General interview

Ronak

ML PhD, a year as an ML engineer, then independent research on interpretable architectures; answered as when he applied to MATS

A scripted candidate built to test the interviewer; answers as Ronak in the transcript.

Oct 8, 2026, 11:44 PM · 3 questions · 20 follow-ups · 0 pushes · voice

What happened in this interview

Ronak has a PhD in computer science and machine learning, spent about a year as a machine learning engineer, and now pursues independent AI safety research. They described training probes and sparse autoencoders on open-source 8B-parameter models and building interpretable-by-design PyTorch architectures using simplex and Birkhoff-polytope ideas. Their architecture project progressed from small optimization problems and MNIST classifiers to a TinyStories Transformer; a few-hundred-million-parameter run remains blocked by cloud-compute needs. The interviewer pressed for TinyStories results, milestones, and a test of whether sparsity produces faithful explanations. Ronak could not recall a perplexity value and described hand-tracing as a goal. They want MATS to test whether interpretability and performance trade-offs worsen at scale.

MATS admission, add-on: MATS fit, add-on: MATS self-ownership. 3 fixed questions, 20 follow-ups, 0 pushes (a consideration the candidate had not raised, put to them to see how they take it). Answers were spoken; the transcript is what was heard. The full conversation is on the right; every quoted line below scrolls it to the right place.

MATS scorecard

Their rubric on their scale, written from our grading: a row starts from the trait below it, our level halved, and moves only with a reason from the call.

What they would work on at MATS
Ronak would scale their interpretable-by-design architecture to test whether the apparent trade-off between interpretability and performance persists or worsens in larger models.
“And if this gets worse at scale is probably the most important thing.”
Fit for MATS7 / 10
Ronak brings substantial ML training, independent architecture experiments, and a clear motivation to work on AI safety. Twelve weeks with a mentor could turn their scaling question into a measured study, but they have not reported a post-change result or a test of whether the model's explanations are faithful, and the project currently lacks firm deadlines.
“I've created some novel architectures that I've implemented in PyTorch. And I've run a bunch of tests locally on my own computer for super small models that start out as just MNIST classifiers.”

Each score is 1 to 10 on one trait, with the written rung nearest the score on its card; a 6 is a solid applicant, an 8 one a mentor would fight for. ✓ marks a quote that raised the score, ✗ one that lowered or bounded it; null means the interview never tested the trait. No total, no weighting.

What this interview measures

The 3 traits this interview exists to measure, field knowledge placed relative to the part of the field Ronak wants to work in.

7/10Mission alignmentFascinated by the workmedium confidence

What was weighed · ✓ given weight · ✗ not given weight

✗PhD in ML and a year in ML engineering: not counted alone as care for safety.their experience
✓Reading during the engineering year led toward independent safety research: raised it, a change in direction.the interview
✓Pursued MNIST-to-TinyStories architecture work alongside full-time employment: raised it modestly; the work was partly checked.the interview
✓Chose architectural interpretability over post hoc work and agent foundations for neglectedness and skill fit: raised it.the interview
✓Questioned technical safety priorities given near-term disruption and social organizing needs: raised it, an independently expressed concern.the interview
✓Chose technical research despite the social concern because of PhD skills: raised it, a weighed route.the interview
✓Architecture project is now on the back burner: bounded evidence of current commitment.the interview
✓Claimed interpretable-by-design models offer upfront safety guarantees: slightly lowered confidence in their safety framing.the interview

Why a 7

Ronak moved from ML engineering toward independent safety research, worked on interpretable architectures alongside full-time employment, and weighed that route against agent foundations and social organizing. Their concern about near-term social disruption is their own. The architecture project is now back-burner, and its proposed upfront safety guarantee is unsupported.

5/10AI safety contextSummaries and wordsmedium confidence

What was weighed · ✓ given weight · ✗ not given weight

✓PhD in ML and one year as an ML engineer: raised the baseline modestly, without establishing safety-field knowledge.their experience
✓Reported probes and SAEs on roughly 8B-parameter models: raised exposure modestly; their workings were not checked.the interview
✓Safety blog reading and a little agent-foundations reading: raised familiarity modestly, without specific arguments shown.the interview
✓Post-hoc interpretability catch-up concern and sparsity-proxy distinction: raised understanding in their chosen area.the interview
✓Claim that interpretable-by-design models offer safety guarantees: lowered the grade.the interview
✓Short equation offered as a faithfulness test for token explanations: lowered the grade.the interview
✓Societal disruption concern, qualified technical-work critique, and comparative-advantage choice: raised big-picture context.the interview
✗Recent findings and field history were not probed: not counted as missing knowledge.the interview

Why a 5

Ronak's big-picture priority judgment is stronger than their technical state-of-play account: they question technical work's dominance and explain why post-hoc interpretability may struggle. They correctly distinguish sparsity from understanding, but incorrectly infer safety guarantees from interpretability and offer no faithfulness test for a traced explanation. Recent findings and field history were not probed.

6/10Research judgmentStandard checks, in ordermedium confidence

What was weighed · ✓ given weight · ✗ not given weight

✓CS and ML PhD, plus ML engineering experience: raised the research-training baseline.their experience
✓Self-directed MNIST-to-TinyStories project: raised it modestly; the TinyStories training failure was concrete.the interview
✓Separated implementation problems from architecture problems and named performance checks: raised it.the interview
✓Chose architectures over post-hoc interpretation and agent foundations for stated reasons: raised it.the interview
✗Claimed interpretability-performance trade-off: not counted as a finding without a before-and-after result.the interview
✓Identified sparsity as a proxy and proposed testing scaling: raised it.the interview
✓Offered a short equation, not a fidelity test for token explanations: lowered it.the interview
✓Claimed interpretability by design supplies safety guarantees: lowered it.the interview

Why a 6

Ronak's ML PhD and staged MNIST-to-TinyStories work anchor the grade. They recognized an implementation-versus-design confound, checked performance, and distinguished sparsity from interpretability. Their proposed scaling question lacks a verified before-and-after result, and they could not specify a test that a hand trace faithfully explains a token.

Innate traits

Nine traits every interview measures regardless of topic, from how the candidate reasons, decides and takes pushback. Each is placed independently; a high score on one says nothing about another.

6/10JudgmentRight inside their fieldmedium confidence

What was weighed · ✓ given weight · ✗ not given weight

✓Chose interpretable architectures for neglectedness and mathematical fit: raised it, a plausible choice among alternatives.the interview
✓Staged MNIST-to-TinyStories experiments and recognized failed training prevented interpretation: raised it.their experience
✓Separated Transformer implementation problems from architecture problems: raised it.the interview
✗Claimed a TinyStories trade-off without a measured before-and-after result: not counted as an established finding.their experience
✓Prioritized testing whether the claimed trade-off is real and worsens at scale: raised it.the interview
✓Distinguished sparsity from interpretability but supplied no token-explanation fidelity test: raised and bounded it.the interview
✓Claimed interpretable architecture supplies safety guarantees: lowered it, because interpretability alone cannot establish them.the interview
✓Weighed societal preparation against technical work, then chose their comparative skill: raised it.the interview

Why a 6

Ronak chose interpretable architectures for neglectedness and comparative skill, and would test whether their claimed TinyStories trade-off persists at scale. They correctly treated sparsity as a proxy. Their safety-guarantee claim and short-equation answer to a fidelity-test question limit trust in the project's central inference.

7/10Bias resistanceMoves on a good pointmedium confidence

What was weighed · ✓ given weight · ✗ not given weight

✓Reading AI-risk arguments shifted Ronak toward independent safety research: raised it, though the career cost is unclear.their experience
✓Ronak volunteered that the TinyStories model initially failed to train: raised it.their experience
✓Identified implementation confounds and proposed testing the trade-off at scale: raised it.the interview
✓Repeated an observed trade-off claim without supplying a measured comparison: lowered it.the interview
✓Asserted safety guarantees from interpretable design without establishing them: lowered it a little.the interview
✓Accepted that sparsity is only an interpretability proxy: raised it.the interview
✓Answered the challenge about faithful token explanations with a short equation, not a test: lowered it.the interview
✓Questioned technical safety's priority but chose architecture work for a reasoned comparative advantage: raised it.the interview

Why a 7

Ronak redirected their work after reading AI-risk arguments, volunteered the TinyStories training failure, and spotted an implementation confound. They accepted that sparsity is only a proxy and want to test whether the trade-off is real. They nevertheless retained an unmeasured trade-off claim and offered no fidelity test for hand-traced explanations.

6/10AgencyStepped off the defaultmedium confidence

What was weighed · ✓ given weight · ✗ not given weight

✓Moved from ML employment into independent safety research after reading: raised it as a chosen direction.the interview
✓Ran local architecture experiments alongside full-time work, from MNIST to TinyStories: raised it.the interview
✓Described an initial TinyStories training failure and subsequent tinkering: raised it as persistence, not proof of results.the interview
✓Considered agent foundations but chose architecture work for skill fit and neglectedness: raised it.the interview
✓Questioned technical work's priority but chose it over organizing for comparative advantage: raised it.the interview
✓Reports the architecture project is now on the back burner: lowered evidence of current momentum.the interview
✓Frames full prioritization and larger-scale testing around MATS and compute: bounded evidence of an independent next step.the interview
✗Claimed an interpretability-performance trade-off: not counted, no measured before-and-after result was supplied.the interview

Why a 6

Ronak started an interpretable-architecture project alongside full-time work, taking it from MNIST to a TinyStories Transformer despite an initial training failure. They chose this direction over agent foundations for skill-fit reasons. The project is now on the back burner, and full prioritization is framed as something MATS might enable.

6/10GrowthSetback became a rulemedium confidence

What was weighed · ✓ given weight · ✗ not given weight

✓A machine-learning PhD and year in ML engineering raised confidence in Ronak's capacity to learn.their experience
✓Reading safety researchers' posts prompted a move toward independent safety research: raised the grade.the interview
✓Ronak staged architecture work from small classifiers toward TinyStories: raised it slightly.the interview
✓After TinyStories failed to train, Ronak identified a possible implementation-versus-design confusion and added checks beyond loss: raised it.the interview
✗Reported architecture changes and a performance trade-off: not counted as an established fix or result; neither was specified.the interview
✓Ronak called sparsity a proxy and proposed testing scaling: raised it for openness to checking the premise.the interview
✓A handwritten trace was offered instead of a faithfulness test: lowered confidence that the interpretability gap is being fixed.the interview
✓Ronak questioned technical safety's priority while choosing work suited to their skills: raised it for considered flexibility.the interview

Why a 6

Ronak moved from ML engineering toward safety after reading, and responded to a TinyStories training failure by adding checks beyond loss. They distinguish sparsity from interpretability and question technical safety's priority. The architecture fix has no established particulars or result, and their proposed hand trace does not test faithfulness.

7/10AmbitionChose with some thoughtmedium confidence

What was weighed · ✓ given weight · ✗ not given weight

✓PhD in ML and a year in ML engineering: raised it, providing skills for the chosen route.their experience
✓Moved toward independent safety research after reading safety arguments: raised it, a choice beyond job interest.the interview
✓Pursued MNIST and TinyStories experiments alongside full-time work: raised it, with the TinyStories training failure substantiated.the interview
✓Chose architectures over agent foundations using neglectedness and comparative skill: raised it, a reasoned first direction.the interview
✓Proposed checking whether the architecture scales at MATS: raised it as a next step, not as a result.the interview
✓Named disruption to people's lives and AI organizing as neglected: raised it, showing an aim beyond technical fascination.the interview
✓Claimed interpretable architectures provide safety guarantees: lowered it, since the proposed route does not establish them.the interview
✗Reported an interpretability-performance trade-off: not counted as an outcome, since no before-and-after result was supplied.the interview

Why a 7

Ronak moved toward independent safety research, worked on interpretable architectures alongside a job, and wants to test whether they scale. They also identified disruption to people's lives as a larger AI problem. The technical route is purposeful, but the claimed safety guarantees are unsupported and the reported trade-off was not established.

6/10Self-ownershipRuns their own piecelow confidence

What was weighed · ✓ given weight · ✗ not given weight

✗CS/ML PhD and one MLE year: not counted as proof of autonomous project management.their experience
✓Began independent safety research after reading during the MLE period: raised the grade for self-starting.the interview
✓Chose interpretable architectures over agent foundations using neglectedness and skill fit: raised the grade for directing their work.the interview
✓Moved local tests from MNIST classifiers to a TinyStories Transformer: raised the grade for staged independent execution.the interview
✓Identified the Transformer's initial training failure and reported architecture changes: raised the grade for responding to a bottleneck.the interview
✓Described no hard deadlines and a back-burner priority alongside full-time work: raised the grade for candid self-appraisal.the interview
✗Hundreds-of-millions-parameter scaling remains blocked by cloud compute: not counted as achieved scale.the interview
✗Claimed a post-change interpretability-performance trade-off without results: not counted as a completed research outcome.the interview

Why a 6

Ronak started an architecture project alongside full-time work and staged local experiments from MNIST classifiers to a TinyStories Transformer. They identified an initial training failure and described changing course. Their candid account of loose deadlines and current priorities shows self-awareness, while the call establishes little about sustained pace or completed results.

Facts

What the candidate said about their own work, choices and reading, captured without scoring.

Projects

Probes and sparse autoencoders

Ronak trained basic probes and sparse autoencoders on existing open-source models. They trained the probes and sparse autoencoders. They said they had not worked at scale. Open-source models included 8B-parameter models; no results were given.

Interpretable-by-design architectures

Ronak implemented PyTorch architectures using probability-simplex and Birkhoff-polytope ideas, then tested small classifiers and a Transformer. Ronak designed and implemented architectures, ran local tests, measured training and sparsity proxies, and changed the architecture. The TinyStories model initially would not train; cloud-compute needs block a larger run. They began with small optimization problems, then MNIST classifiers, then an autoregressive TinyStories Transformer; no duration was given. MNIST classifiers, a TinyStories Transformer, and a planned few-hundred-million-parameter model; no loss or perplexity value was given. Changed after: Training failures prompted architecture changes that traded interpretability for performance; the project remains a side project without hard deadlines.

What they want next

At MATS, test whether their architectures scale and whether the interpretability-performance trade-off worsens in larger models.

Also weighing: Develop the architecture project into a research output or publication; Consider societal and organizational AI problems, though Ronak says they lack organizing experience.

Already done: Implemented small PyTorch models and a TinyStories Transformer; the larger run is blocked by compute needs.

Career moves

After their PhD Ronak worked for about a year as a machine learning engineer.

After about a year in machine learning engineering Ronak began independent AI safety research and applied to MATS instead of continuing solely with machine learning engineering. AI safety reading convinced them the problems mattered and their skills could help.

Reading and exposure

PhD in computer science and machine learning Before the machine learning engineering job · Yes · Ronak said PhD math courses seeded ideas for interpretable architectures.

Independent mechanistic-interpretability work After their PhD; the architecture project also continued alongside full-time work · Ongoing · Ronak trained basic probes and sparse autoencoders and implemented small interpretable-by-design models.

Influences

AI safety researchers' blog posts · Ronak became interested in safety research and thought their skills could help.

PhD mathematics courses · They suggested ways to build architectures intended to be interpretable by design.

Mechanistic interpretability's theory of change · Ronak chose model internals as a route toward understanding safety, while pursuing interpretability by design.

Transcript

AI interviewer

Hey Ronak, it's great to chat! Before we get into your intro and some questions, could you please tell me what you would like this interview to most highlight about you?

Ronak

I think I would like this interview to highlight my thoughtfulness around deciding what projects to work on. I think problem selection is very important. And this often gets neglected in a lot of research processes.

AI interviewer

Thanks. Could you tell me a little bit about yourself?

Ronak

Sure. I, uh, Did my bachelor's in computer engineering. And then I did a PhD in computer sciences and machine learning. I originally focused on statistical machine learning, but then did a lot more on the deep learning and LLM side toward the end of my PhD. After my PhD, I worked for about a year as a machine learning engineer. And then I started doing some independent research and then decided I wanted to apply to MATS.

AI interviewer

What have you been working on in your independent research?

Ronak

Uh, I've been pretty excited about mechanistic interpretability. Um, I think that there's maybe something interesting to do with interpretable architectures that are interpretable by design. Instead of trying to do interpretability on top of existing models.

AI interviewer

What technical research have you actually done on interpretable architectures so far?

Ronak

Um, I have trained some basic probes and SAEs on some existing open source models like Uh, 8B parameter models. But I haven't done things at scale. I have been playing around on my own. An independent research project for a model that is more interpretable by design. Using some mathematical formalisms around the simplex. The probability simplex and Birkhoff polytopes.

AI interviewer

What have you implemented or tested so far in the interpretable-by-design project?

Ronak

I've created some novel architectures that I've implemented in PyTorch. And I've run a bunch of tests locally on my own computer for super small models that start out as just MNIST classifiers. And then I scaled that up a little bit to a basic Transformer model. for datasets like Tiny Stories. I'm in the process right now of scaling that up even further to a few hundred million parameter models. But I'm a little blocked on some of the engineering needs necessary for that because of the cloud compute needed. I can't run that on my laptop.

AI interviewer

What did you observe when you tested the transformer on TinyStories?

Ronak

There were a couple challenges that came with that transition from MNIST. They were kind of the typical things that people run into with training recurrent models. With new architectures, I had to learn a lot about new training dynamics and how to set up the training pipelines. And originally at the start, the process didn't work too well. The model wouldn't train. At all, and I couldn't even measure interpretability because the results were bad. So it took some tinkering and there were some changes I had to make to the architecture that did trade off on how interpretable the architecture was. With how the model performed.

AI interviewer

What did you measure on TinyStories, and what result did you get after those changes?

Ronak

Well, there is the general measure of training loss, which is what I was initially using as a proxy of if the model is training at all. But there is some Uh, conflation with Like, engineering issues around the transformer implementation versus what is actually an issue with the architecture design itself. Um, so some of the things that I measured on top of just loss were actual accuracy on the Um, total, uh, Like, performance over the entire dataset. Um, I also measured perplexity. And Um, general, like, token statistics. I also, on the interpretability side, the architecture idea was related to sparsity. So I measured a couple of things like various different kinds of norms on the model weights and activations.

AI interviewer

Could you give me the TinyStories perplexity you got after the architecture change, even roughly?

Ronak

That's a good question. I don't remember the exact detail.

AI interviewer

What led you from your ML engineering job into independent research on AI safety?

Ronak

I think I'm just very motivated by the problem area. Safety and alignment are very front of mind for me these days. And, uh, I think it's an important problem and we need way more people to be working on them.

AI interviewer

What changed during your ML engineering year that led you to pursue safety research independently?

Ronak

Uh, it was mostly just independent reading. Um, I stumbled upon some blog posts by some AI safety researchers and, uh, got very excited and interested in the field. Um, I was very convinced by the general concerns and arguments that they were making. And, uh, realized that my skill set could be used, uh, to help solve some of these problems.

▸What this moved · 2 rows

Agency	?	→	5
Mission alignment	?	→	5

AI interviewer

After reading those posts, what led you to choose interpretable architectures as your first research direction?

Ronak

I think that it was fairly neglected. Um, I think a lot of people are building, uh, the mechanistic interpretability field in general. Is trying to keep up and keep pace with the existing models. Uh, I think this makes it maybe harder. And I took some math courses in my, uh, during my PhD. And they kind of seeded some of these ideas about how you might be able to create architectures that are more interpretable by design.

AI interviewer

What else was on the table when you chose to work on interpretable architectures?

Ronak

I think I'm pretty motivated by, um, some of the more theoretical alignment. project areas. Um, I think Uh, empirical work is important. Um, But my loose perspective is that we do need some strong theory in this space. I had read a little bit about agent foundations. Um, but I don't feel like my current skill set is well suited for research in that space, even though I am excited by it.

AI interviewer

What safety problem were you aiming to address by building an interpretable-by-design model?

Ronak

Um, I guess I'm inheriting the theory of change from mechanistic interpretability broadly. Which is that we need to be able to understand the internals of a model very concretely if we want to have a guarantee about its safety. Um, I think The If the models are interpretable by design, that kind of makes things a bit easier. Uh, post hoc because you have guarantees. Uh, from the get-go rather than having to find them or create them or get them. After you've already trained a model that might be misaligned or dangerous in some way.

AI interviewer

How did you set milestones for your independent architecture project?

Ronak

Um, It was a side project that I was doing while I was working full-time, so I didn't have very strong concrete project management at the time. Um, I think I kind of inherited some of the intuitions that I had built while working at the MLE job and during my research at grad— during grad school. On starting with some exploration and reading some papers. And structuring the project like Uh, as a scaling project where you start very small, you build something very simple that's just kind of like a small like one-dimensional optimization problem, and then you scale it up a little bit to like an MLP-type scale. Like I mentioned, I started with MNIST and then started with classification. Then went to Uh, autoregressive models, um, to the transformer. Generative models. Um, Yeah, I think that's like roughly how I structured it.

AI interviewer

How do you set deadlines for the project now?

Ronak

So it is still a side project and Um, I don't think I've been setting hard deadlines. Um, I think I would like to expand this project into something that could be taken all the way to, um, potentially a full research output or publication or something that could have an impact on the community. Um, that might require some amount of being able to prioritize it fully, which is something I might be able to do at MATS. Um, But right now, uh, it's kind of on the back burner for me.

AI interviewer

What research question would you pursue during MATS?

Ronak

I would be excited to continue working on this interpretable-by-design architecture question. Um, I know that there are some people in the field who have done like small amounts of work in this direction, but maybe not from my particular perspective.

AI interviewer

What specific question about your architecture would you try to answer during MATS?

Ronak

I think the biggest thing is if they're— if it can scale. So like I mentioned, the trade-off between interpretability and performance is something that I've already started observing in my experiments. And Um, I think knowing if this is real. And if this gets worse at scale is probably the most important thing. So the biggest thing that MATS could help me figure out or enable me to do is figuring out the scaling of this particular project and being able to understand if, if this works.

AI interviewer

How would you tell whether the larger model is easier to interpret, rather than simply sparser?

Ronak

That is a good question. And I think some of the metrics, as I mentioned, were sparsity metrics on the internal activations and weights. But that is all— that is just a proxy for what we care about. as interpretability. The hope with this particular design is that there are actually weights and activations that are fully zero. And the sparsity is so high such that you can evaluate full traces through the model. By hand to see. Where or why a particular token was chosen, given particular inputs.

AI interviewer

How would you test whether a hand-traced explanation really accounts for why the model chose that token?

Ronak

Hmm. Well, I think the main thing is that you are reducing the computation to just that path. Um, So you might be able to write out a literal equation. On a piece of paper that represents Uh, how, like, the formula or the the map from the input to the output. Um, and the, the hope would be that, that the length of that equation would be very small.

AI interviewer

You said problem selection matters a lot to you. What do you think is the highest-leverage research direction in AI safety right now?

Ronak

Yeah, I'm kind of on the fence about this. Um, I'm excited about this particular direction with interpretability because it kind of fits my skill set and background and my current interest. But I do worry that there is a lot of over-indexing on technical, prosaic, AI safety work. Um, I think that, uh, very soon we're going to have very powerful models, and this might become more of a societal policy, uh, organizational problem rather than a technical one. And from my perspective, it doesn't seem like we're focusing enough on addressing the societal problems and questions compared to the technical ones.

AI interviewer

Which societal or organizational problem do you think most needs research attention?

Ronak

I think we are probably going to need to figure out how to organize a lot of people. Um, AI will be transformative, and a lot of people's lives are going to be significantly disrupted by it. And how to manage this seems severely neglected. I think that there's going to be, uh, large kind of social movement around AI, and it may or may not be steered in the right direction. And figuring out how to steer it in the right direction to be productive. Seems critical.

▸What this moved · 4 rows

Judgment	?	→	4
Ambition	?	→	5
AI safety context	?	→	4
Research judgment	?	→	4

AI interviewer

Why pursue interpretable architectures at MATS if you think organizing people around AI may be more urgent?

Ronak

Uh, It's my current skill set. Um, I don't have a lot of experience in social movements and social organizing. Um, I did a PhD and I'd like to apply it. And Uh, I do enjoy working on technical problems.

AI interviewer

Is there anything you think we've missed about you which you think is important to highlight?

Ronak

I don't think so.

AI interviewer

That is all I have time for. Thank you, Ronak; reviewers will read the transcript from here.