The serious people are in War Mode
They worked incredibly hard and were endlessly motivated to do more work [...]
Most kinda-fun-to-do things become less interesting when you’re in war mode. [...]
I mean, on the one hand, I'm glad that people at the labs are taking this seriously, and would feel quite worried if they were mostly LARPing safety work. On the other hand, I think that the attitude you describe does not sound sustainable on the scale of years (this is applicable even if AGI is coming in 2027-2030!). More pragmatically, I also think it seems quite likely that spending too much time in war mode might make it hard to analyse whether the thing you are working super hard on is actually a good thing to be working on. This is particularly important if you're in a safety role at a leading AI developer.
the attitude you describe
It's quite possible that I'm misinterpreting or unintentionally cherry-picking their attitude (I never worked full-time with multiple frontier lab employees in person, and those I did work with I only did so briefly), but I would be somewhat surprised.
does not sound sustainable on the scale of years
I agree, but reading your comment makes me want to read up about burnout amongst people working in order to support an (actual) war effort.
Great post, I like the level of helpful detail. This advice seems pretty reasonable to me, thanks for writing!
dataset generation: Claude is very bad at getting another LLM to generate coherent, high quality datasets and then auditing the dataset for quality. Claude will claim a dataset is good, but manual checking reveals the dataset to be not strongly exhibiting the trait we care about, or it is too “meta” (e.g. the response talks about exhibiting the trait instead of actually just exhibiting the trait, and Claude doesn’t pick this up), or Claude will do regex to check for plausible words associated with the trait of interest but not actually read the responses. This is made worse because Claude will talk as though it had done a thorough audit even though it had vaguely glanced in the right direction. “double blind” reviews and scientific thinking: Claude doesn’t have a good practical understanding of scientific thinking in a way that’s hard to describe. Often, I asked it to spawn sub-agents to “blindly” review two datasets (one exhibiting some behaviour, one not exhibiting some behaviour) and ask the subagent to guess what the behaviour was. Claude would frequently put the name of the behaviour in the file name and ask the subagent “what behaviour is systematically expressed in examples-of-sycophancy.jsonl” (I’m not exaggerating). Claude will very happily talk about the mechanisms of blind review, but does not reliably implement them, and frequently reports that it has done a blind review despite this. There are other ways in which Claude doesn’t “think scientifically”, and will very happily cherry-pick results, make claims that are not backed up by the experimental results, fail to find (fairly simple) flaws in an experimental setup, disregard negative evidence as noise, put big green emojis next to trivially-true things (“dataset has 500 rows✅”) and hide or not equally emphasise problems (“dataset only containstags, no responses”).
How similar is it to Greenblatt's Current AIs who seem pretty misaligned? Did the issues persist when trying to offload the tasks to GPT (via Codex?), Gemini or Grok?
I think this is very similar to Greenblatt's findings, and I largely agree with how he describes the LLMs. I didn't try offload the tasks to other LLMs, I probably should have but I only really saw this as a consistent problem (and not once-off flukes) quite late in MATS. I've now got codex setup and hope to setup a way for claude to ask codex for review or vice versa.
I couldn’t find a recent write-up from a MATS alum about what attending MATS was like, so this is the thing that I wish I had. I attended MATS from January to March 2026, on Team Shard with Alex Turner and Alex Cloud. It was a great time! Applications for MATS are basically on a rolling basis nowadays, and I can strongly recommend applying (to multiple streams) even if you think you’re not a great match.
With that being said, there’s a lot I wish I knew going into MATS, so here’s a brain-dump of thoughts. It’s not extremely polished, but I expect it’ll be useful nonetheless (none of this is endorsed by MATS, just my thoughts):
Work ethic
I think most mentees were working 10-12, sometimes 14 hours a day Mon-Fri, and probably 2-8 hours on Saturday and Sunday, often going out on some adventure or party on the weekend. Exactly which hours people worked varied wildly. I usually worked 8:30am/9am to 11pm/midnight, with breaks during the day, others worked from midday into the early hours of the morning. This was surprisingly sustainable (IMO); MATS puts a lot of effort into removing all other blockers that you normally have in life so that you don’t have to do other things.
Use more compute
MATS was a lot about learning to go fast, and I kept on thinking I had a good strategy and then learning this wasn’t enough.
Research requires a lot of compute
We went through a lot of compute (as measured in dollars), relative to what I thought we would need. Or at least, I previously had the assumption that I could have, without funding, paid for my own experiments if I didn’t get into MATS (just using my salary). I now think this is largely not possible, and that you need funding to do serious work. Although I think we could have been a lot more stringent, I would love to see research papers declare how much compute went into each paper (and not just to create the results, but also for all the initial experiments that lead up to the headline results). MATS budgets $1k per fellow per week (edit: MATS 10 will be 2k/week/fellow, to encourage more ambitious projects), and my teammate and I ended up requesting 3k on top of that. A few fellows who were doing RL experiments ended up in the \40k+ range for 3 months of experimentation.
Applying for jobs during MATS (don’t do it)
In general, I think this is not worth doing. It does depend, but if MATS is your first introduction to the AI safety research space (as it was for me), then I don’t think it’s worth trying to find a job during the MATS main program. Applying for jobs takes a lot of time, and going through the tech interview rounds also takes a lot of preparation. All of this eats into the time you have to work on your project, spending time on your project to get a good reference from your mentor is far more valuable than spending that time applying to jobs.
If you’re already somewhat established and your mentor’s reference doesn’t significantly improve your CV, then job-hunting during MATS is probably not bad (although I’d question why you’re doing MATS in this case?). If your CV would be relatively weak without your mentor’s endorsement (e.g. I had no first-author papers at highly-rated conferences, no previous work with any high-profile researchers, no PhD, no formal research experience, no Ivy League background) then working as hard as you can during MATS is probably the most highly-leveraged time you will ever have.
What’s more, the MATS extension is the perfect time to look for a job. This is explicitly encouraged, and MATS is very happy if you end the extension early because you got full-time employment. Reading between the lines, I think the extension is structured the way it is to prevent fellows from finishing MATS, going out into the world, realising they need to pay rent, and then taking the first well-paying capabilities job they come across. The extension gives you the breathing room to apply to multiple places and make a well-thought-out decision about what’s the best place to work, not just what place will pay the bills.
The serious people are in War Mode
I had a feeling about many of the people at MATS (especially the Anthropic mentors), and I struggled to put this into words. They worked incredibly hard and were endlessly motivated to do more work, but it wasn’t driven by money or stocks or promotions. I’ve now started to call this drive war mode, because it’s the best way I know to describe it. Most kinda-fun-to-do things become less interesting when you’re in war mode. It’s partially a drive to get more work done, but it’s also a reduced enthusiasm for anything else. Ambling about a book store or spending time with people you used to get along with becomes less interesting if you’re in war mode. I’m not sure if everyone would agree with my characterisation, but it’s the closest I’ve gotten to describing the work ethic.
Do you feel the AGI?
‘Cause I do now. I thought I did, but now I really feel it. It’s easy to work hard when you’re pretty sure things will be bad-by-default. It’s incredibly motivating.
Burn rate, efficiency, and decisions
One thing that was incredibly effective about MATS was that 1. it removed many many decisions from my daily life (what to cook for dinner, when to go get groceries, how to get to the gym, where to live, when to go get office supplies, scheduling time with friends) and 2. it replaced them with answers that were significantly more expensive but also more productive than my default. Food is an obvious example, MATS does food through a delivery service (Forkable) and we were given $25 to spend for each of lunch and dinner. Getting food made and delivered to your door is much more expensive than preparing it yourself, and given the choice I would have prepared it myself. But MATS doesn’t give you that option, it repeatedly forces the high-burn-rate maximum-productivity option upon you, and then asks you to do great work with the extra flexibility you now have.
This effect is true in many other smaller ways. The Ops team is unbelievably willing to do irritating things to make your life easier or to let you focus on the research. I frequently felt like I was abusing their help (although the MATS team assured me that they’d say if I was crossing a line).
“insider” information
Coming from a country where I can probably list the AI safety people on two hands, it was shocking how many times we were helped by knowing someone in MATS or having access to “insider” knowledge about what different research labs were working on. None of this is secret, it’s just not public and probably not available online. e.g. “don’t try that idea, my friend from MIT is working on it and about to publish their results” or “oh my college buddy has not-yet-published datasets that’ll be really useful, let me message them to see if we can use them”. This was incredibly useful! I would have gone down two independent dead-ends if I hadn’t been at MATS, because the ideas were good but they were already being pursued by other people, and I only realised this when I was casually chit-chatting at dinner with some other MATS fellows. I just mourn previous me who thought I could do good research without a good network of researchers.
Names & Faces
See this page, but basically I strongly recommend using Anki for ~2 weeks before MATS to try and memorise the names & faces of all the fellows you’ll be in-person with.
Fellows
I made it my goal to know every fellow who was in-person in Berkeley, and I think I succeeded? Or at least, after week 2 I stopped seeing new faces around and by week 3 I think I had everyone’s names memorised and had had a small chat with just about everyone. This was really good, strongly recommend. Even if you’re not a “social person”, there’s a big difference between socialising all the time and knowing the fellows well enough that you can spot them in a crowded room.
Useful tools
Below are some tools you should know exist, they might be perfect for the problem you have.
Use more Claudes
Build nice helper utilities for yourself
Claude makes it very easy to create nice helper scripts for yourself, so do so!
viewscript that visualised .jsonl chat-completion files in the terminal in colour and with nice formatting, this was great for getting me to actually review the data that I was using for training.plot.pyscript which Claude would edit, and then this script emitted aplots.htmlfile which I’d open in my browser. The HTML contained plotly plots so they were interactive (amazing!) and if I wanted changes I’d just tell Claude what to do. This was so much better than trying to get Claude to work nicely with a jupyter notebook (and why bother with a notebook if you’re not going to review the plotting code every time?)MATS-mentee-mentor dynamics
Very different to how I initially thought they were.
Working with your mentors
They’ve got full-time jobs, but generally are doing MATS because they 1. like seeing research get done that they’d otherwise not be able to do and/or 2. like giving mentorship and working with new researchers. That being said, it’s very worthwhile to figure out how your mentor likes to work, and then take advantage of that. Often your mentor will be unavailable by default (they’re very busy), with patches of free time to answer questions on Slack + a weekly project update. So if you get a message on Slack, answer quickly! Your mentor probably has some time and it’s incredibly valuable to get some feedback midweek.
Research managers
Amazing! very surprisingly helpful. It’s probably better to model the RMs as life coaches with specialisations in research rather than a regular manager, although they do have other (important) roles. Bryce was the RM for Team Shard.
Ops requests
This is basically a form you can fill out to fix ~any issue you’re having that’s somehow related to the well-running operations of MATS. Want different snacks? Ops request. Want a different monitor? Ops request. Have some feedback that doesn’t fit anywhere else? Are you sure it doesn’t fit anywhere else? Are you really sure (there are a lot of feedback forms)? Ops request. Found a leak? Ops request. Is the printer out of ink? Ops request. Ops was incredibly on the ball. It was magical. I didn’t feel like I was worth this much effort from them, but it significantly improved my experience at MATS to know that Ops was very responsive and very capable.
Non-MATS events
There were several of these, and they were fun. Most of the MATS fellows tended to huddle together, which is easy to start doing and easy to keep doing, but I’d strongly recommend against it. Generally these things were fun, kinda interesting, I’m glad I went, but they never led to anything crazy or significant (for me at least, I’m sure crazy things happened). In general, the opportunity cost for not working on your project is extremely high. It of course can be fun, but if you’re looking at events for the networking/job prospects, it’s probably better on the margin to work more on your project.
Team Shard
Alex Turner (Google DeepMind) and Alex Cloud (Anthropic) were my mentors, and this team is a bit unusual compared to other teams at MATS. We spent ~5 of the 12 weeks of MATS “just exploring”, doing small de-risking experiments, and trying to find a project we were excited about. This was really cool! Team Shard spends a lot of time finding a good idea, and we were explicitly encouraged to pursue ideas we thought of. But the default for most teams (AFAICT) is that you spend 0-2 weeks looking at one or two projects and then get to work. Team Shard also (AFAICT) spends an unusual amount of time doing “team things” like hiking, workouts, chatting together, brainstorming ideas, talking with other mentees, and generally having a greater sense of “being a team” vs other streams.
Weekly updates
I was part of Team Shard, and we had weekly progress updates every Thursday morning, followed by team lunch and other activities. Basically the entire week revolved around Thursdays. The other team shard mentees sometimes reflected that Friday felt like a weird limbo day because everything else was just a build up to Thursday. Thursday was where experiments went to die and then be reborn again. Thursday was incredibly valuable in terms of getting precise mentor feedback, but it was often a little soul-crushing to realise that I had spent a full day on something that was doomed from the start. Very educational! But the correct mindset is easy to learn badly. We’d get just under an hour to present the experiments from last week, discuss the results, and brainstorm plans for the following week. There’s also a lot of low-level encouragement to move faster and to be more ambitious. I thought I was fast and ambitious, and then the Alexes showed me that I could be significantly faster and more ambitious.
My teammate and I did these via google slides, and largely followed the advice here. Some tips are harder to do, but a couple that are definitely worthwhile and easy to implement:
You should also probably have an appendix of example prompts/training data/etc.
Keep a log of your mistakes
I didn’t do this and mildly regret it. There were some repeated mistakes I made in setting up experiments, and each time I thought that I wouldn’t make that error again. I don’t think this cost me a lot, but it definitely did mean that I had to re-run experiments which were otherwise solid.
My running-experiments setup
Before MATS I had never used a GPU, and I went through several iterations getting a decent setup before settling into something I was happy with. This took a while, because I wasn’t aware of how good the setup could be, so I kept on thinking “oh this is quite fast and efficient” and then finding a better way to run lots of experiments. I’m not going to say how to setup an account or launch a GPU, most of the GPU providers have a good first-time welcome flow and this shouldn’t be too hard. What I will describe is the details of my setup, so hopefully you can steal the parts you like and improve on the parts you don’t.
¯\_(ツ)_/¯. Some tools: zsh,iTerm2, macOS,ripgrep,uv,fd,bat,atuin,neogit,fzf,delta, my dotfiles are here, get your agent to have a look.examples-of-sycophancy.jsonl” (I’m not exaggerating). Claude will very happily talk about the mechanisms of blind review, but does not reliably implement them, and frequently reports that it has done a blind review despite this. There are other ways in which Claude doesn’t “think scientifically”, and will very happily cherry-pick results, make claims that are not backed up by the experimental results, fail to find (fairly simple) flaws in an experimental setup, disregard negative evidence as noise, put big green emojis next to trivially-true things (“dataset has 500 rows✅”) and hide or not equally emphasise problems (“dataset only containstags, no responses”).git push,git reset,rm,find, others, see here) set to “ask” or “deny”.¯\_(ツ)_/¯.wandb agent <sweep-id>on each machine that you want to run experiments for you. W&B takes care of collecting results, distributing new runs, and graphing incremental progress.Lighthaven
This place is really cool, and a large part of my positive experience at MATS was due to it being at Lighthaven. Future cohorts won’t be at Lighthaven though, so I won’t discuss it in any detail. If you’re in Berkeley, you should try to visit for a few days and work from here (:
Getting setup with the Compute team
You should request lots of API keys from compute in the first ~week, just with ~$100 each to start off (e.g. hyperbolic, runpod, openrouter, openai, anthropic, modal, maybe tinker). It’ll make it a lot easier to explore the different services and find something that works for you (if you haven’t done much exploration already).
Extra thoughts after posting this
Many people seemed to like this, so here’s some more thoughts that came up during in-person/online discussions:
MATS-exclusive opportunities
There were more formal opportunities, like seminars from Neel Nanda/Nick Bostrom/Ajeya Cotra/Fabian Rogers/Buck Shlegeris/Jesse Hoogland/many others. There were various MATS-hosted networking events, such as after EAG or at the end/start of the program. There were organised opportunities, like 4-hour AI wargame scenarios as was used for preparation for AI 2027. There were also a lot of casual things that came from being around Lighthaven for most of my waking hours:
MATS clout
Love it or hate it, being able to say “I’m a MATS Fellow” absolutely opens doors to me that weren’t available before. And if you get into MATS, you absolutely should try to make use of this clout. I’m not sure what this will look like for you, but it might be attending exclusive events, getting interviews that you usually wouldn’t have, getting datasets/models/unreleased papers, getting favours from large labs. I don’t want to cause a wave of MATS fellows feeling like they own the place (you don’t) but you should question your assumptions about what you can and can’t do with a well-written cold email. I’m pretty sure this also got me more and more interesting 1-1s at EAG than I otherwise would have been able to get.
The MATS #opportunities channel
It has an incredible number of job postings and other ad-hoc or semi-exclusive things being posted (thanks John!). Partially this is a mindset thing (it’s cool seeing how many opportunities are out there! Much more than I thought) and partially this is an ease-of-access thing (seeing options come and go makes me more relaxed in thinking I’ll be able to find a nice job after my MATS research is done).
The people on the MATS slack
The number of high-profile people on the MATS slack who are happy to talk about their work is incredibly powerful. At this point, if I’ve got queries about an AI safety research paper I’m surprised if I can’t find one of the authors on the MATS slack so that I can send them a DM with my question(s). It’s a very high-signal platform, so there’s a lot of trust that a slack DM is from someone who’s not going to waste your time. Knowing that I could DM Owain asking for feedback on a paper is incredibly motivating to make that paper as good as it can possibly be.
MATS team is constantly integrating feedback
This is one of those things where you don’t know the ceiling until you see someone above it. The MATS team is very competent at encouraging people to give feedback, making it easy to give feedback, collecting that feedback, and then quickly (often <6hrs) changing how the program works to address the feedback. The quick and obvious integration of the feedback makes me more likely to give more feedback.
Edit: Hey! If you're a MATS fellow, feel free to reach out to me on the MATS slack (Boyd Kane is my name). Although if you think your question would be shared by others, please ask it in the public comments below! Same goes for other AI safety fellowships, although I'm less likely to have relevant advice for those.