Company Guide
Frontier Labs
OpenAI Frontier
- Process: 4-6 rounds over 4-6 weeks. LeetCode Medium-Hard (Blind 75)
- Tests: GPT internals, scaling laws (Chinchilla), RLHF, KV cache, inference optimization, distributed training
- Coding: implement attention / a Transformer from scratch in NumPy/PyTorch
- System: design an inference platform at 10M users, train a 100B+ model
- Papers: GPT-3, InstructGPT, Codex, CLIP, Whisper
- Vibe: product-oriented, pragmatic, coding-heavy
Anthropic Frontier
- Process: 90-min CodeSignal (near-perfect score required), then 4-5 rounds
- Tests: Constitutional AI, RLHF vs DPO, reward modeling, mechanistic interpretability
- Safety: articulate the views on AI alignment. They ask directly
- System: guardrail design, content moderation, safety pipelines
- Papers: Constitutional AI (2022), Claude papers, Superposition (Elhage 2022)
- Tip: read anthropic.com/research, all papers are public
Google DeepMind Frontier
- Process: "PhD defense mixed with rigorous engineering exam", <1% acceptance
- Tests: rapid-fire linear algebra, probability, optimization. RL fundamentals for research roles
- Coding: attention backward pass, implement Adam, custom CUDA kernels, math proofs
- System: TPU-scale training, 3D parallelism, pipeline bubbles
- Papers: AlphaFold, Gemini, PaLM, Flamingo, Chinchilla
- Tip: know linear algebra cold. They ask about eigenvalues in ML contexts
Meta AI (FAIR) Frontier
- Focus: open-source LLM engineering (Llama stack) plus deep CV heritage
- Tests: LLaMA architecture, FSDP, PyTorch internals, distributed training
- CV angle: SAM, DINO, Detectron, MAE, segmentation/detection benchmarks
- Coding: custom CUDA kernels, PyTorch internals
- Papers: LLaMA 1/2/3, SAM, DINO, MAE
- Tip: know PyTorch better than any other framework
Indian AI Startups
Sarvam AI Startup
- Mission: IndiaAI Mission, first sovereign Indian LLM
- Founders: Vivek Raghavan & Pratyush Kumar (AI4Bharat / IIT Madras)
- Tests: multilingual tokenization, BPE for Indian scripts (Devanagari, Tamil), pre-training pipeline
- Data: IndicCorp, AI4Bharat datasets, cleaning low-resource corpora
- Research: IndicBERT, MuRIL, Dhruva TTS, Shuka
- Tip: know the AI4Bharat project history (founders' prior work)
Krutrim Startup
- Founder: Bhavish Aggarwal (Ola)
- Models: Krutrim-1 (7B), Krutrim-2 (12B), Indian language focus
- Tests: fine-tuning at 7B-70B scale, RLHF, instruction tuning, quantization
- Practical: cost-efficient serving (quantization, speculative decoding), Kubernetes ML infra
- Culture: startup pace, ownership, ship fast
- Tip: practical skills over theory. Can you ship?
Other Indian AI Startups Startup
- Uniphore: conversational AI, speech + NLP, enterprise focus
- Yellow.ai: enterprise chatbots, RAG systems, LLM integration
- Observe.AI: speech analytics, call center AI
- CoRover: multilingual chatbots, government contracts
- General pattern: Python + PyTorch + HuggingFace + practical LLM engineering
- Process: 2-4 rounds, faster decisions than frontier labs
Startup Interview Format Startup
- Round 1: screening call, basic ML + Python
- Round 2: technical (LLM concepts, RAG, fine-tuning walk-through)
- Round 3: take-home or live coding (build a mini pipeline)
- Round 4: hiring manager (project discussion, fit)
- Weight: practical ability > theoretical depth (opposite of frontier labs)
- Tip: show a GitHub with real LLM projects