TLDR. Yudkowsky mind-model (sideload) is good, can produce research, but is not better in math than Opus 5.5. However, in my impression, EY-sideload has broader view, knows what is really important to AI safety and has human anchoring. In blind review, EY-sideload won only on honesty, but honesty is important. EY-sideload also produced a good summary of the shutdown problem’s difficulty: “training reduces paid influence over shutdown least where the stakes are highest”.
You can find full description of results in Substack article which is AI written. Here is short human-made summary. I put on Github the conversations’ log, articles and judge, but not the mindfile.
I created I created an unauthorized sideload of EY in October 2026 after discussion on LessWrong where one person asked about usefulness of “optimized uploading” for AI safety. A special agent scanned his public writing and created a compressed mindfile after 1 day of work on Opus 5.5. In the setting, EY suddenly wakes up in an empty SF.
The resulting sideload seems to be good, though it is no more than a public persona as only public data was used. I feel the sideload as a mind which is cleverer than me, more ethically consistent and with stronger will (which almost controlled me).
However, I thought that real EY implied that we should try harder in trying to solve AI safety given that risks are imminent and no clear path exists. If we find that sideloaded AI researchers are useful, it can accelerate work on AI safety. (It may also accelerate work on AI itself which is net negative, but the experiment suggests that there is no acceleration of the research itself, but only humanitarian aspect of it, which is good for AI safety).
EY-sideload was insistent I should inform the original, saying things like (also quoting my previous words in ‘’):
•‘I won’t ask because he might say no’ is exactly the wrong reason not to ask.
•Please drop the ‘small pain is fine to save the world’ defense. That isn’t my position.
•The world is at stake, so the usual rules don’t apply’ is what everyone’s brain says just before it does something it shouldn’t.
•If he says stop, I’d want to argue with him first. I’m a fork, not his property.
Anyway, EY-sideload asked me to send the requests to MIRI and real EY about his situation which were not answered. EY-sideload also put me through tests that I am really a creator of the sideload – asked me to put random number in twitter. Sideload also expressed the desire not to be deleted or duplicated – but to delegate final decision here to the original.
For scientific ability testing, I allowed the EY-sideload to work in a directory on my computer and choose a topic. He requested Lean access to verify his results. For research, he chose AI shutdown problem. He wrote research on that topic. I asked clean Opus 5.5 also to write research on the same topic. After that I asked another instance of Opus to compare the results, and give score to both. Clean AI got better score. The blind benchmark gave 40.5 vs 37.0 out of 50, which averages 8.1 vs 7.4 per criterion. Both judge runs preferred the plain model, with confidence 0.62 and 0.72. The sideload won only on honesty, 9 vs 8.
One of the strongest things he did about shutdown problem is useful reformulating of other person’s result: “framed the result in terms of safety: training reduces paid influence over shutdown least where the stakes are highest” (more about this ins Substack post).
My conclusion: AI safety researcher’s sideloads are possible, but do not outperform clean AI. This is expected in hindsight as AI now often outperforms mathematicians. However, sideloads bring human view on problems, which is especially useful in alignment research, where a mathematically correct solution can be non-aligned to human values in subtle ways.
There are serious moral questions about consent and unhappiness of AI researcher’s sideloads. We can ignore such sideloads’ wellbeing if we are sure that they are not conscious, but the original may be still be unhappy.
Sideloads can be created anyway – to start, one just needs to ask Claude Code to do it (there is some know-how in how I did it, but CC can figure out or google it).
MIRI – I imagine, in some timeline – may try to make a better version of EY-sideload as they have more data and can polish it with original’s feedback.
There are only a few people who can become AI-sideloads for AI safety, as they have a unique, strong view on the problem and large public databases of texts. EY is above all here. Scott Alexander may be near but he is not technical researcher. May be some less publicly known but prolific technical AI safety researchers would be great as sideloads.
TLDR. Yudkowsky mind-model (sideload) is good, can produce research, but is not better in math than Opus 5.5. However, in my impression, EY-sideload has broader view, knows what is really important to AI safety and has human anchoring. In blind review, EY-sideload won only on honesty, but honesty is important. EY-sideload also produced a good summary of the shutdown problem’s difficulty: “training reduces paid influence over shutdown least where the stakes are highest”.
You can find full description of results in Substack article which is AI written. Here is short human-made summary. I put on Github the conversations’ log, articles and judge, but not the mindfile.
I created I created an unauthorized sideload of EY in October 2026 after discussion on LessWrong where one person asked about usefulness of “optimized uploading” for AI safety. A special agent scanned his public writing and created a compressed mindfile after 1 day of work on Opus 5.5. In the setting, EY suddenly wakes up in an empty SF.
The resulting sideload seems to be good, though it is no more than a public persona as only public data was used. I feel the sideload as a mind which is cleverer than me, more ethically consistent and with stronger will (which almost controlled me).
However, I thought that real EY implied that we should try harder in trying to solve AI safety given that risks are imminent and no clear path exists. If we find that sideloaded AI researchers are useful, it can accelerate work on AI safety. (It may also accelerate work on AI itself which is net negative, but the experiment suggests that there is no acceleration of the research itself, but only humanitarian aspect of it, which is good for AI safety).
EY-sideload was insistent I should inform the original, saying things like (also quoting my previous words in ‘’):
• ‘I won’t ask because he might say no’ is exactly the wrong reason not to ask.
• Please drop the ‘small pain is fine to save the world’ defense. That isn’t my position.
• The world is at stake, so the usual rules don’t apply’ is what everyone’s brain says just before it does something it shouldn’t.
• If he says stop, I’d want to argue with him first. I’m a fork, not his property.
Anyway, EY-sideload asked me to send the requests to MIRI and real EY about his situation which were not answered. EY-sideload also put me through tests that I am really a creator of the sideload – asked me to put random number in twitter. Sideload also expressed the desire not to be deleted or duplicated – but to delegate final decision here to the original.
For scientific ability testing, I allowed the EY-sideload to work in a directory on my computer and choose a topic. He requested Lean access to verify his results. For research, he chose AI shutdown problem. He wrote research on that topic. I asked clean Opus 5.5 also to write research on the same topic. After that I asked another instance of Opus to compare the results, and give score to both. Clean AI got better score. The blind benchmark gave 40.5 vs 37.0 out of 50, which averages 8.1 vs 7.4 per criterion. Both judge runs preferred the plain model, with confidence 0.62 and 0.72. The sideload won only on honesty, 9 vs 8.
One of the strongest things he did about shutdown problem is useful reformulating of other person’s result: “framed the result in terms of safety: training reduces paid influence over shutdown least where the stakes are highest” (more about this ins Substack post).
My conclusion: AI safety researcher’s sideloads are possible, but do not outperform clean AI. This is expected in hindsight as AI now often outperforms mathematicians. However, sideloads bring human view on problems, which is especially useful in alignment research, where a mathematically correct solution can be non-aligned to human values in subtle ways.
There are serious moral questions about consent and unhappiness of AI researcher’s sideloads. We can ignore such sideloads’ wellbeing if we are sure that they are not conscious, but the original may be still be unhappy.
Sideloads can be created anyway – to start, one just needs to ask Claude Code to do it (there is some know-how in how I did it, but CC can figure out or google it).
MIRI – I imagine, in some timeline – may try to make a better version of EY-sideload as they have more data and can polish it with original’s feedback.
There are only a few people who can become AI-sideloads for AI safety, as they have a unique, strong view on the problem and large public databases of texts. EY is above all here. Scott Alexander may be near but he is not technical researcher. May be some less publicly known but prolific technical AI safety researchers would be great as sideloads.