Why did the EA organizations find it disappointing? Iâm afraid youâre going to say they didnât like that peer reviewers didnât agree with them, and therefore they decided the peer reviewers were wrong, and peer review is a waste of time.
Not all EA organizations are consistently producing high-quality work. Thatâs part of the problem. For instance, the problems with the METR time horizons graph are numerous and severe. Many of them were entirely avoidable, and should have at least been better disclosed. I canât get over that most of the longer tasks, on which the 2025 segment of the graph depends, donât have empirically measured human baselines. The baselines are just guesses by the authors. Surely if you donât even bother to measure data that doesnât qualify as high-quality? This also wasnât disclosed until 2026 â a major omission.
What I would recommend to people at this point is to not believe any of METRâs claims, research, or analysis going forward unless and until it can be independently verified by a reliable source. You donât know if METRâs data is data or just guesstimates. You donât know that the typical best practices of scientific research have been followed. You donât know that flaws or shortcomings or limitations that METR is aware of will be disclosed with sufficient emphasis, consistently across all communications.
Very few people outside of EA consider EAâs idiosyncratic ideas to be serious and credible. What is the strategy for gaining credibility outside the EA echo chamber? Right now, it seems to be a media strategy that counts on people not fact checking EAâs messaging. This could work â a lot of misinformation misleads a lot of people a lot of the time â but it also might rightly damage EAâs reputation if people eventually learn EA is not telling them the truth. Itâs a risky strategy that depends on being able to fool people, rather than intellectually convince them.
80,000 Hoursâ abysmal video on AI 2027 is an example of this. It misinforms its audience about AI expertsâ views and insinuates there is a consensus in support of AI 2027âs core claims that doesnât exist. Either 80,000 Hours knew this and misled its audience anyway, or it didnât do a proper fact check of its script before producing the video. I was a lifelong fan of 80,000 Hours until that video. Now I no longer trust 80,000 Hours about anything. Not even career advice. I was in the top 1% or 0.1% of biggest supporters of 80,000 Hours. Now Iâve been completely polarized in the opposite direction. This is anecdotal, but, also, most people become angry when they feel as if theyâve been misled. Itâs not a stretch to think this strategy could really blow up in an ugly way.
In my opinion, EA is aggressively burning down its reputation and risks being correctly labelled as a purveyor of misinformation. Steps should be taken to at least stop the bleeding.
I donât know for sure that peer review would help move idiosyncratic EA ideas outside the EA echo chamber. I also donât know that there isnât a better strategy for doing so. It just seems like a good idea to me.
The economist Tyler Cowen was actually the first person who I heard suggest this. I believe he was talking about AGI/âAGI safety. It was on a podcast, either his or someone elseâs. I remember he said: publish, publish, publish.
The provider I originally mentioned in this post definitely looks like a shady company that I definitely wouldnât recommend. I was wrong to mention that company and, in retrospect, the signs were obvious that it wasnât a trustworthy company. I only gave it a few cursory glances. Iâm grateful to Clara for giving it a second look and realizing that both Google Gemini 3.1 Pro (with âExtended thinkingâ) and I had been duped by some devious SEO.
The trustworthiness of that provider â or indeed any similar company offering a convenient, off-the-shelf service â is beside the point of whether peer review is a good idea or not. There are scam companies selling fake Ozempic online. That has nothing to do with whether genuine Ozempic is a good drug or not.
Why did the EA organizations find it disappointing? Iâm afraid youâre going to say they didnât like that peer reviewers didnât agree with them
Nope. There have been a variety of issues. One is speed, and another is the difficulty of finding relevant experts. Thinking back to MIRIâs experience with the Damascus paper, my recollection (possibly incorrect) is their final conclusion was the getting published in a good journal took a lot of time, didnât really improve the fundamental quality of the work much, and also didnât yield a lot of prestige/âoutreach benefits.
Very few people outside of EA consider EAâs idiosyncratic ideas to be serious and credible. What is the strategy for gaining credibility outside the EA echo chamber? Right now, it seems to be a media strategy that counts on people not fact checking EAâs messaging.
Come on, I understand you have objections to METRâs methodologyâthough to my knowledge you have not published those objections in a peer-reviewed journalâbut blithely accusing them of a deliberate strategy of misinformation seems low.
Why did the EA organizations find it disappointing? Iâm afraid youâre going to say they didnât like that peer reviewers didnât agree with them, and therefore they decided the peer reviewers were wrong, and peer review is a waste of time.
Not all EA organizations are consistently producing high-quality work. Thatâs part of the problem. For instance, the problems with the METR time horizons graph are numerous and severe. Many of them were entirely avoidable, and should have at least been better disclosed. I canât get over that most of the longer tasks, on which the 2025 segment of the graph depends, donât have empirically measured human baselines. The baselines are just guesses by the authors. Surely if you donât even bother to measure data that doesnât qualify as high-quality? This also wasnât disclosed until 2026 â a major omission.
What I would recommend to people at this point is to not believe any of METRâs claims, research, or analysis going forward unless and until it can be independently verified by a reliable source. You donât know if METRâs data is data or just guesstimates. You donât know that the typical best practices of scientific research have been followed. You donât know that flaws or shortcomings or limitations that METR is aware of will be disclosed with sufficient emphasis, consistently across all communications.
Very few people outside of EA consider EAâs idiosyncratic ideas to be serious and credible. What is the strategy for gaining credibility outside the EA echo chamber? Right now, it seems to be a media strategy that counts on people not fact checking EAâs messaging. This could work â a lot of misinformation misleads a lot of people a lot of the time â but it also might rightly damage EAâs reputation if people eventually learn EA is not telling them the truth. Itâs a risky strategy that depends on being able to fool people, rather than intellectually convince them.
80,000 Hoursâ abysmal video on AI 2027 is an example of this. It misinforms its audience about AI expertsâ views and insinuates there is a consensus in support of AI 2027âs core claims that doesnât exist. Either 80,000 Hours knew this and misled its audience anyway, or it didnât do a proper fact check of its script before producing the video. I was a lifelong fan of 80,000 Hours until that video. Now I no longer trust 80,000 Hours about anything. Not even career advice. I was in the top 1% or 0.1% of biggest supporters of 80,000 Hours. Now Iâve been completely polarized in the opposite direction. This is anecdotal, but, also, most people become angry when they feel as if theyâve been misled. Itâs not a stretch to think this strategy could really blow up in an ugly way.
In my opinion, EA is aggressively burning down its reputation and risks being correctly labelled as a purveyor of misinformation. Steps should be taken to at least stop the bleeding.
I donât know for sure that peer review would help move idiosyncratic EA ideas outside the EA echo chamber. I also donât know that there isnât a better strategy for doing so. It just seems like a good idea to me.
The economist Tyler Cowen was actually the first person who I heard suggest this. I believe he was talking about AGI/âAGI safety. It was on a podcast, either his or someone elseâs. I remember he said: publish, publish, publish.
The provider I originally mentioned in this post definitely looks like a shady company that I definitely wouldnât recommend. I was wrong to mention that company and, in retrospect, the signs were obvious that it wasnât a trustworthy company. I only gave it a few cursory glances. Iâm grateful to Clara for giving it a second look and realizing that both Google Gemini 3.1 Pro (with âExtended thinkingâ) and I had been duped by some devious SEO.
The trustworthiness of that provider â or indeed any similar company offering a convenient, off-the-shelf service â is beside the point of whether peer review is a good idea or not. There are scam companies selling fake Ozempic online. That has nothing to do with whether genuine Ozempic is a good drug or not.
Nope. There have been a variety of issues. One is speed, and another is the difficulty of finding relevant experts. Thinking back to MIRIâs experience with the Damascus paper, my recollection (possibly incorrect) is their final conclusion was the getting published in a good journal took a lot of time, didnât really improve the fundamental quality of the work much, and also didnât yield a lot of prestige/âoutreach benefits.
Come on, I understand you have objections to METRâs methodologyâthough to my knowledge you have not published those objections in a peer-reviewed journalâbut blithely accusing them of a deliberate strategy of misinformation seems low.