Hi Siobhan. Thanks for the post. I broadly agree with the sentiment you express in it.

Forecasting is a dangerous activity, particularly because it is a fun, game-like activity that is nearly perfectly designed to be very attractive to EA/rationalist types because you get to be right when others are wrong, bet on your beliefs, and partake in the cultural practice.

I like bets involving donations, and investments as alternatives to forecasting without money on the line.

Hi Marcus. Thanks for the post. I broadly agree.

Coefficient Giving's (CG's) Forecasting Fund has recently been closed.

As of March 30, the Forecasting Fund is no longer active, though we continue to make key forecasting grants through other funds, such as Navigating Transformative AI. This page will be maintained until the end of 2026 as a record of the fund’s work.

I think this is more likely to make forecasting grants useful. They will presumably be assessed with the criteria used to evaluate the non-forecasting grants of the respective fund.

@NunoSempere wrote about the end of CG's Forecasting Fund in the last edition of the Forecasting Newsletter. Only paid subscribers can check the relevant section.

We are always in triage

Right.

That makes sense. Can I crosspost to the EA Forum arguments from Computational Functionalism Debate (linking to this post too)? I would like to share the Pen & Paper Argument, which is among the ones against CF which I find most persuasive.

I've also donated to the “inverts” project of RP!

Great. You may also be interested in donating to Arthropoda Foundation. I donated a few k$ to them last year. You are most likely aware of them. If relevant to readers, here is the post announcing their launch, and here is their post during the last Marginal Funding Week. They have been funding research informing how to increase the welfare of farmed arthropods, and "are particularly interested in research with a clear path to impact, whether by shaping future science or informing real-world decision-making".

Your updated estimates have huge credible intervals! What is the main source of the uncertainty in the model, or the main sources?

The ranges only account for uncertainty in the individual welfare per fully-healthy-animal-year. I consider this roughly proportional to the expected welfare range. So you can say I am just accounting for uncertainty in the expected welfare range. I believe this the overwhelming driver of the overall uncertainty. Here are the expected welfare ranges as a fraction of that of humans as a function of the exponent of the individual number of neurons.

The 90k (or 90.3k) figure is based on this sentence from this article

Got it. I consider the above ratio reasonable too, but my current best guess is much lower, as I commented above.

Thanks for sharing your thoughts, Derek. I did not know about your paper.

One would ideally recognise the existing uncertainty, and become less uncertain with further research. However, I think it is very easy to underestimate uncertainty. So I can easily see further research making one more uncertain. Yet, this would only lead to incomparability if one is open to impresice probabilities (I am not)?

Hi Dawn. Thanks for looking into soil invertebrates.

I recommend research on the welfare of soil invertebrates, and welfare comparisons across species. I do not recommend pursuing whatever land use changes seem to increase the welfare of soil invertebrates the most cost-effectively.

Part 3: Welfare Biology and AI: What We Can Do Now turns to interventions. I unpack Vasco Grilo’s finding that the effects of GiveWell’s top charities on soil animals are ~90,000× larger than their effects on humans

[...]

Most recently, Vasco Grilo’s series of posts on the EA Forum – “Cost-effectiveness accounting for soil nematodes, mites, and springtails” (June 2025), “Animal farming impacts soil nematodes, mites, and springtails hugely more than directly affected animals?” (June 2025), and “More animal farming increases animal welfare if soil animals have negative lives?” (October 2025) – has brought soil fauna into the center of EA cost-effectiveness debates. His key finding: the effects of almost any intervention on soil nematodes, mites, and springtails are orders of magnitude larger than the effects on the intervention’s intended beneficiaries. I’ll unpack his analysis in detail in parts 2 and 3.

This does not describe my current views well. I can see effects on soil invertebrates being anything from negligible to all that matters in terms of changes in welfare. Below are some numbers illustrating this.

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, I estimate that GiveWell's top charities change the welfare of soil ants, termites, springtails, mites, and nematodes 1.08*10^-5 to 10.9 billion times as much as they increase the welfare of humans. For my preferred exponent of 1, the change in the welfare on those soil invertebrates is 41.5 times as large as the increase in the welfare of humans (much less than the ratio of 90 k you mention in the quote above).

In addition, I have no idea about whether interventions targeting invertebrates increase the welfare of their target beneficiaries more or less cost-effectively than ones targeting humans. 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, I estimate that the Shrimp Welfare Project's (SWP's) Humane Slaughter Initiative (HSI) has increased the welfare of shrimps 1.68*10^-6 to 1.68 M times as cost-effectively as GiveWell's top charities increase the welfare of humans.

Rethink Priorities (RP) has done rigorous work on invertebrate sentience and moral weights, estimating welfare ranges for various species based on neuroscientific evidence. Their moral weight project produced estimates that are widely used in EA cost-effectiveness analyses – including the welfare range of 6.68 × 10⁻⁶ (relative to humans) that Vasco Grilo extrapolated for nematodes. He deferred the caculation to Gemini 2.5, so the derivation is opaque, and the result should not be taken literally.

To clarify, RP did not get welfare ranges for nematodes, and Gemini's guess that a modal soil nematode has 240 neurons has a negligible impact on my estimate for the expected welfare range of nematodes. Adult caenorhabditis elegans have 302 neurons, and the modal soil nematode has fewer neurons than this because new neurons are formed until adulthood. I do not know what is the exact number of neurons of random soil nematodes. However, the conclusions I would take from the calculations would be the same regardless of whether I used 302 or 30 neurons. So I just went with a guess from Gemini.

Hi Matthew.

Our actions have lots of unpredictable effects. If you drive to the store, you will delay everyone behind you in traffic. This will change when they next have sex, thus completely changing the identity of their future child. A different sperm and egg will fuse. This new child will go on to take a staggeringly large number of actions, each of which will change the identity of still more people. For this reason, because of your decision to drive to the store, the world hundreds of years down the line will be completely different.

I agree small actions like driving to the store may have large actual consequences. However, I believe their expected consequences are very small. I think the probability of any given child being born will be practically the same regardless of whether one drives to the store or not. One could tell a story where driving to the store leads to A being born instead of B. However, one could tell a story practically as convincing where driving to the store leads to B being born instead of A. So one should practically stick to the prior that driving to the store does change the probability of A or B being born. Likewise, driving to the store could cause a given hurricane H, but is almost as likely to prevent it. So the probability of hurricane H is practically the same regardless of whether one drives to the store or not.

@Derek Shiller, I would be curious to know your thoughts on the above.

Thanks, Spiarrow. I just set up a periodic reminder to check the post as a result of your comment.

Here is some more context about the exponent-based approach.