Hello!
My favourite links this month include:
- A leading AI company has decided not to release a frontier model, citing the risk it would pose to the internet's infrastructure. Rob Wiblin's 21-minute explainer is the best place to start.
- The science behind stunning fish before slaughter isn't as settled as recent commitments assume, according to a Rethink Priorities literature review.
- In Development, a new global health and development magazine, launched with a first article from GiveDirectly co-founder Paul Niehaus on the organisation's journey from startup to delivering over $1 billion in cash transfers.
|
|
|
CEA is running online career transition bootcamps for professionals looking to pivot into high-impact work — with a specialised option for operations professionals and organisational leaders. Various dates running April–June; apply by April 21st. Today is also the last day to apply for EAGxDC (May 2–3), and applications for EA Global London (May 29–31) close May 10th. As always, we're sharing jobs, opportunities, and great articles below.
|
 — Toby, for the EA Newsletter Team |
|
|
Articles
For the first time in seven years, a leading AI company has decided not to release its newest model publicly, citing safety concerns. Here's what you need to know:
- During testing, Anthropic's new model, Mythos, was shown to be extremely capable at both discovering bugs in critical software and exploiting them. As part of testing, it found a 27-year-old flaw in a widely used operating system, long considered secure. If released, Mythos could enable anyone to find and exploit such vulnerabilities, to steal data, interrupt electricity grids, etc…
- As a result, Anthropic is launching 'Project Glasswing', a coalition of companies that will be given access to Mythos to help them prepare the internet's infrastructure before a model like this is released publicly.
- On a range of tests, Mythos shows itself to be more aligned than previous models from Anthropic. However, there are some causes for concern. For example, Mythos can accurately tell whether it's being evaluated 78% of the time, and when it doesn't think it is being tested, it acts worse.
While concerning, it's healthy to take this with a pinch of salt. Anthropic is a company enjoying some particularly good press at the moment. This news makes them appear responsible and powerful. Internal testing generally shows models to be more impressive than external tests. If you'd like to learn more, I strongly recommend Rob Wiblin's 21-min summary of the research. You can read, listen, or watch it.
|
|
|
Over the past few years, animal welfare organisations have campaigned for fish to be stunned prior to slaughter. It's worked: major producers have made commitments, and the Aquaculture Stewardship Council's new Farm Standard will require stunning for all certified farms. When the alternative is suffocation or suffocation in ice slurry, this seems like a no-brainer. But it isn't.
A recent literature review from Rethink Priorities found that the evidence doesn't show that seabream and seabass are rendered insensible by stunning. Seabream and seabass are slaughtered in their hundreds of millions by European fish farmers every year.
The research review found that the efficacy of stunning varies by fish species, water conditions, equipment settings and upkeep, and more. Worse, much of the research measures bad proxies for consciousness (i.e. are the fish still moving), which can't reliably distinguish a paralysed fish from an insensible one. The information we have on the efficacy of stunning is badly lacking.
|
📝 Editor’s note (added April 22, 2026): Rethink Priorities wants to emphasize that the evidence for electrical stunning varies by species and is cautiously encouraging for rainbow trout, but weaker for gilthead seabream and European seabass. They are calling for better evidence and implementation so that industry commitments translate into significant welfare gains, not challenging those commitments themselves.
|
|
|
|
A new magazine just launched, which I expect to feature in this newsletter often. The editor shared an early look at their first article, which went live today. The article is by co-founder of GiveDirectly, Paul Niehaus. He writes about the organisation's history of using randomised control trials (RCTs). My key takeaways:
- GiveDirectly, a charity that gives cash directly to people in extreme poverty, has now delivered over $1 billion to more than 2 million people.
- Cash transfers and vouchers in general make up 20.6% of international humanitarian assistance as of 2022, up 50% from the previous five years. Niehaus wants much more. GiveDirectly is undergoing ambitious scale-ups, which still seem to be delivering results. Currently, it is doing a district-wide rollout in Malawi, which is reaching 185,000 people, and it wants to do bigger projects in the future.
- GiveDirectly's early RCTs seemed to be more about confirming what it already thought it knew. Later ones have led to insights that are useful outside GiveDirectly and its donor base. For example, a study designed to answer donors' worries about inflation ended up yielding the first experimental estimate of the transfer multiplier, the effect that mass cash transfers have on the wider local economy.
There's much more to see in the article itself. Also, watch out for In Development's collaboration with the EA Forum next month, where we will highlight the first few articles from the magazine and invite the authors to answer your questions. You can subscribe to our Google Calendar if you want to be reminded.
|
|
|
Resources
Links we share every time — they're just that good!
|
|
Jobs
Boards and resources:
- The 80,000 Hours Job Board features almost 800 positions.
- The EA Opportunity Board now includes full-time roles, alongside internships, part-time opportunities, and more.
- Probably Good's job board features a range of impactful roles (and recently had a makeover).
Selection of jobs
Centre for Effective Altruism
Giving What We Can
- Research Associate (Remote, USD $65K–$84K / GBP £46K–£59K / EUR €39K–€50K / AUD $80K–$120K, apply by April 27th)
Rethink Priorities
Animal Charity Evaluators
Effective Altruism Netherlands
- Co-director (Amsterdam, EUR €50K–€70K, apply by April 19th)
Coefficient Giving
|
|
|
Announcements
Conferences and Events
- Today is the last day to apply for EAGxDC (May 2–3).
- The next EA Global is London, May 29–31. Apply here by May 10th.
- Sentient Futures Summit (May 22–24) brings together AI experts, consciousness researchers, and animal advocates to discuss steering technology for the benefit of animals and potentially sentient AI. Friday conference plus weekend unconference. Scholarships and volunteer tickets available.
Courses and Fellowships
- Generator Residency is a 3-month AI safety residency in Berkeley (June 15–August 28) where residents pitch, build, and ship projects that fill gaps in AI safety infrastructure — then get support landing full-time roles. For generalists of any background. USD $6K/month stipend plus housing and travel. 15–30 residents. Apply by April 27th.
- CEA is running more career transition bootcamps. Identify high-impact careers that match your skills, then build momentum for your next steps with exercises, advisor feedback, and peer accountability. Various bootcamp dates running April–June, including a specialised option for operations professionals and organisational leaders. Apply by April 21st.
Funding and Prizes
|
|
Timeless Classic
Our featured article from GiveDirectly's co-founder reminded me of this interview from 2021. In what is sadly the last published episode of Rationally Speaking, a podcast I'd recommend exploring further, Julia Galef interviews Michael Faye, another GiveDirectly cofounder, about the case for just transferring cash. Together, they cover all the key arguments for and against cash transfers. |
|
We hope you found this edition useful!
If you've taken action because of the Newsletter and haven't taken our impact survey, please do — it helps us improve future editions.
Finally, if you have any feedback for us, positive or negative, let us know!
– Toby, for the Effective Altruism Newsletter Team |
|
|
|
|
