A linear probe that scores 98% on held-out data can still be measuring the wrong thing, and much of this work is finding out which. This opening sits among the few remote AI alignment researcher jobs: full-time, home-based in any country, and paid $125,000 a year (USD). You would design interpretability and reward modeling experiments, look inside open models with TransformerLens, and write up what held and what did not. It suits people with a PhD, an equivalent research record, or a finished program such as MATS.
What the team is trying to find out, and why this seat exists
The project team studies why language models behave as they do. Two threads run through the work: interpretability, which asks what a model represents internally, and reward modeling, which asks whether a training signal rewards what people want.
This AI alignment researcher job description exists because the team needs someone who can turn a vague worry, such as "the reward model seems to like long answers", into an experiment with honest error bars.
A week of experiments, write-ups, and review
- Design interpretability experiments in PyTorch or JAX, including probes, activation patching, and ablations on open models.
- Run reward modeling studies that test whether a reward signal tracks the intended behavior or a cheap proxy for it.
- Analyze model internals with TransformerLens, tracing which components drive a given output.
- Write internal notes after each experiment, including results that failed or stayed ambiguous.
- Turn the strongest findings into public research posts, papers, or open-source code, with colleague review.
- Collaborate on methods that make models more reliably aligned, from evaluation design to training changes.
The research record the hiring team looks for: PhD, MATS, or a body of work
A PhD is usually expected, but it is not the only way in: an equivalent research record counts, and so does completing a program such as MATS. No fixed number of years is required.
- A PhD or equivalent research record in ML, CS, or a related field, or a completed program such as MATS.
- Strong PyTorch or JAX, enough to write and debug a training loop.
- Sound experimental design: controls, baselines, and held-out data.
- Clear technical writing that others can check.
Nice to have
- Hands-on experience with TransformerLens.
- Published or open-source alignment work.
The AI alignment researcher skills that matter most are less about knowing every technique and more about skepticism towards your own results.
Pay, time zones, and the uneven rhythm of research weeks
The salary is $125,000 a year in US dollars for a full-time role, or close to ₹1.19 crore at ₹95 to the dollar, a number that shifts as currencies move. As with other AI alignment researcher jobs remote worldwide, the work runs online: your own time zone, a shared repository, and video calls for planning and reading groups. The AI alignment researcher salary per year varies widely by lab and seniority.
Weeks are uneven; nothing promises a publication count or promotion.
From research question to experiment: a worked probe on GPT-2 small
A small early study; the numbers are illustrative, not real results.
Question
Does GPT-2 small linearly represent whether a subject is singular or plural, and use that to choose "is" or "are"?
Hypothesis
A logistic regression probe on the residual stream at the subject token should decode number from early layers, and patching that position from a plural run into a singular run should push the prediction towards "are".
Method
- Generate 2,000 templated sentences such as "The key to the cabinets", where the distractor noun can disagree with the subject.
- Load the model as a HookedTransformer and use run_with_cache to collect the subject-token residual stream at all 12 layers.
- Train one probe per layer on 80% of templates and test on the rest, with no noun in both sets.
- Patch the subject position layer by layer from each plural prompt into its singular twin and measure the logit difference between " is" and " are".
Result (illustrative)
Probe accuracy passes 95% by layer 2. Patching the subject position restores about 70% of the logit difference at layers 3 to 6, then falls off as the information seems to move to the final token.
Caveats a reviewer would raise
- Most plurals here end in "s", so the probe may read spelling, not number; irregular plurals such as "children" are needed as a control.
- Decodable is not the same as used. Patching is the stronger evidence, and it covers one template family.
- Without a random-direction baseline and several seeds, 70% has no reference point.
- GPT-2 small has 124 million parameters; whether this holds at scale is not yet known.
Next step
Add the control, ablate the probe direction, and repeat across several Pythia sizes. A typical review comment: "Interesting, but I would not believe it until the control is in." That is normal, not a rejection.
Interpretability, reward modeling, and the tools on your machine
You will work in Python with PyTorch or JAX, TransformerLens for model internals, Jupyter and Git. Large runs go to cloud GPUs arranged through the project; at home you need reliable internet, a decent screen and quiet time for writing.
AI alignment researcher work-from-home jobs need one habit above others: post a short progress note, even "no result yet", so nobody loses three days on a bug alone.
Routes into remote AI alignment researcher jobs without a PhD
On how to become an alignment researcher without a doctorate: replicate a published interpretability result, extend it slightly, and publish it with code. MATS, a mentored alignment research program, is one well-known path; the Anthropic Fellows Program is another fellowship that funds safety research for a fixed period.
No single AI alignment researcher course gets anyone hired, and no recognized AI alignment researcher certification exists. Your own experiments count for more. Searches for "AI alignment researcher jobs no experience" mostly find fellowships, a sensible order: entry-level remote AI alignment researcher jobs usually go to people with one solid project behind them.
Resume, research sample, and the questions a panel tends to ask
An AI alignment researcher resume should lead with outputs: papers, posts, repositories, and what each found. Name the methods used and say where a result did not replicate.
Sample AI alignment researcher interview questions
- Your probe gets 97% accuracy. What would convince you the model actually uses that feature?
- A reward model prefers longer answers. How would you test whether that preference is a proxy or a real signal of quality?
- Which of your results turned out wrong, and how did you find out?
Alignment researcher or AI safety engineer: where the line falls
A researcher answers open questions, and a well-written null result is a legitimate outcome. An AI safety engineer builds the systems around that research: evaluation pipelines, monitoring and safeguards that ship with a product. People often move between the two.
Funding, demand and what nobody can promise
Funding comes from AI labs, nonprofits such as Redwood Research and FAR.AI, and philanthropic grants. Demand has grown, but the field is small and its shape in five years is unclear. If you track remote AI alignment researcher jobs hiring now, keep building public work; it stays useful no matter what any one funder does.
Fee requests and fake research offers, and how to apply here
Walk away from any "research fellowship" asking for a registration fee, a paid "starter kit", a one-time code from your bank, a UPI transfer, or one recruiting only over Telegram or WhatsApp. An offer with no research sample or interview is another warning sign. No fee applies here at any point.
Apply at no cost by sending your CV, links to your research or code, and a short note on one question you would test first. Not every applicant will be shortlisted.
