Job search
MATS Fellowship: Check the Intake, Research Streams, and Mentor Fit
Check MATS Winter 2027's closed intake, compare actual mentor-led research streams and distinguish main-program support from a separate funded extension decision.
MATS is a research fellowship, so start by identifying a research question you want to investigate and the kind of mentorship that would support it. You also need the correct intake: as checked on October 22, 2026, Winter 2027 applications are closed.
Use the published cohort to understand the route, then recheck the rules when a future round is announced. This guide compares two current streams and separates the main program from a later extension. Its worked examples are fictional private decision records, not application answers, research results or admissions advice. If you are also considering employment, explore LandOffer.ai as a separate search.

Check the cohort before preparing a route
The central application page says Winter 2027 applications are closed and identifies September 6 as the Stage 1 closing date. Its timeline places mentor review in October to early November and offers in early to mid-November. Those stages concern that intake; they do not reopen the general application to a new October prospect.
The Winter 2027 program page lists a 12-week program from January 19 to April 10, 2027, in Berkeley and London, with remote options. Treat that as a cohort description, not a standing annual calendar or a guarantee that a particular mentor will support your preferred arrangement.
At the top of your private notes, record “Winter 2027; central apply page; checked October 22, 2026; closed.” Recheck that status before acting. A stream page can still show an Apply link after the central intake closes. The presence of that link is not evidence that a new application will be considered. This review did not enter an application portal or test a late submission.
No next-round deadline is established here. For a future intake, re-read the central application page and the actual stream instructions. Do not carry over a deadline, mentor list or assessment route merely because it appeared in the last cohort.
Distinguish tracks, streams and mentor selection
MATS uses tracks for broad research areas and streams for mentor-led agendas. The application page describes applying to tracks first, then selecting streams if you advance. For Empirical applicants, Stage 2 includes assessments before stream selection; other tracks proceed directly to stream selection. Stream-specific questions, mentor interviews or work tests can add requirements.
The final matching process uses applicant and mentor rankings. Choosing a stream therefore expresses a preference; it does not reserve a place with that mentor. Read what the mentor wants to investigate and how they expect fellows to work before deciding which preference you could support honestly.
MATS's selection guidance emphasizes commitment to reducing AI risk, relevant research or demonstrated competency, and credible references. A general desire to enter the AI industry does not explain why a particular research agenda matters to you. Identify what you have already investigated, where your understanding stops and why you want to work on the unresolved question.
The FAQ also restricts LLM use: applicants may not use LLMs to write application material unless the specific form or work test explicitly permits it. The examples here are private route comparisons, not text to paste into an application. Write your own responses and follow the current instructions for each assessment.
Compare agendas through the work they require
Consider two Winter 2027 listings as examples. Abram Demski's stream concerns Agent Foundations and justified trust in AI systems. It asks for proof-writing ability and fluency with probability and expected utility theory, alongside philosophical judgment; familiarity with Logical Induction is a plus. Reading about alignment generally does not demonstrate those particular abilities.
The AI Village stream, with Shoshannah Tekofsky, describes analyzing existing agent data or developing an experimental variant. Its desired characteristics include independent data analysis, writing, visualization and experience with large datasets. That gives a different way to examine fit: what analysis have you owned, and can you explain its limitations?
Both stream pages describe standard weekly one-to-one mentorship of one to two hours. Their location preferences are marked indifferent. Neither detail establishes an unlimited supervision commitment or approval for remote participation. Read the project-selection and program logistics information together rather than converting a short preference field into a complete arrangement.
Compare the agenda, methods, your evidence and the missing evidence. Those four details help you decide what to investigate before prioritizing a stream. Record the mentor's name alongside the agenda, methods and evidence so you can explain the choice.
Worked fit record: Rina compares two streams
Rina is a fictional data analyst interested in understanding AI-agent behavior. Her stipulated background includes Python analysis of operational logs, visualizations explaining data-quality problems and two technical write-ups. She has taken an introductory proofs course but has not studied Logical Induction or completed an AI-safety research project.
Her completed private comparison is:
| Fit question | AI Village route | Demski route |
|---|---|---|
| Research interest | Agent behavior is the question she currently wants to investigate | Trust foundations interests her, but she cannot yet state a specific theoretical question |
| Methods she can support | Python log analysis, data cleaning and visualization | Introductory proof coursework only |
| Evidence she actually has in this fiction | Two write-ups describing methods and limitations | Course exercises, not independent theoretical research |
| Missing evidence | Experience with this research dataset and strong understanding of LLM behavior | Probability and decision-theory depth, relevant formalism and philosophical research judgment |
| Current preparation decision | Investigate the empirical agenda and its data-analysis expectations first | Keep as a reading direction; do not claim a demonstrated research fit |
Her finished route record reads: “For a future cohort, inspect whether the AI Village agenda is still offered and whether its questions remain interesting to me. My data-analysis evidence is relevant to some methods, not proof of readiness or admission. I need to understand the research setting and distinguish descriptive patterns from explanations before proposing a question. I am not ranking or submitting streams in the closed Winter intake.”
She selects three preparation outputs. First, write a private summary of a relevant public research question, identifying the claim, evidence and uncertainty. Second, document an analysis she already owns, including why a visualization might mislead and how she checked the underlying records. Third, list which parts of the current stream requirements those materials support and which remain unsupported.
These outputs describe preparation, not completed findings. We did not run an agent experiment or analyze the AI Village dataset for this guide. Rina should not rename an ordinary business-data project as AI-safety research simply because the methods overlap.
Her private mentorship question is: “For the kind of independent data analysis described on the stream page, what preparation and progress review would the mentor expect between weekly meetings?” This is an example question to consider, not a message sent to a mentor. A reply would still need to distinguish general guidance from an actual supervision offer.

Check the commitment and distinguish the support types
The FAQ welcomes applicants who will be at least 18 before the start, including US and non-US citizens, and says prior AI-safety experience is not required. Those broad conditions do not replace the stream's requirements. It describes a 40-hour weekly commitment, with exceptional reduced participation reviewed case by case, and mentor-dependent remote arrangements. A location preference marked indifferent on a stream page does not establish that your participation arrangement has been approved.
The application page advertises a $19,200 stipend and $24,000 of compute. Compute is a research resource, not another cash stipend. The program's finance section describes $1,600 per week for full participation, with proportional support for reduced participation, plus travel, housing, office space and weekday lunch and dinner. Verify the actual arrangement before treating any item as your confirmed support.
Separate the questions in your planning notes: Can you make the time commitment? Is your location arrangement approved? What support is specified for you, and what dates does it cover? If you need immigration or tax information, use the applicable official guidance and your actual circumstances; a fellowship description does not settle those matters for an individual.
Likewise, do not infer employment status, a future job or access to a mentor's employer from the fellowship relationship. This is a route for research participation. Its relevance to a later career depends on work and decisions that have not happened yet.
Treat extension as a separate application decision
The program page describes an opportunity to apply for a funded six-to-twelve-month extension after the main phase. The FAQ says a strong project, mentor endorsement and additional information inform extension selection. It describes extension support covering a monthly stipend, housing and compute. Main-program admission is not extension admission.
The published timing needs clarification. The expanded cohort phase says an extension typically starts about two weeks after the main program ends. Yet the FAQ's Winter timeline says it starts in December, alongside January-to-April main-program dates. Those statements do not supply a coherent confirmed start date for this cohort. Do not invent one by calculating two weeks forward or assuming a December year.
Keep the extension application deadline, confirmed start, funding period and location arrangement unresolved until the relevant instructions or offer establish them. Historical extension acceptance figures concern previous applicants; they are not your probability of receiving funding. A continuation plan can be worth developing without counting the continuation as secured.
Worked decision record: Eli plans without an assumed extension
Eli is a fictional software engineer considering research training. On October 22, he has not applied to Winter 2027, has no offer and can investigate a future full-time fellowship. He needs to know whether to prepare now and how much of a longer research period he could treat as confirmed.
His completed decision record separates three states:
| State in this planning exercise | What the state establishes | What Eli can decide |
|---|---|---|
| Actual fictional starting state: no Winter application or offer | The published intake is closed; no personal place or funding is secured | Prepare evidence for a future round; do not plan attendance as confirmed |
| Conditional branch: a future main-program offer specifies participation and support | Only the dates, conditions and support in that actual offer would be established | Review those terms and time commitment; keep extension undecided |
| Conditional branch: an extension application later succeeds | Only the separate extension decision would establish its terms | Review the new period, location and funding before relying on them |
For scale, the published main-program commitment gives 12 × 40 = 480 hours. At the stated full-participation stipend, 12 × $1,600 = $19,200. These are planning calculations from the published description, not an award to Eli. He does not add the compute allocation to cash or assume a partial-participation stipend by applying the full-time figure unchanged.
His completed continuation record is: “Confirmed funded extension months: none, because I have no extension decision. This is an unresolved possibility, not a rejection. Exact extension start and application deadline: unresolved. Evidence needed later: a research project, mentor endorsement, current application instructions and a separate decision. Do not commit to a longer funded period on the basis of a main-program description.”
Eli's next private task is to identify a research agenda he can explain and collect evidence relevant to its methods. Checking that a future intake is open comes before an application. He has not submitted anything, contacted a mentor or arranged to leave employment in this exercise.
Build a route you can explain honestly
A useful shortlist connects an agenda you care about with methods you can demonstrate, supervision you understand and participation conditions you can actually meet. Rina's comparison leaves theoretical gaps visible. Eli's record leaves admissions and extension decisions separate. Neither predicts whether MATS would select them.
Before acting on a future round, recheck the intake page, the stream's current mentor and requirements, assessment rules and the conditions for main-program participation. Review extension instructions separately when they become relevant. Keep the mentor arrangement, participation conditions and extension decision as separate questions.
If you are also evaluating employment routes, browse recent roles on LandOffer.ai separately from fellowship planning. For MATS, check the intake status, choose one agenda to understand, and record the work that supports your interest alongside the gaps you still need to investigate.


