{"items": [{"author": "Jan-Willem", "source_link": "https://plus.google.com/100580955183019057735", "anchor": "gp-1419793731812", "service": "gp", "text": "Moore's law slowing down suggests we'll have to have some breakthroughs in \nhow\n we do the computation as well as \nwhat\n we compute, which is what the IBM folks looking at neural computing are doing (and I think what David Dalrymple was doing the last I checked a while back). \u00a0But of course there are a lot of questions about whether the neural behavior selected (thresholds) even makes sense. \u00a0I'm assuming GPGPU computing is being used for this stuff now. \u00a0If it isn't, that's probably a serious mistake, but maybe the C. Elegans connectome is just too small to benefit from it.", "timestamp": 1419793731}, {"author": "Jeff Kaufman", "source_link": "https://www.facebook.com/jefftk/posts/704150818992?comment_id=10100102108700182", "anchor": "fb-10100102108700182", "service": "fb", "text": "Raymond asked whether anything has changed in the last ~5y here, and I found \"Why is There No Successful Whole Brain Simulation (Yet)?\" [1] which just came out. While I've only skimmed it and its reference list, if there had been something new here I think they would have cited it.
I think we're still stuck on both (a) we can't read weights from real worms (and so can only model a generic worm) and (b) we don't understand how weights are changed in real worms (and so can't model learning).
[1] sci-hub.tw/10.1007/s13752-019-00319-5", "timestamp": "1563213198"}, {"author": "Jess", "source_link": "https://www.facebook.com/jefftk/posts/704150818992?comment_id=10100102108700182&reply_comment_id=10100102120895742", "anchor": "fb-10100102108700182_10100102120895742", "service": "fb", "text": "→ Great find, thanks Jeff.", "timestamp": "1563218720"}, {"author": "Danner", "source_link": "https://www.facebook.com/jefftk/posts/704150818992?comment_id=10100102251518972", "anchor": "fb-10100102251518972", "service": "fb", "text": "I like the process in westworld, where they run the human through a calibration, then run a series of robots through the calibration until they react the same way. If the robot can get through the calibration the same way as the human, they call the robot indistinguishable from the human.
Of course, it fails, but I like the idea.
Take a worm, run it through tests. This defines it's neural weights. Run simulations through tests till you find the one that responds and changes in the same way, and you've found that worm. You'd also be finding most of the other worms too, but that's how permutations work, right?", "timestamp": "1563287665"}, {"author": "Jeff Kaufman", "source_link": "https://www.facebook.com/jefftk/posts/704150818992?comment_id=10100102251518972&reply_comment_id=10100102255760472", "anchor": "fb-10100102251518972_10100102255760472", "service": "fb", "text": "→ That sort of brute force approach is maybe workable for something with as few weights as c elegans, but can't possibly determine the weights for something as complex as a human.
A rough model is that to identify N bits you need N tests. There is at least one bit of weight information per neuron, which gives a lower bound of 302 tests for c elegans but 100B for a human.", "timestamp": "1563289888"}, {"author": "Danner", "source_link": "https://www.facebook.com/jefftk/posts/704150818992?comment_id=10100102251518972&reply_comment_id=10100102256014962", "anchor": "fb-10100102251518972_10100102256014962", "service": "fb", "text": "→ I assume there's some smart trimming of unlikely data points. Honestly, 100B? Not so bad?", "timestamp": "1563290078"}, {"author": "Jeff Kaufman", "source_link": "https://www.facebook.com/jefftk/posts/704150818992?comment_id=10100102251518972&reply_comment_id=10100102261254462", "anchor": "fb-10100102251518972_10100102261254462", "service": "fb", "text": "→ 100B is a very simple lower bound; weights are a pairwise property of neurons, so it's really far high.
And I don't see how you would run 100B tests on a person?", "timestamp": "1563293474"}, {"author": "Danner", "source_link": "https://www.facebook.com/jefftk/posts/704150818992?comment_id=10100102251518972&reply_comment_id=10100102269497942", "anchor": "fb-10100102251518972_10100102269497942", "service": "fb", "text": "→ Jeff Kaufman hm, I guess I was imagining a single(or small number of) complex scenario for the human, which is then used to make many many tests for the simulations. Like having a racecar run a loop around a track, then generating simulations that try to copy the path of that car exactly. Are you saying that the track loop only denotes one test?", "timestamp": "1563297855"}, {"author": "Jeff Kaufman", "source_link": "https://www.facebook.com/jefftk/posts/704150818992?comment_id=10100102251518972&reply_comment_id=10100102269962012", "anchor": "fb-10100102251518972_10100102269962012", "service": "fb", "text": "→ Danner tests can give multiple bits of information, yes, but we have so many unrelated components that you'd still need an enormous number of tests. Consider all the things you've encountered over your life that you can recall given an appropriate prompt. Sufficient prompts to collect all that data would take more than the rest of your life.", "timestamp": "1563298217"}, {"author": "Danner", "source_link": "https://www.facebook.com/jefftk/posts/704150818992?comment_id=10100102251518972&reply_comment_id=10100102270431072", "anchor": "fb-10100102251518972_10100102270431072", "service": "fb", "text": "→ I believe there's a smaller number of values and weights that actually influence a person's behavior, people forget and rearrange their memories and experiences to streamline them.
I do agree that I have no idea what those things are, and that they are probably all different per person, those \"formative memories\", so yes, it's still absurd to try this approach, but ... I think we need to get close enough that if \"Jeff simulation 1\" stated that his favorite vegetable was corn, everyone would believe it and it wouldn't matter, even if the real Jeff would be absolutely disgusted with that choice. The important bit would be when someone confronted you with \"but you hate corn, remember that picnic?\" that your simulation would handle it like you would.", "timestamp": "1563298726"}, {"author": "Jeff Kaufman", "source_link": "https://www.facebook.com/jefftk/posts/704150818992?comment_id=10100102251518972&reply_comment_id=10100102272427072", "anchor": "fb-10100102251518972_10100102272427072", "service": "fb", "text": "→ What about \"that error message looks like one I saw once when credentials were corrupted\" sorts of memories? We have tons of memories like this which normally are completely silent but are super useful when they apply.", "timestamp": "1563300292"}, {"author": "Danner", "source_link": "https://www.facebook.com/jefftk/posts/704150818992?comment_id=10100102251518972&reply_comment_id=10100102277566772", "anchor": "fb-10100102251518972_10100102277566772", "service": "fb", "text": "→ I guess it depends how vivid those memories are. How much does that change you as a person, and how many times do you remember correctly, vs getting a suspicious feeling and having a sense/hunch about where to look? I guess in that situation I'm expecting a simulation to get a percentage of hits on what to do when an error message is found, but expecting a person to remember one specific one of these messages is impressive. Sometimes I can remember what I ate for lunch the day I solved an error if I think hard enough. Often it doesn't matter though.
But no, you're right, What I'm asking is definitely to have a higher level of intellect available, which then chooses how often \"Jeff simulation 1\" uses that higher intellect to simulate the real Jeff.", "timestamp": "1563302615"}, {"author": "Jeff Kaufman", "source_link": "https://www.facebook.com/jefftk/posts/704150818992?comment_id=10100102251518972&reply_comment_id=10100102294832172", "anchor": "fb-10100102251518972_10100102294832172", "service": "fb", "text": "→ Danner I don't think vividity is the issue. A lot of how I respond to things will have to do with how previous things went. A version of me without my memory and things I've learned would be very different. While you might be able to extract this information from someone's brain through destructive scanning in the far future, I don't think you can get anywhere near enough of it through behavioral tests.", "timestamp": "1563311182"}, {"author": "Danner", "source_link": "https://www.facebook.com/jefftk/posts/704150818992?comment_id=10100102251518972&reply_comment_id=10100102296089652", "anchor": "fb-10100102251518972_10100102296089652", "service": "fb", "text": "→ I don't think you'll be the first person to be recreated. Many people are much less based on singular memory recall. I guess I'm asking and answering a different question.", "timestamp": "1563311811"}, {"author": "Jeff Kaufman", "source_link": "https://www.facebook.com/jefftk/posts/704150818992?comment_id=10100102251518972&reply_comment_id=10100102307586612", "anchor": "fb-10100102251518972_10100102307586612", "service": "fb", "text": "→ Danner I guess I'm surprised you don't think most people use their memories much!? Every time you do something you've done before, navigate, have a conversation with someone, that's informed by memories you have", "timestamp": "1563313689"}, {"author": "Danner", "source_link": "https://www.facebook.com/jefftk/posts/704150818992?comment_id=10100102251518972&reply_comment_id=10100102315904942", "anchor": "fb-10100102251518972_10100102315904942", "service": "fb", "text": "→ Yes, but I was surprised to learn that people don't remember where they saw a thing as often as you think. For me it was formative when I 1) Realized that synthesis was a valid form of creativity and 2) realized that people who come up with ideas \"out of the blue\" often have seen things that inspired their ideas, but then forgot where they saw it.
I remember where I saw things, so my remixes can be cited to their source. Many other folks just forget, so can't cite their sources.
Also, listening to people tell a story, then finding out that the incident was recorded and didn't play out how they said leads me to not trust people's memory much. I must include myself in this as well, I'm as unreliable a narrator as anyone.
From that, I think for day to day personas, there are many less memories accessed than we've got, and many memories can be made up and still seem to be right, as long as they don't clash with reality. The trick is also modeling how these clashes with reality get resolved, each person has ways of doing that.", "timestamp": "1563316027"}, {"author": "Jeff Kaufman", "source_link": "https://www.facebook.com/jefftk/posts/704150818992?comment_id=10100102251518972&reply_comment_id=10100102322516692", "anchor": "fb-10100102251518972_10100102322516692", "service": "fb", "text": "→ Danner even though people's memories are messy and inaccurate, and we don't track provenance well, that doesn't mean that what's going on inside is simple or easy to reverse engineer from behavior.", "timestamp": "1563319360"}, {"author": "Danner", "source_link": "https://www.facebook.com/jefftk/posts/704150818992?comment_id=10100102251518972&reply_comment_id=10100102371124282", "anchor": "fb-10100102251518972_10100102371124282", "service": "fb", "text": "→ I totally agree.", "timestamp": "1563367328"}, {"author": "Jeff Kaufman", "source_link": "https://www.facebook.com/jefftk/posts/704150818992?comment_id=10100106633567312", "anchor": "fb-10100106633567312", "service": "fb", "text": "The Drosophila (fruit fly) connectome is now complete: https://www.biorxiv.org/.../early/2019/04/11/605634.full.pdf", "timestamp": "1565894884"}]}