davidshapiro_youtube_transcripts
/
ChatGPT4 Sparse Priming Representations Hierarchical Memory Consolidation and Implied Cognition_transcript.csv
text,start,duration | |
hey everybody David Shapiro here with,0.62,3.76 | |
video,3.12,3.06 | |
um I'm actually pretty tired so today's,4.38,4.92 | |
video will be short but this is some,6.18,5.339 | |
really important information to share so,9.3,5.219 | |
basically what I wanted to do was share,11.519,5.581 | |
with you three new repos that I just,14.519,4.141 | |
published,17.1,3.42 | |
um so one is on sparse priming,18.66,4.98 | |
representations I realized that just a,20.52,5.099 | |
YouTube video with a transcript is,23.64,5.1 | |
probably not enough so I have this out,25.619,5.761 | |
here it's just a um,28.74,6.18 | |
a very uh high level overview with a few,31.38,5.519 | |
examples of what the sparse priming,34.92,4.74 | |
representations are so for instance I I,36.899,6.66 | |
had I had it write an spr of spr's so,39.66,5.76 | |
sparse priming representation concise,43.559,4.081 | |
context driven memory summaries enables,45.42,5.22 | |
smes or llms to reconstruct ideas short,47.64,4.759 | |
complete sentences provide context,50.64,3.96 | |
effective for memory organization and,52.399,4.66 | |
retrieval reduces information to,54.6,4.799 | |
Essential Elements facilitates quick,57.059,3.66 | |
understanding and recall designed to,59.399,3.421 | |
mimic human memory structure so just,60.719,6.361 | |
with this short eight uh eight lines of,62.82,5.82 | |
assertions or statements you probably,67.08,5.88 | |
get a pretty good idea of what an spr is,68.64,6.0 | |
um so that's an example of an spr and,72.96,4.5 | |
here is the hierarchical hierarchical,74.64,5.519 | |
memory consolidation system which is the,77.46,5.58 | |
autonomous cognitive entity uh memory,80.159,4.32 | |
system that I've been working on and,83.04,3.24 | |
it's 11 lines,84.479,3.96 | |
um but again it I won't read the whole,86.28,4.32 | |
thing to you but you get the idea so,88.439,5.22 | |
here is here is this um if anyone wants,90.6,4.559 | |
to use this and adapt this to an actual,93.659,4.14 | |
paper feel free it's um all published,95.159,4.981 | |
under the MIT license so this is free,97.799,4.201 | |
for the world,100.14,5.96 | |
um so that's the spr uh repo slash paper,102.0,6.78 | |
next is the hierarchical memory,106.1,4.839 | |
consolidation system I talked about this,108.78,4.199 | |
and I showed you guys that I had a,110.939,4.621 | |
really long chat conversation I'm going,112.979,5.161 | |
with chat GPT and so what I realized is,115.56,4.379 | |
one you guys can't read through this,118.14,4.619 | |
conversation and two again just a video,119.939,5.401 | |
is probably not enough,122.759,4.5 | |
um so I showed you guys this,125.34,5.04 | |
conversation before but here I've got,127.259,5.761 | |
um the most Salient bits,130.38,5.579 | |
um held out here I probably will try and,133.02,5.04 | |
get a little bit more information,135.959,4.321 | |
um because it gives you an overview,138.06,3.42 | |
um it gives you some of the theory and,140.28,3.179 | |
reasoning and it tells you the basics of,141.48,3.899 | |
how to implement it but I don't have any,143.459,3.901 | |
examples yet,145.379,5.881 | |
um so it might be more difficult but uh,147.36,5.76 | |
one of the reasons that I don't have,151.26,3.6 | |
examples is because I haven't fully,153.12,3.3 | |
implemented this yet,154.86,5.459 | |
um but it is here in theory also hmcs is,156.42,6.3 | |
not the easiest thing to say and if,160.319,5.041 | |
we've learned anything from chat GPT is,162.72,5.099 | |
that naming something that's easier to,165.36,4.44 | |
say is better so we might choose,167.819,4.801 | |
different names AKA I like this adaptive,169.8,5.34 | |
knowledge archive rolling episodic,172.62,4.199 | |
memory organizer this is actually like,175.14,5.28 | |
the most on the nose so Remo so we might,176.819,5.341 | |
call this Remo who knows,180.42,6.06 | |
anyways it's a good uh it's a good start,182.16,6.18 | |
um to this oh one thing that I did want,186.48,3.24 | |
to say is that I've got the discussions,188.34,4.2 | |
enabled for all of these,189.72,5.4 | |
um because these concepts are really,192.54,4.44 | |
important and really critical,195.12,3.18 | |
um and so we can discuss them on Reddit,196.98,3.78 | |
as well but you can also discuss them,198.3,4.56 | |
directly on GitHub if you'd like and,200.76,3.78 | |
then finally this is the most exciting,202.86,5.64 | |
one so uh four weeks ago the paper about,204.54,6.24 | |
large language model theory of Mind came,208.5,4.319 | |
out and since then a lot of people have,210.78,4.92 | |
been using Bing and chat gbt and gpt4 to,212.819,5.28 | |
do experiments with theory of mind and,215.7,5.22 | |
one thing occurred to me when I was,218.099,4.14 | |
working on sparse priming,220.92,3.12 | |
representations the idea that there's,222.239,4.56 | |
enough cognition going on inside the,224.04,5.339 | |
model to reconstruct something I,226.799,4.321 | |
realized that what I was banking on was,229.379,5.341 | |
implied cognition and so I just spent,231.12,7.02 | |
some time with chat gpt4 to articulate,234.72,5.76 | |
implied cognition and start to come up,238.14,4.92 | |
with some tests for it so I've got the,240.48,4.679 | |
full transcript of the whole,243.06,3.72 | |
conversation in here and you can read it,245.159,3.66 | |
it's pretty impressive,246.78,4.62 | |
um what what it was able to do so one of,248.819,4.14 | |
the biggest highlights of this,251.4,3.479 | |
conversation was as I was talking with,252.959,4.321 | |
chat GPT I said okay look over this,254.879,4.681 | |
conversation and look for evidence of,257.28,4.079 | |
implied cognition and it was able to,259.56,4.079 | |
look back through the conversation and,261.359,4.261 | |
give me evidence of its own implied,263.639,5.0 | |
cognition and even how to test itself,265.62,5.22 | |
and not only that but it did it much,268.639,3.761 | |
faster than a human could do it so it's,270.84,3.26 | |
like all right this is,272.4,4.64 | |
basically you know we're bordering on,274.1,5.02 | |
metacognitive abilities and we even,277.04,3.879 | |
addressed that as well I said how do how,279.12,4.1 | |
will we discern the difference between,280.919,3.901 | |
self-explication like true,283.22,4.539 | |
self-explication versus confabulation,284.82,5.34 | |
and it has some ideas on that too some,287.759,4.861 | |
testable hypotheses,290.16,4.259 | |
um so that's all here this is Far and,292.62,3.84 | |
Away the most interesting thing,294.419,3.961 | |
um that I was working on today,296.46,3.84 | |
um and then uh finally at the very end,298.38,4.379 | |
of the conversation I asked chat GPT if,300.3,4.14 | |
there was anything that it wanted me to,302.759,3.361 | |
document and share with the world and,304.44,4.92 | |
this is Verbatim what it said,306.12,5.7 | |
um about you know its own perspective on,309.36,5.88 | |
this and then desires moving forward,311.82,5.34 | |
um but yeah so what I wanted to do is I,315.24,3.0 | |
wanted to actually show you this,317.16,2.52 | |
conversation so you know I didn't just,318.24,4.62 | |
make this up this is this is right in,319.68,5.82 | |
chat GPT so I talked about theory of,322.86,4.02 | |
mind,325.5,2.52 | |
um I asked do you have any questions,326.88,3.24 | |
about what we're talking about,328.02,3.84 | |
um and it asks for clarification on,330.12,4.139 | |
sparse priming representations implied,331.86,6.119 | |
cognition so that was already evidence,334.259,5.581 | |
of implied cognition because it was,337.979,4.44 | |
aware of what it didn't know it was able,339.84,3.84 | |
to say I'm actually not sure what you're,342.419,4.321 | |
talking about and so by by virtue of,343.68,5.9 | |
chat gbt recognizing novel information,346.74,6.36 | |
that implies some kind of cognition and,349.58,4.72 | |
I don't mean cognition like human,353.1,3.36 | |
cognition and that's why I have this,354.3,5.16 | |
labeled implied cognition or you could,356.46,5.04 | |
even call it a facsimile of cognition,359.46,4.739 | |
and so then I gave it an example of an,361.5,4.8 | |
spr so I said sure for your first,364.199,5.041 | |
question here's an example of an spr,366.3,4.619 | |
um I said given that list of statements,369.24,3.899 | |
you can imagine what the concept is and,370.919,3.421 | |
unpack it does that make sense,373.139,3.661 | |
furthermore you can J you can even,374.34,4.02 | |
generate the highly Salient questions,376.8,4.2 | |
above implies a lot of cognition so I,378.36,4.26 | |
already recognize the fact that it can,381.0,4.62 | |
generate relevant questions implies some,382.62,5.579 | |
level of cognition,385.62,5.46 | |
um so it was it was uh it was okay with,388.199,4.741 | |
that and then it came up with some some,391.08,4.44 | |
initial tests so ask for logical,392.94,4.14 | |
reasoning understanding ambiguity,395.52,3.78 | |
generating relevant questions counter,397.08,4.26 | |
factual thinking so on and so forth and,399.3,3.48 | |
then that's when I as I was reading this,401.34,4.199 | |
I was like oh yeah self-explication,402.78,5.52 | |
um so the ability to explain itself uh,405.539,8.041 | |
plausibly is another uh potential,408.3,8.76 | |
um aspect of implied cognition,413.58,5.459 | |
um so then I asked it to analyze the,417.06,3.359 | |
conversation instead analyzing our,419.039,2.88 | |
conversation I can't identify a few,420.419,3.84 | |
instances wherein where implied,421.919,4.921 | |
cognition might be at play and so then,424.259,5.28 | |
it says context awareness so basically,426.84,4.62 | |
reading the context of the conversation,429.539,5.22 | |
it is able to understand what it means,431.46,6.299 | |
but also it's able to infer a lot about,434.759,4.921 | |
what's going on just by virtue of,437.759,3.961 | |
looking at the language,439.68,4.38 | |
um it adapts its communication now this,441.72,4.5 | |
is in part due to the fact that this is,444.06,4.44 | |
how the model is trained,446.22,4.68 | |
um but the fact of the matter is it does,448.5,4.08 | |
adapt its communication depending on,450.9,4.38 | |
what I'm trying to do conceptual,452.58,6.119 | |
integration so this is actually,455.28,6.06 | |
probably the most one of the these last,458.699,5.461 | |
two are the most important because not,461.34,5.04 | |
only was it able to really quickly,464.16,4.92 | |
understand the concept of spr and,466.38,5.159 | |
implied cognition it was then able to,469.08,5.82 | |
use it and synthesize more and build on,471.539,6.06 | |
it so the ability to to use novel,474.9,6.419 | |
information is the essence of fluid,477.599,6.121 | |
intelligence which up until now only,481.319,5.1 | |
humans have been capable of so just the,483.72,4.62 | |
fact that it is able to to recognize,486.419,4.321 | |
novel information and use it this,488.34,5.34 | |
quickly implies a lot,490.74,6.66 | |
so moving forward that's when I ask okay,493.68,5.04 | |
how do we discern the difference between,497.4,3.479 | |
self-expo explication and confabulation,498.72,5.34 | |
and then on on on being goal oriented,500.879,6.0 | |
that reminded me that goal tracking and,504.06,5.579 | |
figuring out one where you are,506.879,4.38 | |
um in terms of solving a problem,509.639,3.541 | |
figuring out where you need to go and,511.259,3.66 | |
measuring how close you are to solving,513.18,4.44 | |
that goal this is part of executive,514.919,4.98 | |
function and cognitive control and so I,517.62,4.74 | |
decided to just throw in a test for goal,519.899,5.221 | |
tracking as we go so then it came up,522.36,4.979 | |
with some really good ideas about,525.12,2.76 | |
um,527.339,3.541 | |
about uh testing for self-explication,527.88,5.579 | |
versus confabulation so checking for,530.88,4.74 | |
consistency over time external,533.459,5.041 | |
validation such as using another system,535.62,5.04 | |
um and then probing questions so ask,538.5,4.32 | |
follow-up questions or so on and again,540.66,4.08 | |
humans are not really capable of,542.82,4.92 | |
self-explication anyways we confabulate,544.74,6.06 | |
our reasoning post facto by and large,547.74,5.76 | |
unless we are very explicit when we,550.8,4.44 | |
bring an unconscious thought to,553.5,3.6 | |
Consciousness and we talk through I'm,555.24,4.5 | |
going to do this because and even then,557.1,4.62 | |
the reasoning that you think that you,559.74,4.5 | |
used is still not gonna going to be a,561.72,4.26 | |
hundred percent accurate,564.24,3.24 | |
um just ask any psychologist or,565.98,3.539 | |
philosopher we think we know why we do,567.48,4.799 | |
stuff but we really don't and so I even,569.519,4.32 | |
point that out in the conversation that,572.279,3.18 | |
self-explication is not something that,573.839,4.201 | |
humans are even really capable of,575.459,5.281 | |
um so having a expecting a machine to be,578.04,4.859 | |
fully capable of self-explication is,580.74,3.9 | |
kind of a moot point,582.899,4.44 | |
um but it still has some good ideas and,584.64,4.74 | |
of course explain ability and AI is,587.339,3.661 | |
important,589.38,4.98 | |
um and then it goes and unpacks a lot,591.0,7.74 | |
about goal tracking so it says um to,594.36,5.7 | |
evaluate our progress towards achieving,598.74,2.82 | |
our goals let's recap the goals that,600.06,3.06 | |
you've articulated,601.56,3.06 | |
um develop the concept of implied,603.12,3.659 | |
cognition establish tests and create,604.62,4.14 | |
criterias and protocol for using implied,606.779,4.441 | |
cognition progress so I was able to say,608.76,4.86 | |
okay this is what we've achieved,611.22,4.2 | |
we have so we have proposed several,613.62,3.98 | |
tests we have yet to create criteria,615.42,4.919 | |
next steps further refine the concept,617.6,4.6 | |
develop proposed test begin creating,620.339,4.861 | |
criteria and protocols for using implied,622.2,4.44 | |
cognitions,625.2,3.84 | |
um and we actually didn't get this far,626.64,4.74 | |
um but still the fact of the matter is,629.04,4.5 | |
it was able to kind of keep track of our,631.38,3.66 | |
goals,633.54,2.34 | |
um,635.04,3.6 | |
and then I made the observation that on,635.88,5.1 | |
conceptual integration it was,638.64,3.5 | |
incorporating new information,640.98,3.84 | |
synthesizing new ideas and Novelty so,642.14,5.439 | |
then I asked about novelty so,644.82,5.519 | |
recognizing new information is actually,647.579,6.361 | |
uh creates really distinctive patterns,650.339,5.701 | |
in the human brain and so then I was,653.94,5.1 | |
like okay it's a it was already implied,656.04,5.16 | |
that it you know just at the very outset,659.04,5.46 | |
right here where it said hey I don't,661.2,4.8 | |
recognize this can you tell me more,664.5,4.74 | |
about it so it can recognize novelty,666.0,4.98 | |
which is really important and really,669.24,4.68 | |
fascinating and so then I said can you,670.98,5.4 | |
discern what is novel and what isn't is,673.92,4.56 | |
it able to explain what it recognizes,676.38,5.16 | |
and what's new so it said in our,678.48,5.039 | |
conversation the following concepts and,681.54,3.72 | |
ideas would be considered a priori,683.519,4.141 | |
familiar so it understood theory of mind,685.26,4.92 | |
it already knew that language models and,687.66,3.78 | |
their potential for reasoning problem,690.18,2.76 | |
solving and cognition it said that it,691.44,2.82 | |
already understood that which is,692.94,3.54 | |
interesting General know Notions of,694.26,3.66 | |
memory and representation and efficiency,696.48,3.78 | |
the following concepts and ideas would,697.92,5.52 | |
be considered novel to me sparse primary,700.26,6.18 | |
representation or spot this spr's sparse,703.44,5.22 | |
priming representations and implied,706.44,4.139 | |
cognition so again it was able to,708.66,3.66 | |
restate these are things that I'm not,710.579,3.841 | |
familiar with,712.32,2.88 | |
um,714.42,2.64 | |
so then I told it that my hypothesis,715.2,4.319 | |
that perhaps what it's doing is that it,717.06,4.56 | |
actually gets a unique signal or a,719.519,4.741 | |
unique uh basically flow of tensors,721.62,5.76 | |
through the uh or mathematical patterns,724.26,4.86 | |
of tensors when there's novel,727.38,3.899 | |
information and it said oh yeah that's,729.12,4.26 | |
that's interesting and it comes up with,731.279,6.981 | |
a few ideas to um to explore how llms,733.38,7.86 | |
handle novel information and then we get,738.26,5.079 | |
into writing the repo which is out here,741.24,3.839 | |
so anyways and then I've got the whole,743.339,3.601 | |
conversation here so you can read it in,745.079,3.961 | |
Greater depth if you'd like,746.94,5.94 | |
um but yeah so that's it for today these,749.04,6.72 | |
are all ideas that I'm working on,752.88,6.12 | |
um and when I when I get so just if,755.76,4.98 | |
you're if if you're satisfied that's,759.0,3.42 | |
fine now I'm just going to ramble about,760.74,4.02 | |
kind of my own process so what happens,762.42,5.34 | |
is in the past when I get to this point,764.76,5.46 | |
I would start to write a new book but,767.76,5.579 | |
one writing a book is slow and two,770.22,4.32 | |
um well I mean that's the primary,773.339,3.721 | |
problem it it's slow but also we've got,774.54,4.32 | |
a good platform like there's my YouTube,777.06,4.079 | |
there's Reddit there's GitHub so it's,778.86,4.74 | |
like let me just go ahead and start,781.139,4.561 | |
um sharing this stuff as as soon as I've,783.6,3.0 | |
got it,785.7,3.42 | |
so that's that,786.6,4.979 | |
um yeah like I said uh all these are,789.12,3.959 | |
public they've all got the discussions I,791.579,2.82 | |
think I'll also go ahead and post these,793.079,3.301 | |
on Reddit for uh for discussion's sake,794.399,5.901 | |
anyways thanks for watching take care,796.38,3.92 | |