01Why India matters
Most AI safety research happens in the San Francisco Bay Area and London. It cannot stay there. India has over 5.4 million software engineers and a large pool of STEM graduates. Researchers in India can help most in three areas:
- Interpretability. Building open tools, such as sparse autoencoders, circuit analysis and activation patching, to see what happens inside neural networks.
- Red-teaming in Indian languages. Testing whether safety filters fail when the language or the cultural context changes.
- Compute governance. Advising MeitY and international bodies on how to monitor and verify the hardware that trains frontier models.
02Four career tracks
AI safety work falls into four tracks. Each asks for different skills and a different kind of portfolio.
Mechanistic interpretability
Working out what a transformer has learned, with sparse autoencoders, residual stream analysis and activation patching. Tools: PyTorch, TransformerLens, SAELens, einops and CUDA.
Scalable oversight and evaluations
Building red-teaming benchmarks, catching models that sandbag or hide a backdoor, and working on RLHF, DPO and debate. Tools: Inspect, Hugging Face, vLLM, DeepSpeed and LangChain.
AI governance
Research on compute thresholds, licensing for frontier models and safety cases, and advice to MeitY, NITI Aayog and international bodies. Skills: policy memos, economics, game theory and legal analysis. Start with our AI Governance course.
Field-building
Running university groups, hackathons and reading groups, and mentoring new researchers across Indian colleges. Skills: running a community, writing grants and designing courses. See run a group.
03A six-month plan
Moving from web, cloud or ML engineering into safety research takes 3 to 6 months of focused work. Go through four stages, and build public work as you go.
Stage 1, months 1 and 2: foundations
Learn how a transformer works inside: attention, the residual stream, layer norm and positional embeddings. Work through ARENA chapters 0 and 1, or build GPT-2 from scratch in PyTorch.
Milestone: join our free six-week Technical AI Safety course and finish the weekly labs.
Stage 2, month 3: replicate a paper
Pick a recent interpretability or red-teaming paper, such as indirect object identification circuits, representation engineering or jailbreak transfer. Run its experiments again on an open model such as Gemma 2 2B or Llama 3 8B.
Milestone: a clean GitHub repository with notebooks that run, and a write-up of about 1,500 words on LessWrong or the Alignment Forum.
Stage 3, months 4 and 5: your own research
Start a short project of your own. From 2027, top graduates of our courses can apply for our 12-week research fellowship with a stipend. You can also ask the Long-Term Future Fund for a grant for independent research and compute.
Milestone: a preprint on arXiv, a talk at one of our sessions, or an evaluation tool others can use.
Stage 4, month 6 on: fellowships and jobs
Use your public work to apply to MATS, SPAR or Astra, or for safety roles at Anthropic, METR, Apollo Research and the UK and US AI Safety Institutes. Our hiring page lists open roles at AI Safety India, and our directory of AI safety organisations in India shows the research groups, think tanks and policy bodies working here.
Milestone: a funded fellowship or a research job.
04Fellowships and programmes
Several funded programmes take applicants from India, remote or with travel paid. Amounts are in US dollars. Check each programme's own site for its current dates and terms.
| Programme | Format | Support | Focus |
|---|---|---|---|
| AI Safety India research fellowship | Remote, India, from 2027 | Stipend and compute | Interpretability, Indic evaluations, governance |
| MATS | Berkeley, London or remote | $5,000 to $7,000 a month, and a J-1 visa | Mentored alignment research |
| ARENA | London or Berkeley | Travel, housing and a stipend | Intensive PyTorch and interpretability bootcamp |
| SPAR | Remote or hybrid | Research stipend | Supervised alignment research with senior researchers |
| Open Philanthropy undergraduate scholarship | University, worldwide | Full tuition and living costs | Outstanding STEM undergraduates |
05Pay and stipends
Funded AI safety work often pays as well as a software job in India, or better. Typical ranges, in US dollars:
- Independent grants: $2,500 to $5,000 a month, from LTFF, Manifund or SFF, for 3 to 6 months of your own research and its compute.
- Fellowships such as MATS and SPAR: $4,000 to $7,000 a month, often with travel, health cover and compute.
- Frontier labs: $150,000 to $400,000 a year or more, for technical staff at Anthropic, Google DeepMind, OpenAI, METR and Apollo Research.
06Three traps to avoid
The most common reason people do not make the move: they keep reading and never publish code or results.
Reading without building
Reading 50 papers without writing any PyTorch or testing a model. Fix: for every 5 hours of reading, spend 5 hours coding with TransformerLens or SAELens.
Keeping your work private
Waiting until your project is as strong as a lab paper before you share it. Fix: post negative results, small replications and short write-ups on the Alignment Forum or LessWrong.
Working alone
Studying with no one to check in with. Fix: join our weekly discussions, or book a free advising call and we connect you with a research group.
07What to read first
These papers are the base of today's alignment, interpretability and risk work.
- AGI safety from first principles, Richard Ngo. Theory.
- A mathematical framework for transformer circuits, Nelson Elhage and others, Anthropic. Interpretability.
- Scaling monosemanticity: extracting interpretable features, Adly Templeton and others, Anthropic. Sparse autoencoders.
- Sleeper agents: training deceptive LLMs that persist through safety training, Evan Hubinger and others. Threat models.
- Frontier AI regulation: managing emerging risks to public safety, Markus Anderljung and others. Governance.
08Next steps
- Book one-to-one career advising. It is free. We help you pick a project, review fellowship applications and meet mentors.
- Join the Technical AI Safety course, six weeks, free.
- Join the AI Governance course, six weeks, free.
- See open roles at AI Safety India.
- Browse AI safety organisations in India, and add yours if it is missing.
09Questions
Can I move into technical AI safety with only a software engineering background?
Yes. Many interpretability and red-teaming roles value strong engineering, such as systems, CUDA, PyTorch and distributed training, more than pure maths. A six-week course and an open-source replication with TransformerLens or sparse autoencoders is often enough to apply to fellowships such as MATS or ARENA.
How does pay in AI safety compare with big tech jobs in India?
Fellowships such as MATS, SPAR or FAR AI pay stipends of about $3,000 to $7,000 a month, remote or in person. Frontier labs such as Anthropic, Google DeepMind, OpenAI and METR pay market rates for ML staff: $150,000 to $400,000 a year or more for senior roles, plus equity.
Do I need a PhD to work in AI safety or AI governance?
No. Empirical alignment research values public output, such as research, open-source tools and write-ups on arXiv, LessWrong or the Alignment Forum, more than degrees.
What is the best first step for an undergraduate in India?
Apply to our free six-week Technical AI Safety course, join or start an AI safety group at your college, and look at student funding such as the Open Philanthropy undergraduate scholarship or the Kairos fellowship.
How is AI safety work in India different from San Francisco or London?
Those cities host the labs that train frontier models. India has one of the largest developer populations in the world, a non-aligned voice in international talks on compute, and urgent needs of its own: evaluation benchmarks in Indian languages, and advice on governance.
What should my first 30 days look like?
Week 1: read AGI safety from first principles by Richard Ngo, and Neel Nanda's 200 Concrete Open Problems in Mechanistic Interpretability. Week 2: work through ARENA chapter 1 and build a transformer from scratch in PyTorch. Week 3: join our weekly discussions and a reading group. Week 4: book career advising to choose a replication project.