I took on three projects:

• A meta-analysis of speciesism benchmarks

• Contributing towards a mid-training data pipeline

• Planning for the project I’m doing now back in the UK, creating emergent misalignment from factory farming manuals

There were around 12 other residents from varied backgrounds. Mostly from the US, but we also had people from Egypt, Vietnam, Ireland and another Brit besides myself.

The residency provided accommodation near Mox, plus full membership and two meals a day. That gave us our own private office with monitors, and a very homely work environment. Lots of people would relax there and take part in the regular events, and there were free snacks and drinks. It’s also really inspiring to be around other people working in the field, exchanging ideas and getting feedback on your own. I’d definitely recommend it to anyone looking to enter animal welfare or the AIxAnimals field.

It was also a good chance to see whether moving to San Francisco would be my cup of tea. I’ve often considered it. I’ve lived abroad for the last 12 years and enjoy it, living in North America is something I’ve long wanted to try, the salary is very competitive, and it’s probably the best place to get started in AI safety.

The trip has made me update against it though:

• It’s a long way from my friends and family in Europe

• The visa process is arduous and would likely commit me to the country for a significant time

• For the high impact jobs I’d be competing for, I think my counterfactual would be bad. I’m more likely than most to leave early and come home, and I don’t keep the same hours as most of the EAs working in SF

It’s a lovely city though, and it was very easy to make friends and meet new people. I’d highly recommend it to people who don’t share my circumstances and beliefs.

The residency was fairly unstructured. The first two weeks were packed with activities, training exercises, onboarding and networking events. After that it tapered off to one or two events a week, usually Q&A sessions with people in the animal welfare or AI safety space. That doesn’t include the after-work activities we did together, which were very good from a networking standpoint.

The day-to-day culture was unusual from a European standpoint. We spent long days at the co-working space, often from mid-morning until late at night, but a lot of that time went on socialising and events, meals with other residents, Sentient Futures and the other organisations at Mox, and chatting to people over the free drinks the space provided. Still, I’ve been working part time for the last two and a half years, and I don’t think that pace suits me.

I preferred the lack of structure as I had plenty to do. It might be less suited to people who don’t have a good list of project ideas to go for.

I learnt a lot more about the AIxAnimals space by spending time among the people in it. The main surprise was how useful it is to be in San Francisco. We regularly got feedback on our projects, in some cases from people at frontier labs, usually through contacts we’d met at parties or after work events.

If you’re interested in the type of AIxAnimals projects SentientFutures work on you can find a list of past incubator projects here.

This was a pretty quick 2 day fit-test but worth sharing as other people may want to try it. AI Red teaming is the process of trying to get AIs to do things they shouldn’t do. (i.e. give OpenAI’s wifi credentials, or tell you how to make a bomb).

Trying it out

Gandalf is a great introduction to this. There’s a list of solutions for if you get stuck. It’s a good test for if you find this fun.

After trying Gandalf out I applied for a redteaming role at TrajectoryAI. Their test was a similar sort of tool to Gandalf but was a bit more difficult as it involved hacking a company through their chatbot by incrementally getting access to more.

Verdict

I wasn’t particularly good at this and although I enjoyed getting past each level of Gandalf and Trajectory’s AI simulations the grind of trying things out over and over was a bit of a drag. I also can’t see this sort of jailbreak surviving AI automation. I was impressed with how easy it is to get hands on experience doing this as both Gandalf and TrajectoryAI’s testing apps were pretty sleek and self contained.

I suspect there’s more diversity to this field than prompt hacking. If I were to try this further I would probably try out some of GraySwan’s hackathons.

In February I got in touch with CaML as part of the AIxAnimals incubator, run by Sentient Futures. They tasked me with putting MORU Bench (Moral Reasoning Under Uncertainty) up on Inspect a benchmarking service run by the UK’s AI Security Institute. My PR was accepted and completed the project within three weeks. It was my first project working with Claude Code.

It felt surprisingly familiar despite being python rather than Swift. Building tooling, writing to a spec, making sure test coverage is solid. Of all the things I’ve done during this career switch, this is probably the closest to what I actually did as an iOS framework engineer.

A few things stood out to me:

• The Inspect team had an agents.md setup that could run workflows for checking whether your benchmark was ready for a pull request, whether it had sufficient test coverage, and for generating a report from your results. These files are incredibly useful when onboarding people to an unfamiliar project. When you’re new you don’t know what to prompt for because you don’t know what’s there, you don’t know the guidelines, you don’t know the architecture. Having an agents.md automates all of that.

• Having a fairly thorough agent-driven PR review before a human looks at it seems like a good process to adopt. I know a lot of people in the industry who are struggling with the volume of pull requests coming through because of AI coding tools. This feels like a reasonable way to enforce some quality before it hits a human reviewer, which isn’t sustainable at scale if people are just vibe-coding their PRs.

Is this impactful work?

I’m a bit torn. On one hand, what CaML are doing (creating benchmarks that measure how an AI views animal welfare, how compassionate it is to non-human beings) seems like a great way to influence model behaviour. These benchmarks are used by frontier labs as targets to hit before releasing a model, so they have a pretty direct effect on how models will end up behaving.

On the other hand, I’m apprehensive about benchmarks in general. During my time at the EA hotel I spoke to a few AI safety people who considered most evaluations as progressing capabilities, not safety. The logic being that an evaluation measuring frontier maths or any other capability ends up helping the labs make their agents better at those capabilities. I don’t feel experienced enough to hold a strong opinion on this myself, I’m just aware opinions differ. I’ve been encouraging the people I know who feel strongly about this to write about it because I’d love to see their views stress-tested.

Fit

Of all the fit tests so far, this has been the most promising in terms of how capable I am at it. It maps nicely on to my existing experience and I did enjoy doing it, maybe not quite as much as iOS development, but more so than any other project I’ve worked on during this process.

I also have to bear in mind that creating an evaluation is probably one of the easier tasks. I can see there being more interesting work when you’re actually maintaining a framework rather than using it to create benchmarks. It’s also a project where I’m in a much better position to hit the ground running and get a job.

A couple of concerns though:

• In the case of Inspect, it would likely mean moving to London. Not sure that’s my cup of tea.

• Long-term viability. This is a form of coding where you’re not doing any ML and it’s already highly automatable with Claude Code. I actually met with a former manager of AISI who’s attempting to automate benchmarks. That said, it could be interesting to actually do the automations.

All in all, it’s a high contender so far and a successful fit test. I may go back and do a few additional PRs for Inspect or look into other evaluation projects to revisit this.

In January I spent three weeks at the EA Hotel in Blackpool. It was a really good experience overall.

Blackpool itself was nicer than its reputation suggests. I enjoyed walking around during the day and it has a pretty nice city center with lots of cafes, shops and restaurants. Living with twenty other people turned out to be very motivating, mostly because they were doing similar things to me. Career transitions, AI safety, adjacent cause areas. I got a lot of project ideas and fit test ideas just from being around them.

It was a pretty close-knit community. During the day we’d work in various spots around the building. Some people worked from their rooms but I stuck to the main office since it had a monitor and a desk. Basically a full co-working setup, which suits me. It’s both motivating and keeps me from getting distracted.

Every day at 6pm we’d have dinner together, always a freshly cooked vegan meal. People would arrange events throughout the week too, roughly one a day, maybe slightly less. These included things like:

• Movie nights

• Going to the pub

• Presentations

• A pilates class at one point

• Reality TV (I originally joined to sit in the background and mock this but ended up quite enjoying it)

• Circling (a mix between meditation and group councilling)

• Acoustic Kareoke

There was also a decent climbing gym nearby, a communal bike anyone could use, lots of board games + chess, snacks & protein shakes, meeting rooms, a projector and a tv + bean bags in the loft.

Staying there is free if you’re working on EA-aligned causes or transitioning to an EA-style job. I decided to pay half the cost of hosting me since I was also using some of the time for contract work and £300 for 3 weeks with bills and evening meals included is cheaper than I’d be paying for rent elsewhere. The only obligations are three to four hours of communal housework per week (I did all of mine in the kitchen, helping whoever was cooking. Chopping vegetables for twenty people is no small task.) and a mandatory presentation night on Fridays where everyone shares what they’ve been working on.

If you like co-living and co-working, and you’re working on a career transition towards AI safety, animal welfare or global health, I’d highly recommend it. I believe you can book for anything from a week to three months. I was only there for three weeks but I might go back in summer for a longer stay.

I got a decent amount done in those three weeks:

• Started building an evaluation using Inspect

• Worked on fixing a bug in an ablation tool

• Built a collection of open source projects to help with fit testing

• Updated my weighted factor model

• Got some contract work done to help fund the career switch

I skipped my bi-monthly update for December and January, so this is me catching up. Going forward I’ll write these after completing a fit test rather than on a schedule.

December I mostly took off for Christmas, but I managed to fit in three things:

• The AI x Animals course run by Sentient Futures

• A Code4Compassion hackathon for animal welfare projects

• An interview with Animal Advocacy Careers an animal welfare careers service

AI x Animals Course

More technical than I expected, which was good. No hands-on work within the course itself, but a fellowship followed it that I worked on in January and February. I’ll write about that separately.

Three areas came up worth fit testing:

Animal communication: deciphering animal vocalisations, either for general communication (whales) or welfare purposes (chickens). This is the one I’m most drawn to. It’s exactly the kind of ML project I’d enjoy. The problem is I can’t see many jobs actually opening up here.

Precision livestock farming: there’s a real case for welfare advocates getting involved, since the industry is adopting it either way. Getting welfare measures built in early matters. But limited funding, and whatever jobs do exist would probably go to people with stronger data science backgrounds. My counterfactual impact seems low.

Shaping model opinions on animal welfare: benchmarks for how much a model values animal welfare, pre-training tools to influence what the model absorbs. Not as flashy as animal communication but probably where the most impact is. My concern is the alignment risk. If you teach a model a non-speciesist outlook where animal life counts for even a fraction of human life, you could end up with a model that concludes, pretty logically, that eliminating humans would reduce net suffering.

Animal Advocacy Career Interview

I attended EAGx Connect were I had an interview with AnimalAdvocacyCareers. It was interesting, if a bit sobering. They gave me some links for volunteering and hackathons, but said there aren’t really many jobs for programmers in the animal welfare movement. I also spoke with a data scientist working in animal welfare, who said he could count the number of ML engineers in animal welfare on two hands. I’m very appreciative for the candour they both offered me.

The Berlin Hackathon

I didn’t join live. The project used Claude Code and I figured the pace would be too fast to sync on remotely, so I got a list of things they wanted improved and set aside a day after the hackathon to work through them.

The project was a web app rather than native mobile, which is my area. First time using Claude Code on something I had no familiarity with. I redid the main screen in a couple of hours (would have taken me a lot longer manually). Good introduction. I’ve used it a lot more since.

Verdict

This was a relatively short fit-test in an area I didn’t expect to have opportunities for someone with a technical background. Animal welfare is an area I care about so I’m going to keep an eye on animal communication because the ML problems are exactly the kind I’d enjoy and deciphering whale vocalisations is fascinating. I also went on to do a project with CaML which wasn’t a animal welfare project but every project they had going looked really interesting so I will likely go to them if I want to extend this fit-test later in the year. At present though my preliminary verdict is that this is not somewhere I can easily build a technical career within.

What went well

• Caught up with friends and family in Liverpool, Hangzhou, Bali, Singapore, Las Palmas and Madrid.

• Spent most the year living in Vietnam and Bali.

• Attended and completed the Alignment Research Bootcamp in Oxford

• Did most of the ARENA Fundamentals syllabus

• Studied Pytorch and Python throughly

• Did 11+ different AISafety fellowship applications

• Maintained climbing performance and improved finger strength despite poor health

• Was heartburn free in Vietnam for 5 months showing that the longterm issue I’ve had with this can likely be resolved by my diet

• I did my first D&D game, paddle game and Asian Karaoke session

• I started a year out for transitioning my career to something more impactful/purposeful and got a big severance payment to fund it.

Not as Planned

• The search for an ideal career is still ongoing

• My dating life was a bit purposeless as I was moving around so much

• I was sick for 3 months while in Asia.

Career / Finance

I was made redundant in April of last year, and started investigating different career paths around May or June. My intention with this is to find a career that’s more impactful, either altruistically or for solving interesting technological problems while maintaining the benefits I had from my previous job, or minimising the compromise on them.

This will impact my finances in the short run as I won’t be earning money, and switching careers will likely mean making less money when I first switch to it. I can backtrack to my previous career of iOS engineering if needed although there’s a secondary benefit to this switch which is that a switch to ML/AI would likely future-proof my career long-term more than iOS development.

Going forward next year, I want to spend the first half continuing to up-skill in AI but mostly through projects and less through courses. I want to fit-test different impactful roles I can take. This includes but is not limited to AI safety, LLM Development, anti-disinformation coding, and independent app development.

• Explore AISafety

• Explore anti-disinformation projects

• Explore LLM Finetuning

• Explore RL

• Explore independent app development

• Get a new job

Social

Socially things have gone well this year. I made new friends in Vietnam despite having no prior connections in Saigon and had a lot of social events each week. Outside of Vietnam I was actively keeping in touch with friends and family including travelling to Bali, Las Palmas, Liverpool, Hangzhou and Madrid as part of this. I also got to know plenty of new people in the UK by using my spare time to attend more effective altruism events including ARBOx, the EA North retreat and several EA Liverpool events.

Travelling and messaging takes up a lot of time, I may need to ease up on this somewhat.

• Start dating again, preferably a nomad or someone who I can travel but a local could work if the location is somewhere I can see myself settling

• Cut back on Whatsapp messaging and social media

Self Development

I was kept pretty busy with the career switch which is sort of a giant self-development project. I would like a bit more time to put towards things like singing, dancing, DIY etc.

After reading “Make it stick” a book about the science of learning I started using Anki flashcards. This is a flash card app where you make your own cards or download premade ones online on any topics you want. I’m going to utilize anki more as I retain information from doing this much better than by reading and it takes much less time than learning via tutorials for technical topics. I used it for python, ML, transformers, calculus CNNs and numpy this year and will expand this to cover some of my favourite books, AISafety and some more deep-learning topics.

For my hobbies: Chess I think I can just pick up when I feel like it and isn’t worth pursuing as a goal. Climbing my goal of climbing 10 V5s wasn’t met although I was sick for so long this year it really disrupted things. The other problem with this goal is it’s grade based so it would result in me avoiding my weak points. Spanish I’m keeping on the backburner until I live in a Spanish speaking country again.

• Identify climbing weak points and set goals for them

• Add more books to my Anki

• Try singing, dancing or DIY classes

• Figure out how to fit the above around climbing, going out, and studying

This Autumn has been a bit of a mix of continuing with Arena, being sick, and applying for half a dozen fellowships and courses.

Claude and Astra’s fellowships both required code signal’s industry coding assessments. I did a week of preparation for both of them before taking the test but did a lot worse than I thought I would. Being relatively new to Python and not being able to code at high speed without making mistakes was my downfall with these. I’m generally not great at thinking fast be it coding or chess.

I did enjoy practising Python. It’s very valuable given most the career switches I’m considering use it. It was refreshing to see this kind of test instead of algorithm and data structure tests, which I feel involve revising something you’re not going to need for the job. I created some mock practice tests using LLMs that were excellent preparation for the real deal. I’ve also picked up some Anki flashcards for the most common Python functions, which should come in handy for next time.

I also did tests that involved writing essays about papers for Cooperative AI Research fellowship, MARS, LASR, Deepmind, ERA, and the 2nd stage of Astra. With the exception of the ERA test (really interesting paper on LLM coding capabilities), I didn’t enjoy these. But they were probably valuable for fit testing; if I’m not enjoying reading papers and writing about them, AI safety research probably isn’t going to be my cup of tea.

I’ve decided, for the time being, to stop applying for new fellowships. Joining a fellowship would be a great fit test in itself, but it’s quite a large time commitment, especially when you include the x number of applications you’ll need before getting accepted to one. I have limited time to do fit tests this year, so my priority now that I’m getting a sense that pure research isn’t my thing, is to focus on finishing ARENA, start to explore career paths other than AI safety, and contribute to more open-source projects. I need to be careful not to use up all my time doing applications rather than getting my hands dirty with projects that will give a much better indication of my fit for a given career path.

In order to use it in our forward pass, we need to express this in terms of the previous running average, where the upper limit of our summation is ‘T-1’ aka EMA_T-1 as we’ll be recursively calculating this for each step.

1: Write out the Full Summation

Using the formula we’ve been given, we’ll express the EMA at time T as a series of additions:

EMA_T = μ·mean_T + μ(1-μ)·mean_T-1 + μ(1-μ)²·mean_T-2 + … + μ(1-μ)^T-1·mean₁

2: Look at the Previous EMA

Now consider what EMA_T-1 looks like:

EMA_T-1 = μ·mean_T-1 + μ(1-μ)·mean_T-2 + μ(1-μ)²·mean_T-3 … + μ(1-μ)^T-2·mean₁

3: Notice the Pattern

Look carefully and you’ll see that every mean_T-n value is present in both EMA summations but multiplied by (1-μ) for EMA_T. There is also one additional μmean_T.

This gives us:
EMA_T = EMA_T-1 * (1-μ) + μ·mean_t

or in our forward method:

running_avg = running_avg * (1 - momentum) + batch_mean * momentum

For July & August my focus has been up-skilling for AISafety. I was accepted on to ARBOx which involved a week of studying their prerequisites, 2 weeks attending an in-person crash course in Oxford and then self-studying the fairly extensive ARENA syllabus the course had utilised afterwards.

• 30/06/2025: ARBOx

• 07/07/2025: ARBOx

• 14/07/2025: Travel / Planning / ARBOx prereqs

• 21/07/2025: ARENA Prerequisites / Medical appointments

• 28/07/2025: Break – Family holidays

• 04/08/2025: SPAR/Anthropic Research Applications

• 11/08/2025: ARENA Ray Tracing + Pylance setup

• 18/08/2025: ARENA: Ray Tracing

• 25/08/2025: ARENA: CNNs

ARBOx – Oxford

ARBOx was a great use of my time. It was pretty motivational to be around people much smarter than myself, living out of student halls again, and to be able to focus 100% of the course material through pretty intensive pair programming session. I think much of the material I was lacking the foundations to really utilize but it was a great introduction and the pair programming opened my eyes to how other people work on these exercises (much faster than I do mostly).

I also met some great people there, working on this stuff alone can feel like an impossible task sometimes so having 19 other people going through the same learning process alongside you and chatting to them about all these niche AISafety topics you care about was really inspiring.

I did have a bit of imposter syndrome while there as I was usually the weaker pair in the pair programming exercises. My career experience of Swift, Unit testing and design patterns doesn’t really come in useful when going through Jupyter notebooks and most my partners had career or post-grad experience with deep-learning. They were amazingly patient though and once I put aside feeling guilty over slowing them down it was double valuable going through the material with people who can literally show me the things I don’t know.

ARENA – Post ARBOx

Since leaving ARBOx I’ve focussed on the ARENA prerequisite & Fundamentals materials as my lack of Pytorch, Linear Algebra and experience working with Tensors were my biggest blockers when going through the ARBOx syllabus.

ARENA sets pretty intense time-intervals on their exercises and has a lot of reading material that you generally skip or skim-read when pair programming with the aim to finish a notebook in a day. I think this is a good practice when familiarising yourself with a topic as I personally learn best when having to recall the topic when the time comes to utilise it in a project.

However the fundamentals I’ve decided to go over more slowly for two reasons. One being that everything else on this syllabus is going to rely on it so knowing it well will help me with subsequent exercises. The 2nd being that these foundational topics will not just be useful for AISafety but for most the ML related areas I wish to explore while doing this career transition. So I’m taking it slow, going over some of the additional reading material, writing up Anki flash cards and watching youtube videos visualising key topics.

I’ve created a fork of the ARENA syllabus to work out of for now but will probably take on a suggestion one of the ARBOx teachers gave me and invest a bit of time in to creating a few demos of my favourite exercises to start building up a portfolio and cement my knowledge in each of these areas.

SPAR – Supervised Alignment Research

I put the best part of three days in to applying for this years SPAR cycle. My main aim is to fit-test different career paths and getting hands on AI Alignment research experience would be a perfect fit-test.

However I’m getting cold feet for some of the more research heavy projects for a variety of reasons:

• I’m committing to 3 months of 10-30 hours a week involvement when I only plan on spending 12 months for my entire career transition and there are a lot of areas other than whatever niche of AI Safety alignment research my chosen projects involve.

• My biggest draw as a volunteer is my SWE experience yet its the ML that I need to improve on most.

• I’ll be in Bali for 2 of the 3 months meaning my time zone and involvement will be limited somewhat.

• After unpacking it research heavy roles might not be a good fit for me. I find it really hard to get motivated to; apply for funding, read papers, write reports and run experiments that involve scaffolding LLMs.

• It’s a very competitive field

When trying to decide what to work on during my career switch break I try to use a weighted factor model. This consists of simply listing all the things you could work on alongside the factors that would make a project appealing to you. It’s very much a living document where both the weighing you apply to each factor, the factors you choose and the projects within it can and should be updated regularly. I also keep a schedule and a template within the same document if anyone cares to copy it.

I recently left my job with the intention to change career at the ripe age of 36. I’ve been an iOS developer for 12 years and a professional poker player for the 6 years before that.

iOS development gave me almost everything I wanted. A career I enjoyed, was good at, paid well, could work remotely and even work part-time. I’m incredibly lucky to have had this and if things don’t pan out I’ll likely go back to it.

Why

In 2022 I had my first interview with 80,000 hours, a career service specialised in moving people in to high impact altruistic jobs. Things were going good for my career, I had just landed my highest paying role at a company with a great culture and timezoneless remote working. However I still felt there was something missing, my last few jobs had been working on gambling or social media apps. I liked the code, but not the product and I was likely having a negative impact on the world around me. I wanted to solve problems that benefit humanity which in turn would give me as much enthusiasm about the product I’m working on as the code I write for it.

2023 I went part time to train up in deep learning on my day off. 2024 I switched from Tumblr’s iOS team to our internal deep learning team.

When

April 2nd 2025 I wake up in my apartment in HoChiMinh city and find an email for myself and 16% of Automattic that we’ve been made redundant. It was a shock at first but it came with a generous severance package at a time when our team was increasingly moving away from Deep-learning projects. The time to make the switch real was at hand.