous AI agent would be constantly creating new bugs while fixing old ones. Snapshot benchmarks can't see this. SWE-CI can. We're still far from trusting AI with unsupervised codebase ownership.

Here's the guide: unsloth.ai/docs/basics/cl…

Claude Code can run entirely on your local GPU now. Unsloth AI published the complete guide. The setup itself is straightforward - llama.cpp serves Qwen3.5 or GLM-4.7-Flash, one environment variable redirects Claude Code to localhost. But the guide is valuable because of what it explains beyond the setup: Why local inference feels impossibly slow: Claude Code adds an attribution header that breaks KV caching. Every request recomputes the full context. The fix requires editing settings.json - export doesn't work. Why Qwen3.5 outputs seem off: f16 KV cache degrades accuracy, and it's llama.cpp's default. Multiple reports confirm this. Use q8_0 or bf16 instead. Why responses take forever: Thinking mode is great for reasoning but slow for agentic tasks. The guide shows how to disable it. The proof it all works: Claude Code autonomously fine-tuning a model with Unsloth. Start to finish. No API dependency. Fits on 24GB. RTX 4090, Mac unified memory.

Read the paper here: arxiv.org/pdf/2603.11798

Why SQL beats attention for multi-document reasoning. Long-context LLMs suffer from "contextual dilution" - key entities get lost due to attention saturation. DocSage's solution: Don't reason with attention. Reason with SQL. Here's what happens in the reasoning module: Step 1: Query Compilation → Natural language question → optimized SQL query → Schema provides join keys and relationship definitions → Compiler pushes down filters, chooses efficient join order Step 2: Execution → SQL runs on structured database → Returns structured result set Step 3: Evidence Traceback → Each row traced to originating tuples → Tuples mapped to specific document locations → Full provenance chain Step 4: Answer Synthesis → LLM generates natural language answer → Based on result set + complete provenance Every claim is verifiable. Every answer is traceable. Multi-hop reasoning becomes deterministic database operations. Link to the paper in comments.

Here's the GitHub repo - github.com/HandsOnLLM/Han…

Jay Alammar is the best teacher in AI. Period. If you have ever seen "The Illustrated Transformer," you know his diagrams are legendary. He also open-sourced the entire codebase for his O'Reilly book: Hands-On Large Language Models. It’s effectively a visual masterclass in LLMs for free. Chapter 1: Introduction to Language Models Chapter 2: Tokens and Embeddings Chapter 3: Looking Inside Transformer LLMs Chapter 4: Text Classification Chapter 5: Text Clustering and Topic Modeling Chapter 6: Prompt Engineering Chapter 7: Advanced Text Generation Techniques and Tools Chapter 8: Semantic Search and Retrieval-Augmented Generation Chapter 9: Multimodal Large Language Models Chapter 10: Creating Text Embedding Models Chapter 11: Fine-tuning Representation Models for Classification Chapter 12: Fine-tuning Generation Models I will put the repo link in the comments.

@devXritesh Great! :)

@techNmak Thanks for sharing Mak, Helpful for us.

Most people try to learn AI randomly. I mapped the entire AI engineering journey into a metro system. The problem with most AI roadmaps: They're linear. Step 1, Step 2, Step 3. As if everyone starts at the same place and wants the same destination. But AI engineering isn't linear. It's a network. → A software engineer skips Python basics, jumps straight to LangChain → A data analyst already knows Pandas, needs Transformers next → A product manager wants RAG and Agentic AI, not CNNs → A researcher needs Ethics & Safety before deployment A metro map captures this reality. Generative AI Hub (Line 4) connects to: → Machine Learning Loop (you need Transformers first) → Applied AI Sector (where RAG becomes chatbots) → Tooling & Deployment (where demos become products) Career Launchpad (Line 8) connects to: → Every other line (skills from any track convert to job offers) Ethics & Safety (Line 7) connects to: → Deployment (you can't ship without guardrails) → Applied AI (real-world projects need fairness and privacy) The 8 lines: 🟠 Foundations - Python, Math, Git (boarding passes) 🔵 Machine Learning - Neural Nets, CNNs, Transformers (the heart) 🟡 Deep Learning Express - LLMs, Fine-Tuning, PyTorch (fast track) 🟢 Generative AI Hub - RAG, Diffusion, LangChain (the magic) 🩷 Applied AI - Agentic AI, Healthcare, Chatbots (real projects) 🟣 Tooling & Deployment - Cloud, Kubernetes, MLOps (production) 🔴 Ethics & Safety - Bias, Privacy, Governance (guardrails) 🟢 Career Launchpad - Portfolio, Interviews, Networking (job offers) You don't take every line. You don't visit every stop. Find where you are. Pick your destination. Transfer as needed. Bookmark this. Start today.

@rathan_pmr Glad to share :)

@techNmak that's very valuable. thanks for the share