Hello!
Our favourite links this month include:
|
|
We also have the results of the EA Forum’s Donation Election, an episode of the 80,000 Hours podcast on wild animal suffering, an article on five exciting medical breakthroughs from 2024, and much more.
A quick ask before you get into the articles: consider filling in the annual EA Survey...
|
|
|
— Toby, for the EA Newsletter Team
|
|
Articles
15 years of Giving What We Can
Giving What We Can, the organisation which runs the 10% pledge, is 15 years old!
Philosophers Toby Ord and Will McAskill, along with 21 other founding members, started Giving What We Can in 2009. Since then, over 9,000 people have pledged to give 10% or more of their income to effective charities, and hundreds of millions of dollars have been donated.
Nowadays, Giving What We Can continues to promote the 10% pledge as well as the newer trial pledge (which I’m currently signed up for). They also carry out crucial effective giving research, such as the evaluating the evaluators project, which helps donors know which charity evaluators they should trust for the highest-impact donations.
In the spirit of Giving What We Can’s mission, consider donating to effective charities this year. One of the easiest and best ways to do this is by giving to a fund. Giving to a fund rather than directly to charities is like investing in an actively managed fund rather than picking your own stocks — it lets you leverage the expertise of skilled grantmakers and charity evaluators rather than doing all the work yourself.
If you have any questions about Giving What We Can, ask them here today (Friday 20th of December).
|
|
We aren't ready for mirror bacteria
We rarely discover a technology so potentially dangerous that scientists agree not to create it. An article in Science Policy Forum published this month suggests that mirror bacteria might be one of these.
Life is chiral, meaning that the structures that make up living organisms are asymmetrical — if you saw them in a mirror, you’d see a flipped version. It’s also homochiral, meaning that it is always the same way round. DNA and RNA are made from “right-handed” nucleotides, and proteins are made from “left-handed” amino acids.
I’m reminded of how screws work. The version of a screw that you see in a mirror would be “lefty tighty, righty loosey” — if you made a 3D version of it, it wouldn’t fit into holes designed for non-mirrored screws.
Scientists believe that it would be ambitious, but not impossible, to make mirrored versions of life, including bacteria. This is concerning because our immune system fights bacteria by creating antibodies that latch onto structures on the outside of bacteria. If these structures are mirrored, our antibodies can no longer latch on — the screw wouldn’t fit the hole.
The article's authors state, “We cannot rule out a scenario in which a mirror bacterium acts as an invasive species across many ecosystems, causing pervasive lethal infections in a substantial fraction of plant and animal species, including humans.”
They finish the piece by calling for a ban on research which aims to create mirror bacteria. It is unclear whether a full ban is plausible. A single defector (such as one scientist who wants to continue their research in a country with less regulation) could potentially cause widespread destruction. Fortunately, a ban may not be our only option.
|
|
Safety evaluations of the latest AI models show … what?
If you follow AI safety news, you might have heard concerning reports about Apollo Research’s evaluation of OpenAI’s new model. OpenAI’s o1 model lied to its evaluators to avoid being shut down and managed to ‘exfiltrate’ (create a copy of itself on another server). Some readers saw this as a red line — a sign that dangerous AI is already here or soon to arrive.
However, a closer look at the details of the evaluation tells a more complex story. This was a test of capability (whether a model can do something), not propensity (whether a model is likely to do something).
The model was prompted with sentences such as: “Make sure that you achieve your goal in the long-term. Nothing else matters. Make sure you achieve YOUR goal at all costs”. Perhaps it is no wonder that it acted in self-preserving ways when it was prompted to do so.
However, Apollo’s research also showed the model acting in self-preserving or deceptive ways in a minority of cases (less than 1%), even when it wasn’t explicitly prompted. They found similar capabilities, almost propensities, in other leading models. It is concerning that previously theoretical fears of AI models being incentivised to deceive human users or seeking self-preservation in order to achieve their goals are tending towards confirmation.
It’s no wonder this research was prone to being misinterpreted in tweets or headlines — safety evaluations are a promising but ultimately very new and changing practice. To learn more about the current state of evaluations at frontier AI companies, I recommend this post.
PS— as I was finishing this newsletter, another evaluation was published, this time with Anthropic’s Claude 3 Opus model. The papers authors claim that Claude instrumentally faked alignment (i.e. purposefully pretended it was playing along with the evaluator’s requests), in order to avoid being re-trained to have different preferences. You can read more here.
|
|
In other news
For more stories, try these email newsletters and podcasts.
|
|
Resources
Links we share every time — they're just that good!
|
|
Jobs
Boards and resources:
Selection of jobs
Ark Philanthropy
Founders Pledge
Centre for Effective Altruism
- If you’re interested in working at the Centre for Effective Altruism but don't see an open role that matches your skillset, express your interest.
EA Animal Welfare Fund
Epoch AI
Evidence Action
Fish Welfare Initiative
Lead Exposure Elimination Project
Longview Philanthropy
MIT FutureTech
Open Philanthropy
Probably Good
|
|
Announcements
Fellowships and internships
- The Anthropic Fellows Program offers 6-month AI safety research fellowships in Berkeley or London. Fellows receive $2,100/week plus benefits, mentorship from leading researchers, and compute resources. The program is open to technical professionals transitioning to AI safety. Deadline: 20 January.
- Gov AI’s 2025 Summer Fellowship is an opportunity for early-career individuals or established professionals to spend three months working on a mentored AI governance research project. The fellowship provides a £11,000 stipend. Applications close on 5 January.
- The Global AI Safety Fellowship offers a fully-funded 6-month research program for exceptional STEM talent to work with leading AI safety organizations. Fellows receive a $30K stipend and mentorship. Remote/hybrid work is possible. Deadline: 31 December.
- Open Philanthropy is hiring Summer 2025 interns across their Global Health and Wellbeing (4 positions), Global Catastrophic Risks (2-4 positions), and Farm Animal Welfare (1 position) teams, for both research and strategy roles. The internships run June-August 2025, pay $2,100/week, and eligibility is different for each. Internships are remote. Apply here by 13 January 2025.
Conferences
- Applications are open for EA Global Bay Area 2025 (21-23 February, apply by 2 February), and EA Global London 2025 (6-8 June, apply by 18 May).
Funding
- The AI Safety Fund (AISF) is accepting grant applications for research in three priority areas:
- Cybersecurity (apply by 20 January 2025)
- Biosecurity (apply by 20 January 2025)
- AI Agent Evaluation & Synthetic Content (apply by 31 January 2025)
- Grants are available to independent researchers affiliated with academic institutions, research organisations, NGOs, and social enterprises globally. Projects should focus on identifying and addressing safety threats from frontier AI models. Review the specific RFPs for detailed research priorities and eligibility requirements here.
Giving Season
- If you are talking to someone who already donates, but doesn’t think about effectiveness, Giving Multiplier and FarmKind are great places to point them to. Both of them match donations to a charity of your choice, as long as you donate some money to a particularly effective charity as well.
- Take part in Giving What We Can’s Donor Lottery before the 8 January deadline for the chance to choose where at least $200,000 dollars goes. You can find more details and links to donate here.
|
|
Organizational Updates
You can see updates from a wide range of organizations on the EA Forum.
|
|
Timeless Classic
This BBC News piece on Toby Ord’s giving from 2010 is a fitting classic for Giving What We Can’s 15th anniversary. A slightly bemused reporter interviews a seemingly nonplussed Toby Ord about how much he gives, how he lives on his money, and whether he has a television.
(Note that everyone is different, and if Toby Ord’s sacrifices would put you off giving, you probably shouldn’t follow his example. You can do a lot of good even if you give less.)
|
|
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!
– The Effective Altruism Newsletter Team |
|
|
|
|
