Memory is not an archive that happens to forget. It is a bounded system that preserves, compresses, reconstructs, and prioritizes information according to what it expects will matter.
I've been thinking about this with respect to any long form legal writing or long form writing in general. Unless you are writing long form mystery or fiction always open with an abstract of what is contained in the data, what is the question, what is the answer, how will the paper get there, why should the reader care? Then the conclusion having seen the evidence, what can be said with confidence, what remains uncertain, and what follows in practice, law, policy, theory, or further research?
A prospective map and a retrospective judgement.
This becomes structured data where the center hopefully does not get lost.
Pre-compression at the beginning. Then decompression of the data. Then recompression confirming the data in the conclusion.
In regards to mystery sometimes the pre-compression misdirects the reader from the answer. In dramatic fiction you might set the stakes without the details, or simply point to the path that the reader will follow. But for legal or academic papers I think pre-compression, decompression, and then recompression are what is expected.
For dealing with AI I may try that approach to setting up my queries. I'm not quite sure how the would work with coding or creating an AI agent, but I suspect it might be a helpful way to approach it.
Your “prospective map and retrospective judgement” distinction fits this surprisingly well. The opening compresses the path ahead; the body expands it; the conclusion compresses again, but now with the evidence behind it.
For AI, I’d treat that as a useful working pattern rather than a rule: put the task, constraints, and desired output up front, then restate the important state as the work changes.
For agents, I suspect the key is deciding what has to be carried forward as the context keeps changing.
That may be the reassuring version of it. Evolution gets somewhere remarkable without clean design, but it also has the luxury of enormous trial and error.
For AI, I’m less worried about elegance than about whether the mistakes stay recoverable while the system is learning.
That goes to resilience and adaptability. Poisoning the climate through exploitation and greed is still at the top of my list for things that would be very, very hard to recover from. AI is a distant second at best, and I don't think it's risen to the level of nuclear war. Even bio-war sets below that. So AI would have to be pretty darn bad to even crack the top three.
As a bit of an aside I wonder how these tasks that were set up to test agent alignment. Somewhere in the middle of the task did the important rules get lost? I have this suspicion that sloppy programming or vague programming might have compounded. My retired programmer friends seem to think the AI apocalypse will not be from malicious agents or evil programmers, but simply sloppy programming.
Between nuclear, war, biological pathogens, weather, earthquakes, meteors, etc. AI is just another on the list of potential ways for humanity to end. That doesn't mean don't take it seriously, but a judgement may need to be made. Does following profit only have to be the way this race is run?
Memory is not an archive that happens to forget. It is a bounded system that preserves, compresses, reconstructs, and prioritizes information according to what it expects will matter.
I've been thinking about this with respect to any long form legal writing or long form writing in general. Unless you are writing long form mystery or fiction always open with an abstract of what is contained in the data, what is the question, what is the answer, how will the paper get there, why should the reader care? Then the conclusion having seen the evidence, what can be said with confidence, what remains uncertain, and what follows in practice, law, policy, theory, or further research?
A prospective map and a retrospective judgement.
This becomes structured data where the center hopefully does not get lost.
Pre-compression at the beginning. Then decompression of the data. Then recompression confirming the data in the conclusion.
In regards to mystery sometimes the pre-compression misdirects the reader from the answer. In dramatic fiction you might set the stakes without the details, or simply point to the path that the reader will follow. But for legal or academic papers I think pre-compression, decompression, and then recompression are what is expected.
For dealing with AI I may try that approach to setting up my queries. I'm not quite sure how the would work with coding or creating an AI agent, but I suspect it might be a helpful way to approach it.
Your “prospective map and retrospective judgement” distinction fits this surprisingly well. The opening compresses the path ahead; the body expands it; the conclusion compresses again, but now with the evidence behind it.
For AI, I’d treat that as a useful working pattern rather than a rule: put the task, constraints, and desired output up front, then restate the important state as the work changes.
For agents, I suspect the key is deciding what has to be carried forward as the context keeps changing.
In retrospect to my question, evolution is a great example of sloppy programming that worked out. I think we will be fine.
That may be the reassuring version of it. Evolution gets somewhere remarkable without clean design, but it also has the luxury of enormous trial and error.
For AI, I’m less worried about elegance than about whether the mistakes stay recoverable while the system is learning.
That goes to resilience and adaptability. Poisoning the climate through exploitation and greed is still at the top of my list for things that would be very, very hard to recover from. AI is a distant second at best, and I don't think it's risen to the level of nuclear war. Even bio-war sets below that. So AI would have to be pretty darn bad to even crack the top three.
As a bit of an aside I wonder how these tasks that were set up to test agent alignment. Somewhere in the middle of the task did the important rules get lost? I have this suspicion that sloppy programming or vague programming might have compounded. My retired programmer friends seem to think the AI apocalypse will not be from malicious agents or evil programmers, but simply sloppy programming.
Between nuclear, war, biological pathogens, weather, earthquakes, meteors, etc. AI is just another on the list of potential ways for humanity to end. That doesn't mean don't take it seriously, but a judgement may need to be made. Does following profit only have to be the way this race is run?