Iām an iOS and Deep-Learning Engineer. Iām currently on a career transition looking in to a variety of impactful career paths.
Outside of work Iām a British nomad who likes chess, gaming, bouldering, anime and pretty much any other stereotypical hobby for a software engineer.
It seems like there would be a lot of benefits to having public impact assessments.
It verifies itās taken place, assures applicants a given talent programme is worth their time, gives talent programme orgs visibility on whoās most effective and why, can be seen by other potential donors, and would allow the impact assessment process to be scrutinised and improved upon.
Maybe cG canāt do this as making how they make donation decisions might corrupt their process. In this case the independent analysis fulfills a different purpose rather than being better or worse.
I think informing rejected candidates of how many people applied is another quick-win. The only time Iāve got figures on applicants is when Iāve got in to a fellowship or role and never when Iāve been rejected.
So far Iāve seen:
* Spar Projects with 1ā50 applicants and 6ā75 applicants accepted. (Iāve never got in to SPAR but have worked in an office with a SPAR mentor).
* Two fellowships where 20ā140 and 15ā170 were accepted respectively.
These were all unpaid positions (bar the last one which covered living expenses), I can only imagine applying for paid roles would be even more competitive.
I think having rejected candidates know how competitive these roles/ācourses/āfellowships are would likely reduce the drop-out from the AIS talent pool so would be a win-win for both the applicants and the orgs.
The text is a bit hard to read when you zoom in on the images linked in this post.
The google drive link was high enough res, mostly commenting to make your work more accessible to people scanning the article.
Python & Fellowship Tests
Repost from blog (Iād appreciate any feedback)
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 Astraās fellowships both required code signalās 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. Iām generally not great at thinking fast be it coding or chess.
I did enjoy practising Python. Itās very valuable given most the career switches Iām 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 youāre not going to need for the job. I created some mock practice tests using LLMs that were excellent preparation for the real deal. Iāve 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 didnāt enjoy these. But they were probably valuable for fit testing; if Iām not enjoying reading papers and writing about them, AI safety research probably isnāt going to be my cup of tea.
Iāve decided, for the time being, to stop applying for new fellowships. Joining a fellowship would be a great fit test in itself, but itās quite a large time commitment, especially when you include the x number of applications youāll need before getting accepted to one. I have limited time to do fit tests this year, so my priority now that Iām getting a sense that pure research isnāt 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.
Thanks, this is just the feedback I was looking for.
Hereās the original looks like I need to publish drafts for them to be visible, Iāve edited the original post. Iām weighing the AI version vs the 30 minute unedited brain-dump as the thing which puts me off writing these sort of updates is that my perfectionism can cause me to put several hours in to something I intend to publish and I donāt want to put too much time in to this.
On the other hand sloppily written blog posts might be a net negative thing to be publishing in the first place so not doing them or keeping them private is also a valid choice. A fourth choice could be designing a prompt to do less invasive editing. What do you think is the best approach if Iām looking to keep the time I spend writing this sort of thing to a minimum?
Iād like to get feedback on the writing style of this post. I want to try to write up bi-monthly updates but donāt enjoy sinking time into writing.
Iāve never really stuck with blogging despite it being valuable for sharing what Iām working on as Iām a bit of a perfectionist. I end up spending hours combing over the posts I make. Iād like my posts to only take 30 minutes, so my current ideas are to write quickly and post as is, or to have an AI edit out my mistakes.
Which of the two do you prefer? Do you have any suggestions on ways to make quick blog posts without potentially attaching poor communication or AI slop to myself?
My original post.
My prompt.
Claudeās edit of my post:
------
My Two-Month Deep Dive into AI Safety: From Imposter Syndrome to Clarity
How ARBOx and ARENA helped me navigate a career transition into AI Safetyāand what I learned about myself along the way
Two months ago, I committed to spending my summer diving headfirst into AI Safety. As someone with a background in Swift development and traditional software engineering, the world of Transformers, Linear algebra, and AI alignment research felt like an entirely different universe.
Hereās what happened when I threw myself into ARBOx and ARENAāthe good, the challenging, and the surprisingly clarifying moments that helped shape my career transition.
ARBOx accepted me for their intensive program: one week of prerequisites, followed by two weeks of in-person training in Oxford.
The reality check was swift. During pair programming sessions, I was often the weaker partner. While my colleagues brought post-grad experience with deep learning or career backgrounds in ML, I was frantically trying to remember basic PyTorch syntax. My years of Swift development, unit testing, and design patterns suddenly felt irrelevant when staring at Jupyter notebooks full of tensor operations.
However, being the āslowerā partner was actually incredibly valuable. My more experienced partners were amazingly patient, and having someone literally show me concepts I didnāt understand was worth twice the learning compared to struggling alone. Sometimes the best education comes from admitting what you donāt know.
The human element made all the difference. Working on AI Safety material in isolation can feel overwhelmingālike youāre trying to solve impossible problems alone. Being surrounded by 19 other people going through the same learning process, discussing niche AI Safety topics over meals, was genuinely inspiring. It reminded me that this field is built by communities of people, not just individual brilliance.
After Oxford, I faced a choice: rush through ARENAās materials to keep up with their intense timeline, or slow down and actually master the fundamentals.
I chose to slow down, and hereās why that decision was crucial.
Everything builds on the basics. My biggest blockers during ARBOx were gaps in PyTorch, linear algebra, and tensor operations. Rather than continuing to build on shaky foundations, I decided to really nail these prerequisite skills. If Iām transitioning careers, I reasoned, these fundamentals wonāt just help with AI Safetyātheyāll be essential for any ML-related work I pursue.
Active learning over passive consumption. ARENA encourages rapid progress through pair programming, often skimming reading materials to finish notebooks in a day. While this works well for exposure to topics, I found more value in:
Writing Anki flashcards for key concepts
Watching YouTube videos to visualize complex topics
Going through additional reading materials
Actually understanding rather than just completing exercises
Iāve created a fork of the ARENA syllabus to work at my own pace, and Iām considering building demos of my favorite exercises to start developing a deep-learning portfolio.
Midway through this journey, I applied for SPAR (Supervised Alignment Research)āa 3-month program requiring 10-30 hours per week. On paper, it seemed perfect: hands-on AI alignment research experience during my career transition.
But as I worked through the application, something became clear: I was applying for the wrong reasons.
Hereās what gave me pause:
Time commitment vs. exploration breadth: Committing 3 months to one niche area when I only have 12 months total for career transition
Skills mismatch: The program values my software engineering background, but I need to develop ML skills most
Practical constraints: Iāll be working from Bali for 2 of the 3 months, limiting my in-person or synchronous involvement.
Motivation misalignment: I found myself struggling to get excited about reading & writing papers, applying for funding, and running scaffolding experiments
Realising that pure research might not be my best fit helped clarify what actually excites me. AI Safety engineeringābuilding robust systems, tinkering with deep-learning model, creating tools that researchers can useāaligns much better with my background and interests.
Sometimes the most valuable outcome of an application process is discovering what you donāt want to do.
Imposter syndrome is data, not truth. Feeling like the weakest person in the room doesnāt mean you donāt belongāit often means youāre in exactly the right place to learn quickly.
Community matters. The isolation of solo learning can make career transitions feel impossible. Finding your peopleāeven temporarilyāprovides motivation and perspective that no amount of individual study can replace.
When to go slow or fast. The pressure to quickly ācatch upā in a new field can lead to surface-level learning. Sometimes the best strategy is to thoroughly understand the foundations, even if it feels like youāre moving slowly. For less foundational topics a quicker approach can pay dividends so you can get a sense of which topics interest you and go in greater detail later when youāve a better idea of what to focus on or can apply the knowledge in a real project.
Career fit-testing is as important as skill building. Donāt just ask āCan I do this work?ā Ask āDo I want to do this work?ā The difference between research and engineering roles, for instance, isnāt just about skillsāitās about what energises you.
Iām continuing through ARENA fundamentals with a focus on building solid foundations rather than racing through material. Iām also exploring AI Safety engineering opportunities that better align with my background and interests.
The goal isnāt to become an AI Safety researcher in two monthsāitās to understand the landscape well enough to make informed decisions about where I can contribute most effectively.
For anyone considering a similar transition: Donāt underestimate the value of being honest about what you donāt know, finding communities of fellow learners, and taking time to understand not just what you can do, but what you want to do.
Whatās been your experience with career transitions in technical fields? Iād love to hear about your journey in the comments below.
Iām currently taking a career break intended to fit-test what impactful careers suit me. Iāve created a spreadsheet with a weighted factor model (Altruism/āCareer projects tab) and a rough schedule. Iām eager to get feedback on how Iām planning to spend my time and how Iāve prioritized what to work on.
blog-post on topic.
Our best-guess estimate of GWWCās giving multiplier for 2023ā2024 was 6x, implying that for the average $1 we spent on our operations, we caused $6 of value to go to highly effective charities or funds.
The Centre for Exploratory Altruism Research (CEARCH) estimated GWWCās marginal multiplier to be 17.6 % (= 2.18*10^6/ā(12.4*10^6)) of GWWCās multiplier. This suggests GWWCās marginal multiplier from 2023 to 2024 was 1.06
Why is there such a large difference between these multipliers?
After googling I think I understand a marginal multiplier is what the next dollar donated to GWWC returns but would like to clarify and post this so others such as myself donāt mistake this as a large reporting discrepency.
If Iām understanding the marginal multiplier correctly I would also be interested as to why the return on the next dollar donated is so much lower than the return on the average dollar donated.
That would be really useful!
Some of my ideas for forum or blog posts are:
Bi-weekly updates on what Iāve been working on.
Posting stuff Iāve worked on (mostly ML related).
Miscellaneous topics such as productivity and ADD.
Reviews of EA programmes Iāve taken part in or books Iāve read
Dumping my thoughts on a topic
Iām also interested in how you differentiate between content better suited for a blog or better suited for a forum?
Iām a 36 year old iOS Engineer/āSoftware Engineer who switched to working on Image classification systems via Tensorflow a year ago. Last month I was made redundant with a fairly generous severance package and good buffer of savings to get me by while unemployed.
The risky step I had long considered of quitting my non-impactful job was taken for me. Iām hoping to capitalize on my free time by determining what career path to take that best fits my goals. Iām pretty excited about it.
I created a weighted factor model to figure out what projects or learning to take on first. I welcome feedback on it. Thereās also a schedule tab for how Iām planning to spend my time this year and a template if anyone wishes to use this spreadsheet their selves.
I got feedback from my 80K hour advisor to get involved in EA communities more often. Iām also want to learn more publicly be it via forums or by blogging. This somewhat unstructured dumping of my thoughts is a first step towards that.
If you take either of these pledges while you are a student or unemployed, it is within the spirit of the pledge to give 1% of your spending money until you start earning an income, at which time you would then begin giving your pledged amount.
Does the lifetime pledge tracking on your website allow for this 1% unemployment/āstudying period? Or is the idea that you would make up the time where you were only giving 1%?
Iām going to be exploring Berlin that day but will try and join in the evening for dinner. What time do you think youād be carrying on until?
How do you call up HMRC to adjust your tax code? Do you need to know how much you will make /ā donate in advance to do this?
Iāve been filling out self assessments each year to do this and itās been getting more difficult each year. 2 years ago the online repayments stopped going through and I needed to call up their helpline. last year the helpline couldnāt help me and I needed to write a letter to HMRC.