Sulav Ojha

Exploring AI & tech. let's connect

@sama No seat at the GPT-5.5 party, but hoping it still brings enough heat to challenge Claude 😄

we are gonna do something nice for everyone who applied for the GPT-5.5 party and that we didn't have space for. hope you enjoy!

7 Must-Have Skills for Applied AI Engineers in 2026 • LLM Fundamentals (Transformers + Tokenization + Fine-tuning basics) • Python for AI (PyTorch / TensorFlow / JAX basics) • RAG Systems (Retrieval-Augmented Generation design + implementation) • Vector Databases (Embeddings, similarity search, indexing) • AI Deployment (APIs, FastAPI, cloud inference, scaling models) • MLOps Basics (model versioning, pipelines, monitoring, CI/CD for AI) • AI Product Integration (building AI features into real apps/tools)

AI / Machine Learning Job Salaries (2026) • 🧠 AI Research Scientist — $180K–$500K+ • 🧪 Research Engineer (AI Systems) — $160K–$350K+ • 🤖 Machine Learning Engineer (MLE) — $140K–$300K • ⚙️ AI Infrastructure Engineer — $150K–$350K • 📊 Data Scientist (AI/ML-focused) — $120K–$220K • 🧱 MLOps Engineer — $130K–$250K • 🧬 LLM Engineer (Generative AI) — $160K–$400K • 📦 Data Engineer (AI Pipelines) — $120K–$220K • 🧠 Applied Scientist — $150K–$320K

Why is GPU access the real AI bottleneck? Because compute = intelligence. → more GPUs = better models → limited supply = higher cost → Nvidia controls pricing power Everyone is compute-constrained.

@koderayush congrats

It hasn’t even been 3 days on X yet, and I’ve already reached 200+ followers. Thank you for all the support and love ❤️ Manifesting 300 by tomorrow

Tech stack for starting a modern business: • Claude for coding • Supabase for backend • Vercel for deployment • Namecheap for domain name • Stripe for payments • GitHub for version control • Resend for email sending • Clerk for authentication • Cloudflare for DNS management • PostHog for analytics • Sentry for error tracking • Upstash for Redis • Pinecone for vector database You can start your own company with just a laptop and an internet connection. This job isn't as complicated as you might think.

Why is @deepseek_ai 100x cheaper than @AnthropicAI? China is built to be cheap. → Cheap model: designed to use fewer tokens, reuse results (caching), less compute per request → Cheap chips: uses local hardware (no expensive Nvidia dependency) → Cheap energy: lower electricity costs, often supported by the state → Cheap talent: strong engineers, paid much less than US labs → Cheap economics: funded by trading profits, so the AI itself doesn’t need to be highly profitable The only gap is performance. But as models become “good enough,” price starts to matter more. And staying at the frontier keeps getting more expensive.

Most people trying to build AI automations are wasting time. Top agencies aren’t. They’re using tools like this: → 19,000⭐ → 1,500+ nodes → open-source n8n-mcp. Save this before you need it.

If you’re serious, you should recognize these: 🧠 TensorFlow → production-scale ML 🔥 PyTorch → research + flexibility 🤖 Scikit-learn → classic ML foundation ⚡ Keras → fast prototyping 🤗 Hugging Face Transformers → LLM ecosystem 👁️ OpenCV → image/video processing 🌲 XGBoost → structured data winner 🚀 fastai → practical deep learning 🔗 LangChain → agent workflows 📦 ONNX → model portability Which one are you actually using in production? 👇 Save this before you forget. 📌

Everyone is obsessed with “better models.” Meanwhile, the winners are: • reducing latency • improving evals • designing workflows • controlling cost Engineering > model choice.

You’re not an AI Engineer until you understand these terms: • 🧠 Embeddings → Numerical meaning of text/data • 🔍 Vector DB → Similarity search storage • 📚 RAG → Retrieval-Augmented Generation • 🎯 Fine-Tuning → Task-specific model training • 🪶 LoRA → Lightweight fine-tuning method • ⚡ Quantization → Smaller/faster models • 🧵 Context Window → Model memory limit • 🔄 Function Calling → Structured tool usage • 🛡 Guardrails → Output constraints/safety • 📏 Eval Frameworks → Measure model quality • 🧮 Tokenization → How text becomes tokens • 🚀 KV Cache → Faster inference reuse • 🔥 Hallucination → Confident wrong output • 🪝 Prompt Chaining → Multi-step workflows Building demos is easy. Production AI is not.

Now it’s free for 24 hours. Perfect moment to actually ship something, not just scroll.

CONTENT CREATION JUST COLLAPSED. One command now replaces scripts, editors, and ad teams. Pika MCP turns any idea into podcasts, explainers, or UGC ads instantly. Drop input → get finished video. This changes everything.

@XFreeze Soon, robots will be everywhere…

Soon, a billion of these robots will be working around the clock... helping humanity achieve our greatest feats Moving and operating just like us...a machine fully interacting with our physical world

SOMEONE JUST OPEN-SOURCED “Claude Code for finance.” It’s called Dexter. You give it a stock, and it builds a full investment thesis end-to-end. Not just data, it forms a clear opinion with reasoning on growth, margins, and risks. This is institutional-grade research, now running on a laptop.

90% of AI Engineer interviews in 2026 will test these concepts: * Transformers / Attention * Embeddings / Vector Search * RAG Architecture * Fine-Tuning / LoRA / PEFT * Prompt Engineering / Structured Outputs * LLM Evaluation / Benchmarking * Hallucination / Guardrails * Inference / Latency Optimization Not just “build a chatbot.”

SOMEONE BUILT A SELF-HOSTED AI APP THAT HANDLES YOUR RECEIPTS AND INVOICES AUTOMATICALLY . It’s called TaxHacker. Just snap a photo or upload a PDF, and it pulls out amounts, dates, vendors, taxes, and line items into a clean database. 100% open source.

We agree that Google (Gemini) is officially out of the AI race, right ?!