This is a special post for quick takes by Ben Stewart. Only they can create top-level comments. Comments here also appear on the Quick Takes page and All Posts page.
I was excited by ForecastBench and FutureEval both projecting that LLMs would reach superforecaster parity by June 2027.But I didn't realise access to human crowd forecasts might be driving a lot of performance.If it is, that is massively disappointing.
The top LLM performers in ForecastBench have access to the crowd forecast (and it's not clear to me if FutureEval hides crowd forecasts - Metaculus did for the Quarterly Cup in 2025 but I couldn't find info about FutureEval). Skimming the literature with Claude, it seems like most studies either deliberately provide crowd forecasts or don't prevent searching for it, and those that hide it tend to have significantly worse results (still interesting, but less exciting).
To me, the potential wonders of LLM superforecasting is being able to get excellent guesses at any questions I might come up with. If I need to already have a human crowd or market forecast for the guess to be any good, then the kind of LLM superforecasting being projected is about 10% as useful to me. I still expect 'true' parity eventually, but it becomes a story of general timelines rather than empirical projection.
I don't know the field well, and I'm probably misunderstanding something. I'm posting this to find out I'm wrong. If I'm right, then it's worth dampening the expectations of anyone else who was imagining having an instant team of supers at their beck-and-call in ~14 months time.
On the recent post on Manifest, there’s been another instance of a large voting group (30-40ish [edit to clarify: 30-40ish karma, not 30-40ish individuals])arriving and downvoting any progressive-valenced comments (there were upvotes and downvotes prior to this, but in a more stochastic pattern). This is similar to what occured with the eugenics-related posts last year. Wanted to flag it to give a picture to later readers on the dynamics at play.
I would be surprised if it's 30-40 people. My guess is it's more like 5-6 people with reasonably high vote-strengths. Also, I highly doubt that the overall bias of the conversation here leans towards progressive-valenced comments being suppressed. EA is overwhelmingly progressive and has a pretty obvious anti-right bias (which like, I am a bit sympathetic to, but I feel like a warning in the opposite direction would be more appropriate)
My wording was imprecise - I meant 30-40ish in terms of karma. I agree the number of people is more likely to be 5-12. And my point is less about overall bias than just a particular voting dynamic - at first upvotes and downvotes occurring as is pretty typical, then a large and sudden influx of downvotes on everything from a particular camp.
There really should be a limit on the quantity of strong upvotes/downvotes one can deploy on comments to a particular post -- perhaps both "within a specific amount of time" and "in total." A voting group of ~half a dozen users should not be able to exert that much control over the karma distribution on a post. To be clear, I view (at least strong) targeted "downvoting [of] any progressive-valenced comments" as inconsistent with Forum voting norms.
At present, the only semi-practical fix would be for users on the other side of the debate to go back through the comments, guess which ones had been the targets of the voting group, and apply strong upvotes hoping to roughly neutralize the norm-breaking voting behavior of the voting group. Both the universe in which karma counts are corrupted by small voting groups and the universe in which karma counts are significantly determined by a clash between voting groups and self-appointed defenders seem really undesirable.
Thanks for the suggestion Jason! @JP Addison says that he forum-gated it at the time because he wanted to “see how it went over, whether they endorsed it on reflection. They previously wouldn’t have liked users treating votes as a scarce resource.” LW seems happy with how it’s gone, so we’ll go ahead and remove the forum-gating.
What can I read to understand the current and near-term state of drone warfare, especially (semi-)autonomous systems?
I'm looking for an overview of the developments in recent years, and what near-term systems are looking like. I've been enjoying Paul Scharre's 'Army of None', but given it was published in 2018 it's well behind the curve. Thanks!
I really enjoyed this 2022 paper by Rose Cao ("Multiple realizability and the spirit of functionalism"). A common intuition is that the brain is basically a big network of neurons with input on one side and all-or-nothing output on the other, and the rest of it (glia, metabolism, blood) is mainly keeping that network running. The paper's helpful for articulating how that model's impoverished, and argues that the right level for explaining brain activity (and resulting psychological states) might rely on the messy, complex, biological details, such that non-biological substrates for consciousness are implausible. (Some of those details: spatial and temporal determinants of activity, chemical transducers and signals beyond excitation/inhibition, self-modification, plasticity, glia, functional meshing with the physical body, multiplexed functions, generative entrenchment.) The argument doesn't necessarily oppose functionalism, but I think it's a healthy challenge to my previous confidence in multiple realisability within plausible limits of size, speed, and substrate. It's also useful to point to just how different artificial neural networks are from biological brains. This strengthens my feeling of the alien-ness of AI models, and updates me towards greater scepticism of digital sentience. I think the paper's a wonderful example of marrying deeply engaged philosophy with empirical reality.
I was excited by ForecastBench and FutureEval both projecting that LLMs would reach superforecaster parity by June 2027. But I didn't realise access to human crowd forecasts might be driving a lot of performance. If it is, that is massively disappointing.
The top LLM performers in ForecastBench have access to the crowd forecast (and it's not clear to me if FutureEval hides crowd forecasts - Metaculus did for the Quarterly Cup in 2025 but I couldn't find info about FutureEval). Skimming the literature with Claude, it seems like most studies either deliberately provide crowd forecasts or don't prevent searching for it, and those that hide it tend to have significantly worse results (still interesting, but less exciting).
To me, the potential wonders of LLM superforecasting is being able to get excellent guesses at any questions I might come up with. If I need to already have a human crowd or market forecast for the guess to be any good, then the kind of LLM superforecasting being projected is about 10% as useful to me. I still expect 'true' parity eventually, but it becomes a story of general timelines rather than empirical projection.
I don't know the field well, and I'm probably misunderstanding something. I'm posting this to find out I'm wrong. If I'm right, then it's worth dampening the expectations of anyone else who was imagining having an instant team of supers at their beck-and-call in ~14 months time.
On the recent post on Manifest, there’s been another instance of a large voting group (30-40ish [edit to clarify: 30-40ish karma, not 30-40ish individuals])arriving and downvoting any progressive-valenced comments (there were upvotes and downvotes prior to this, but in a more stochastic pattern). This is similar to what occured with the eugenics-related posts last year. Wanted to flag it to give a picture to later readers on the dynamics at play.
Manifold openly offered funding voting rings in their discord:
Just noting for anyone else reading the parent comment but not the screenshot, that said discussion was about Hacker News, not the EA Forum.
Also it was clearly not about Manifest. (Though it is nonetheless very cringe).
I would be surprised if it's 30-40 people. My guess is it's more like 5-6 people with reasonably high vote-strengths. Also, I highly doubt that the overall bias of the conversation here leans towards progressive-valenced comments being suppressed. EA is overwhelmingly progressive and has a pretty obvious anti-right bias (which like, I am a bit sympathetic to, but I feel like a warning in the opposite direction would be more appropriate)
My wording was imprecise - I meant 30-40ish in terms of karma. I agree the number of people is more likely to be 5-12. And my point is less about overall bias than just a particular voting dynamic - at first upvotes and downvotes occurring as is pretty typical, then a large and sudden influx of downvotes on everything from a particular camp.
There really should be a limit on the quantity of strong upvotes/downvotes one can deploy on comments to a particular post -- perhaps both "within a specific amount of time" and "in total." A voting group of ~half a dozen users should not be able to exert that much control over the karma distribution on a post. To be clear, I view (at least strong) targeted "downvoting [of] any progressive-valenced comments" as inconsistent with Forum voting norms.
At present, the only semi-practical fix would be for users on the other side of the debate to go back through the comments, guess which ones had been the targets of the voting group, and apply strong upvotes hoping to roughly neutralize the norm-breaking voting behavior of the voting group. Both the universe in which karma counts are corrupted by small voting groups and the universe in which karma counts are significantly determined by a clash between voting groups and self-appointed defenders seem really undesirable.
We implemented this on LessWrong! (indeed based on some of my own bad experiences with threads like this on the EA Forum)
The EA Forum decided to forum gate the relevant changes, but on LW people would indeed be prevented from voting like I think voting is happening here: https://github.com/ForumMagnum/ForumMagnum/commit/07e0754042f88e1bd002d68f5f2ab12f1f4d4908
Thanks for the suggestion Jason! @JP Addison says that he forum-gated it at the time because he wanted to “see how it went over, whether they endorsed it on reflection. They previously wouldn’t have liked users treating votes as a scarce resource.” LW seems happy with how it’s gone, so we’ll go ahead and remove the forum-gating.
What can I read to understand the current and near-term state of drone warfare, especially (semi-)autonomous systems?
I'm looking for an overview of the developments in recent years, and what near-term systems are looking like. I've been enjoying Paul Scharre's 'Army of None', but given it was published in 2018 it's well behind the curve. Thanks!
FWIW, CNAS (where Paul and I work) are continuing to put out work on drone warfare:
https://www.cnas.org/publications/reports/evolution-not-revolution
https://www.cnas.org/publications/reports/swarms-over-the-strait
https://www.cnas.org/publications/reports/countering-the-swarm
I have been recommended the Drone Ultimatum podcast, but haven't listened to any
I really enjoyed this 2022 paper by Rose Cao ("Multiple realizability and the spirit of functionalism"). A common intuition is that the brain is basically a big network of neurons with input on one side and all-or-nothing output on the other, and the rest of it (glia, metabolism, blood) is mainly keeping that network running.
The paper's helpful for articulating how that model's impoverished, and argues that the right level for explaining brain activity (and resulting psychological states) might rely on the messy, complex, biological details, such that non-biological substrates for consciousness are implausible. (Some of those details: spatial and temporal determinants of activity, chemical transducers and signals beyond excitation/inhibition, self-modification, plasticity, glia, functional meshing with the physical body, multiplexed functions, generative entrenchment.)
The argument doesn't necessarily oppose functionalism, but I think it's a healthy challenge to my previous confidence in multiple realisability within plausible limits of size, speed, and substrate. It's also useful to point to just how different artificial neural networks are from biological brains. This strengthens my feeling of the alien-ness of AI models, and updates me towards greater scepticism of digital sentience.
I think the paper's a wonderful example of marrying deeply engaged philosophy with empirical reality.