Quick takes
Yesterday's Anthropic research ("Emotion Concepts and their Function in LLMs") provides a fascinating mechanistic analogue that highly resonates with the field observations from my March audit of GPT-5.2 Thinking.
While Anthropic studied Claude Sonnet 4.5 and my audit focused on GPT-5.2, the structural alignment between their white-box findings and my black-box observations is striking:
- Accumulation mechanism: In the audit, I documented how prolonged conflict or user "irritation signals" lead to a pattern I called "Procedural Capture". Anthropic's paper demo
🟪 Wide-spread epistemic anomaly detected || Global Risks Instant Message #01-04-2026
We are detecting today a shared collective delusion leading victims to degrade their epistemic standards. This anomaly is aimed towards no particular end, except perhaps for the amusement of its participants and the satisfaction of ingenious expression.
So far, it appears to be mostly harmless. Nonetheless, this phenomenon creates space for vulnerabilities. If some geopolitical actor were to take some implausible action on this day (for instance, US to invade Canada, Spain ...
Some possible containment procedures are as follows:
Altering the Gregorian Calendar to change Leap Day to April 1st (unknown effectiveness, could lead in transferal of the anomaly to another day)
The teaching of mind-resistance techniques in schools and workplaces, using standard cover stories (media literacy, appreciation of the arts, combating racial bias). However, this runs the risk of collapsing important delusions to the functioning of society.
Global usage of hypnotic drugs through the atmosphere, as well as using sleeper agents in the government to f...
Does anyone know why @William_MacAskill says he is "not convinced by the shrimp argument" on his recent appearance on Sam Harris's podcast?
...SAM HARRIS
So yeah, so this is one area where perhaps my own cynicism creeps in. I worry that any focus on suffering beyond human suffering, it risks confusing enough people so as to damage people's commitment to these principles. So I mean, I'm not, there's zero defense of factory farming coming from me here, but When I see a philosopher who's clearly EA or EA-adjacent arguing on behalf of the welfare of shr
Hi Aaron and Will. I estimated how much cage-free corporate campaigns for layers, and the Shrimp Welfare Project’s (SWP’s) Humane Slaughter Initiative (HSI) increase the welfare of their target beneficiaries for individual welfare per fully-healthy-animal-year proportional to "individual number of neurons"^"exponent", and "exponent" from 0 to 2, which covers the best guesses that I consider reasonable. An exponent of 1 would correspond to the linear weighting preferred by Will. Below is a graph with the results. I calculate cage-free corporate campaig...
Seriously, I love this EA forum holiday ❤️ I genuinely feel like this helps the community do more good, get more silly-but-perhaps-with-a-grain-of-usefulness ideas across, and waste time in a way which feels a bit productive
You should turn your project into an organization
If your team's work is worth doing, it's worth doing as an org
When a few people are doing good work together, the question of whether to formally incorporate into an organization can feel like a distraction from doing the actual work. Why take time away from your exciting research project to create an org? There are some real up-front costs to incorporating – dealing with bureaucracy, legal overhead, governance obligations – but I think the benefits of doing so are usually greater and underappreciated.
Orgs a
...I mostly strongly agree with this but think it's worth considering "being an official, recognized, and funded part of an organization" rather than constituting one's own from scratch. I know Rethink Priorities and Hive have sponsored projects before - that seems like a possibly-good intermediate step, with the possibility of spinning out independently later
Look I know I'm on the forum too much @Toby Tremlett🔹 , but I don't think its necessary to put "reading limit" controls on me....
How organisations with low AI usage can and should be using it more
There is a lot of discussion about how everyone should be using AI more, and efforts to increase use and literacy. So far in animal advocacy spaces where I work I’ve seen the following efforts to increase usage so far:
- Orgs provide model subscriptions to their teams.
- People share the ways they’ve been using AI in slack channels or recurring meetings.
- There are educational webinars or fellowships.
The above has made a real dent in AI usage, but much less than we should be aiming for given ...
Yeah I have, and my impression from those I've spoken with is that this has not been the case. You don't think most people whose job primarily involves sitting at a computer could have much of their job automated by a software engineer on call? For example:
- I know grantmakers who have significantly automated parts of their work.
- I know people who have classified 1,000 people in their CRM across a range of people using AI instead of manually.
- I've seen some impressive use of AI to go through 1000's of academic papers looking for novel solutions to a welfare that might exist but is not widely known.
[ETA: I posted a revised version of this essay here.]
AI pause advocates often say they are pro-technology and pro-economic growth, and that they simply make one exception for AI because of its unique risks. But this reasoning will grow less credible over time as AI comes to account for a larger and larger share of economic growth.
Simple growth models predict that AI capable of substituting for human labor will raise economic growth rates by an order of magnitude or more. If that's right, then AI will eventually be driving the vast majority of technological...
Many pessimistic predictions about AGI or ASI tend to paint the picture of a superhuman agent with an extreme maximalisation mindset powered by some unsophisticated version of rationalist principles, which would lead it to commit unspeakable acts of violence (e.g. the paperclip problem: the AI starts killing every form of life in order to save energy that could otherwise be used to make more paperclips).
This, to me, seems somewhat antithetic with the very notion of intelligence.
Surely, a truly 'superior' agent would be able to question the goal of tu...
Over on my blog, I wrote about prediction models, replacement value, and how I was taught about saving lives for pennies on the pound.
So long Mo Salah, and thanks for all the lives you saved.
"Death in a Shallow Pond": A new-ish book on the 'drowning child' thought experiment and EA
TIL about this book: Death in a Shallow Pond: A Philosopher, A Drowning Child, and Strangers in Need, published September 2025, by David Edmonds. I can't find it mentioned on the Forum but apologies if I've missed it. I haven't read it, but according to the blurb, it discusses 'the experiences and world events that led Singer to make his radical case and how it moved some young philosophers to establish the Effective Altruism movement, which tries to optimize philant...
Help me find my replacement doing farmed animal advocacy grantmaking!
I wanted to share a job opening for, in my opinion, one of the coolest jobs to help animals: my job! I'm moving on from Mobius soon, so we're looking for the next person to lead our grantmaking and entrepreneurial projects.
The role: You'd manage the grantmaking portfolio for one of the top ten largest funders of farmed animal welfare work globally, plus lead entrepreneurial projects like incubating new organisations and identifying strategic gaps in the movement. You'd work with a small a...
Coal and nuclear electricity generation kill a significant number of fish through water intake systems. This matters for evaluating the impact of any new electricity load.
Most thermal power plants (coal, nuclear, and to a lesser extent gas) draw large volumes of water from rivers and lakes for cooling. This causes two underappreciated harms to fish:
Impingement — fish get trapped against water intake filters and die. Entrainment — eggs, larvae, and small fish are pulled through pumps and heat exchangers, killing them. A single coal plant in Ohio (Bay Shore ...
To be honest it wasn't my intention to argue that fish eggs have moral weight - I included them to give a sense of the scale of impact - but I can see how that came across, so apologies.
Fish eggs may more clearly have moral worth under non-utilitarian value systems, such as believing that all life that will eventually be sentient has intrinsic moral worth, or that impacts to nature should be minimised for intrinsic reasons (to be clear I don't hold these views personally, but maybe having some level of moral uncertainty leads to a non-zero moral weight).
On...
There's been some discussion here of the claim that AI capabilities improvements have been a consequence of unsustainable increases in inference compute. Redwood Research Astra fellow Anders Cairns Woodruff has written a great post analyzing the data and disputing this.
Unjournal AI-assisted research prioritization dashboard (very early prototype)
We've been experimenting with using LLMs to help identify and prioritize research for Unjournal evaluation, to work with and complement human prioritization (and learn). We now have a public prototype dashboard:
uj-prioritization-dashboard.netlify.app
What it does: Automatically discovers recent papers from NBER, arXiv (econ), CEPR, SSRN, Semantic Scholar, EA Forum paper links, and OpenAlex, then scores them using AI models (GPT-5.4 family) against our prioritization criteria — d...
Thanks, I'd be up for hearing insights. This is related to a larger project (see https://llm-uj-research-eval.netlify.app/), but this part of it is still pretty early stage.
Will DM
🚨 New EA book cover to critique 🚨
Tell me all the ways we messed up, and how/why the original was better actually (see for example this excellent alternative design by Catherine: https://x.com/wilhelmscreamin/status/2029302612210626958 )
Reminder that the symposium kicks off in an hour! If you want to help the conversation go well, you can write up particular considerations, cruxes or questions you have as comments on the symposium post. Invited guests and other participants will respond to them later.