text,start,duration
hey everybody David Shapiro here with a,0.719,4.441
video update,3.54,4.02
um so this is super exciting let me just,5.16,4.14
go ahead and show you right off the bat,7.56,3.42
what I'm working on and I apologize that,9.3,3.06
it's small,10.98,4.44
um this is uh it's a it's a very,12.36,5.999
primitive graphical user interface but,15.42,5.46
as most of you are aware chat GPT has,18.359,6.42
been in um let's shall we say bad shape,20.88,6.6
um network errors your history is gone,24.779,5.401
etc etc and for a lot of people that,27.48,4.86
means like your your work comes to a,30.18,4.74
grinding halt that being said I've been,32.34,5.52
working on learning the chat gbt API,34.92,5.459
I've got access to gpt4,37.86,5.34
um not the 32 000 token one um if anyone,40.379,4.321
knows how to request access to that,43.2,2.699
please let me know in the comments,44.7,4.14
because uh working on a novel I could,45.899,5.301
easily make use of the 32 000 tokens,48.84,5.16
that being said the 8 000 tokens will,51.2,6.879
carry you a really long ways so with,54.0,5.28
that being said,58.079,3.241
um this is the repo it's chat some,59.28,4.32
chapter summarizer gpt4 and it is,61.32,4.199
already completely wrong because it,63.6,3.72
provides developmental feedback Pros,65.519,3.901
feedback and now you can actually chat,67.32,3.299
with it,69.42,4.739
um and it will know your entire story so,70.619,5.701
how did I achieve this,74.159,5.82
um with the larger token window with the,76.32,6.96
8000 token window that is enough that,79.979,4.5
you can,83.28,3.18
um some you can summarize all your,84.479,4.981
chapters and it will fit you generally,86.46,4.799
within it it probably won't fit by the,89.46,3.659
time I finish my novel but hopefully,91.259,5.22
I'll have access to the 32 000 token API,93.119,4.621
endpoint which means that I'll be able,96.479,4.561
to summarize the whole thing and then,97.74,5.4
what I did was I took the scratch Pad,101.04,3.66
here and I've got all this documented,103.14,3.0
too because this is like the best,104.7,4.5
instance of Auto Muse yet so basically,106.14,4.979
you copy and paste whatever you want,109.2,3.84
into the scratch Pad,111.119,5.221
um uh file uh whoops where did it go so,113.04,4.92
you copy it into scratch pad and you see,116.34,4.38
it's 24 kilobytes so this is a summary,117.96,6.9
of the first 19 chapters of my story,120.72,5.999
um and it's from here you know the,124.86,4.44
chapters the text the conversion the,126.719,4.26
chat logs oh I want to show you like,129.3,3.96
this is working right like here's all,130.979,4.201
the chat logs that I've got uh between,133.26,4.8
myself and and and the Muse which also,135.18,4.199
means that you've got a permanent record,138.06,3.78
even if chat gbt goes down,139.379,4.561
um and all this is excluded in the in,141.84,4.92
the git ignore so you see um open AI,143.94,6.0
your key is ignored chapters uh the,146.76,4.8
summaries and the chat logs and the,149.94,3.36
scratch Pad so don't you don't have to,151.56,3.42
worry about accidentally sharing any of,153.3,2.82
your own,154.98,3.36
um any of your own work,156.12,3.9
um it'll just stay local in the repo,158.34,4.74
when you copy it down now I was using,160.02,7.38
chat gpt4 to help write a graphical user,163.08,5.7
interface because I was like you know,167.4,2.64
what,168.78,3.84
um I could do this as a web web thing,170.04,5.279
but honestly I don't like why have,172.62,5.699
multiple apps running like why why use a,175.319,6.241
a web browser when python has a built-in,178.319,5.401
graphical user interface and so that's,181.56,3.3
this here,183.72,3.9
so I was using chat GPT,184.86,4.44
um to help make this and you can see,187.62,4.86
like it dynamically sizes I've got it,189.3,5.76
color coded so that you can just easily,192.48,5.22
see what's going on,195.06,4.92
um it also uses threading so that way,197.7,4.44
because it takes a while for the API to,199.98,5.16
uh to spit out something but yeah so let,202.14,6.12
me show you the uh the The Prompt that I,205.14,5.64
use so the scratch Pad this is just all,208.26,4.74
the information about your story I,210.78,4.5
probably will get this automatically,213.0,4.019
updated in the future but for right now,215.28,4.86
it's just manually copy pasted and then,217.019,5.22
I have a new kind of prompt it's a,220.14,3.599
instead of a prompt it's a system,222.239,2.941
message,223.739,4.08
so in this case it says here let me zoom,225.18,5.339
in I'm a writing assistant and coach the,227.819,4.381
following is the summary of chapters for,230.519,3.36
the novel the user is working on my job,232.2,3.42
is to help them develop their novel in,233.879,3.301
any way that I can including,235.62,3.119
brainstorming plotting outlining and,237.18,3.6
planning we may engage in developmental,238.739,3.72
editing which includes restructuring or,240.78,3.42
modifying character arcs plot threads,242.459,2.7
World building and development,244.2,3.66
developing of themes we almost we may,245.159,4.5
also work on finer grained aspects such,247.86,3.659
as character backstory subplots and so,249.659,4.381
on I should refer to highly professional,251.519,4.68
writing Concepts such as the 3x plot,254.04,4.199
structure the work of Joseph Campbell,256.199,4.801
and CG young save the cat and any other,258.239,5.041
formalized approaches to storytelling I,261.0,4.08
should not seek to push the author but,263.28,3.6
rather teach them when I perceive gaps,265.08,4.14
in their understanding Above All Else my,266.88,3.72
primary mission is to help them produce,269.22,4.14
an amazing work of literature,270.6,4.92
um and then so then I you can populate,273.36,4.38
whatever's in the scratch Pad here and,275.52,4.02
again you can dynamically update this if,277.74,3.06
you want,279.54,2.939
um important note the above material is,280.8,3.42
provided for background information this,282.479,3.72
format is not necessarily intended to be,284.22,4.14
used moving forward lastly be sure to,286.199,3.421
ask plenty of questions so that the,288.36,2.76
author remains firmly in the driver's,289.62,4.92
seat so this huge Constitution or,291.12,6.18
purpose or system message actually works,294.54,3.96
really well,297.3,3.06
so in this case,298.5,4.08
um just right off the bat I can say hey,300.36,3.899
describe the main characters and,302.58,2.94
identifies all four of the main,304.259,4.021
characters and gives a very succinct,305.52,5.7
definition of them and actually this,308.28,5.82
would be really good back matter,311.22,6.78
um for the the you know the little blurb,314.1,6.12
um so you can see that just with one,318.0,4.62
message it is able to,320.22,5.1
um uh read my entire story up to this,322.62,5.1
point while the summaries at least and,325.32,4.5
we can talk about it obviously I don't,327.72,3.78
want to reveal too much more about my,329.82,2.76
story,331.5,3.0
um you guys know the main characters and,332.58,3.839
a little bit about them,334.5,5.46
um but yeah so uh give it a try yourself,336.419,6.181
I'm super happy with this,339.96,4.32
um and the biggest thing that it does is,342.6,4.74
it saves time because again I've got,344.28,5.04
I've got all of my chapters loaded into,347.34,3.96
it and now I can just ask it whatever I,349.32,2.939
need,351.3,2.66
um also I found that this system message,352.259,4.141
it really cuts to the point I'm just,353.96,4.12
like hey let's plan the scenes for the,356.4,3.6
next chapter it's like okay cool let's,358.08,3.059
go,360.0,3.18
um it does it does do a little bit too,361.139,4.321
much for me but the fact that you don't,363.18,3.959
have to spend too much time going back,365.46,4.079
and forth I think is actually better,367.139,4.921
because it will it'll come up with an,369.539,4.261
idea and say hey do you like this let's,372.06,3.96
modify it and then you can say yes let's,373.8,3.72
modify it,376.02,3.66
um so that's probably all you need to,377.52,4.679
know but now we can step through the uh,379.68,4.5
the the function,382.199,4.44
um so I I did a nice clear demarcation,384.18,4.56
this is what I wrote,386.639,4.221
so just these just these three things,388.74,5.1
open file save file the chat GPT,390.86,5.08
completion which you can see it's using,393.84,4.26
gpt4,395.94,4.379
um I did I do still have it nested in a,398.1,5.039
while true Loop which is basically just,400.319,5.181
to capture in case um,403.139,5.461
in case the API has an error which it,405.5,6.039
does all the time lately and then I,408.6,6.36
added this here because if the chat,411.539,5.94
conversation gets too long,414.96,5.16
um it will it will give you a maximum,417.479,4.621
content length there and so in that case,420.12,4.079
I just removed the oldest message,422.1,4.2
not a big deal,424.199,2.761
um,426.3,4.14
it actually uh chat gpt4 said why don't,426.96,5.579
you actually measure the tokens and just,430.44,4.68
uh remove it until it's done but I,432.539,4.021
couldn't actually find anything on this,435.12,4.74
the the open AI API client the clean the,436.56,5.1
underscore clean,439.86,2.76
um,441.66,2.7
I think that's a tokenizer but I,442.62,3.78
couldn't find any documentation and I,444.36,4.32
didn't want to try it just yet but,446.4,4.98
anyway so that's there uh let's see,448.68,4.98
nothing else really changed here,451.38,3.84
um I still have the exponential back off,453.66,3.42
which basically means that it'll wait,455.22,5.28
longer and longer to retry until it will,457.08,5.28
finally give up,460.5,4.819
um and it's just because the the the the,462.36,6.0
API is wonky right now now everything,465.319,7.181
from here down was written by gpt4,468.36,6.48
uh just with some feedback back and,472.5,5.759
forth between myself and gpt4 so,474.84,5.639
basically it came up with the function,478.259,4.321
to send the message the thread to get,480.479,6.06
the response I asked it to uh to ignore,482.58,6.119
like if you if you use shift enter it,486.539,4.38
will just insert the um the new line,488.699,3.601
character otherwise it'll send the,490.919,2.641
message,492.3,3.299
um because I realize like that function,493.56,5.34
is available in chat GPT on the web and,495.599,5.88
then finally uh it did every uh,498.9,5.04
basically all it did was it took I gave,501.479,4.56
it this much and you can see here where,503.94,4.74
we populate the system message by,506.039,4.38
loading what you have in the scratch pad,508.68,3.539
it then populates the system message,510.419,3.841
based on this,512.219,4.861
and starts the conversation with that,514.26,5.88
um so the the the system role will be in,517.08,6.18
index zero meaning index one is going to,520.14,4.259
be the oldest message in the,523.26,2.22
conversation,524.399,3.241
so then it fires up the ticket the,525.48,4.56
tokenter GUI oh and for anyone who's not,527.64,4.379
familiar with tokenter um that's what,530.04,3.96
this is it's just it's a really brain,532.019,4.5
dead simple uh portable,534.0,5.76
um uh graphical user interface and so,536.519,5.341
then it constructs it all there's a few,539.76,4.259
problems when I first created it so one,541.86,5.82
the the window was was static so it,544.019,6.241
wouldn't allow you to expand and then,547.68,4.62
also while it was waiting for the API to,550.26,3.48
respond the whole thing would freeze and,552.3,3.719
say not responding but now it has a,553.74,4.08
little output here that says you know,556.019,3.961
Auto Muse is thinking,557.82,5.34
um so welcome to Auto Muse,559.98,5.58
um it gave the uh you know it set to,563.16,3.9
word wrap,565.56,4.02
um it's yeah so basically it just when I,567.06,4.62
had when I had an issue I just said okay,569.58,4.439
hey can you make this modification,571.68,5.04
um and away it goes and then again one,574.019,5.521
thing that that I haven't integrated is,576.72,4.14
um being able to search the previous,579.54,4.5
logs but because of the work that I've,580.86,7.68
been doing on um on other topics,584.04,6.479
um that that'll be there now,588.54,4.2
it might not be relevant because once we,590.519,4.5
get to the 32 000 token window limit,592.74,4.26
that's going to be like,595.019,5.161
what like a hundred Pages uh worth of,597.0,5.58
chat logs like you won't even need that,600.18,5.04
much uh history right so you could load,602.58,6.0
this entire history and you know most of,605.22,5.4
it's going to be irrelevant uh so,608.58,4.56
anyways we'll see I did have a thought,610.62,5.52
uh one final thought about the kinds of,613.14,5.34
problems that we that we solve just by,616.14,5.04
increasing context window,618.48,5.28
um is amazing and so basically what I'm,621.18,5.04
going to be doing from now on is I'm not,623.76,5.88
even gonna try and and force problems to,626.22,5.52
fit into smaller context Windows because,629.64,4.4
one larger context windows are coming,631.74,5.4
and two the kinds of problems you are,634.04,4.96
able to address once you have that,637.14,3.379
larger amount of working memory,639.0,4.38
completely changes the game,640.519,5.32
um so I was talking with my fiance and,643.38,5.579
some people on Discord and what we're,645.839,5.161
noticing is that we're getting more and,648.959,4.681
more purpose-built models so there are,651.0,4.2
some open source models out there that,653.64,4.08
have token windows up to 64 000 tokens,655.2,5.34
which is crazy some of them have 16 000,657.72,5.22
tokens so on and so forth so what we,660.54,4.08
expect to happen,662.94,4.56
is that uh we're going to end up with,664.62,5.219
many many many models most of them open,667.5,4.92
source that are purpose built for,669.839,5.641
specific tasks so for instance some,672.42,6.24
models are going to be focused on uh,675.48,4.14
they're going to be like window,678.66,3.419
optimized models right so like we might,679.62,4.62
eventually have a window optimized model,682.079,5.521
that can have a token uh window of like,684.24,6.36
a thousand uh not a thousand a million,687.6,5.16
tokens right but then we might also have,690.6,4.38
compute optimized models or cognition,692.76,3.84
optimized models,694.98,2.88
um that sort of stuff and some of the,696.6,2.58
people in Discord are like oh yeah I'm,697.86,2.88
already working on that so that's the,699.18,3.36
way that it's going,700.74,5.099
um and then aside from that uh from just,702.54,5.64
from a psychological perspective my,705.839,4.74
fiance pointed out that working memory,708.18,5.04
is actually one of the biggest factor in,710.579,6.181
intelligence and that by by doubling the,713.22,5.04
working memory the effective working,716.76,4.079
memory by going from gpt3 with a token,718.26,5.519
count of four thousand to gpt4 with a,720.839,6.0
token Window 8 000 doubling the working,723.779,5.521
memory that like basically takes it up,726.839,4.861
in order of magnitude or a standard,729.3,4.14
deviation of IQ,731.7,6.56
so if GPT 3.5 had an IQ of about 115,733.44,7.5
doubling that that window takes it up,738.26,6.04
another 15 points to about 130,740.94,5.88
um and then when you when you actually,744.3,4.5
if you were to test its IQ based on,746.82,5.519
speed it's probably already like,748.8,6.18
in the 160s or something,752.339,5.101
um and the so there's there's from from,754.98,4.859
a from a numbers perspective,757.44,4.86
two of the universal features of,759.839,4.201
intelligence that figure that should,762.3,4.2
figure into a good IQ test one is,764.04,5.28
processing so IQ,766.5,5.519
um most IQ tests are actually speed,769.32,5.819
tests which because speed is a good,772.019,5.94
proxy for a lot of kinds of intelligence,775.139,6.241
so your mental speed is not your IQ but,777.959,5.82
it is a proxy for actual intelligence,781.38,5.16
and so when you measure the speed of,783.779,6.12
gpt3 and gpt4 its IQ is actually really,786.54,7.32
high but then the other universal one of,789.899,5.521
the other Universal things is working,793.86,2.76
memory,795.42,4.68
so uh working memory is also a really,796.62,7.08
good proxy for uh for intelligence and,800.1,5.64
many forms of intelligence so when you,803.7,4.68
combine the large working memory,805.74,4.92
combined with the speed,808.38,4.8
um of reading and output I would say,810.66,5.78
gpt4 probably has an effect IQ of around,813.18,7.08
130 if I had to guess now that being,816.44,5.86
said it doesn't have permanent memory,820.26,3.84
you have to have a secondary system for,822.3,3.96
that but that paper came out that talked,824.1,4.56
about how storage systems combined with,826.26,5.519
llms are are computationally complete so,828.66,5.76
I'm not worried about that,831.779,4.5
um yeah so that's that's the direction,834.42,4.02
it's going and that's where we're at,836.279,5.281
right now so I'm really excited to get,838.44,5.94
my hands on the 32 000 token one I think,841.56,4.26
that that will also really help with the,844.38,3.18
scientific research one,845.82,4.38
so I haven't made any progress on oh and,847.56,3.959
I also wanted to show you how expensive,850.2,3.66
it was so yesterday it cost about four,851.519,4.681
to eight dollars to summarize the first,853.86,4.74
60 of my novel and then the,856.2,4.439
conversations that I had about it so,858.6,3.66
even just reading all the summaries and,860.639,3.541
talking about it cost another two,862.26,4.44
dollars and 62 cents which is still kind,864.18,4.56
of expensive but in the grand scheme of,866.7,5.34
things like it's so fast and it is still,868.74,4.62
a lot more expensive than just the,872.04,3.0
twenty dollars a month,873.36,4.2
um for the like unlimited chat GPT but,875.04,3.84
that price will come down I'm not,877.56,3.54
worried about that but anyways so,878.88,4.8
figuring out these chat interfaces and,881.1,4.62
how to handle large amounts of,883.68,3.779
information and and really how to make,885.72,4.44
the most use of the system window that,887.459,4.981
is what I'm working on now and I will,890.16,5.039
return to the um,892.44,5.22
let's see where is it,895.199,4.38
um the yeah the regenerative medicine,897.66,3.9
thing so I've been thinking about this,899.579,3.241
one,901.56,3.899
um come on so I got a bunch of papers,902.82,4.259
but the thing is is these papers are,905.459,3.12
unrelated wow it's already been two,907.079,3.781
weeks good grief,908.579,4.981
um a lot of these papers are unrelated,910.86,4.44
um which I think,913.56,3.48
um is kind of problematic although you,915.3,3.3
might think the idea that like cost,917.04,3.84
pollinating stuff like correlating stuff,918.6,4.02
from unrelated papers could be really,920.88,5.1
good priming exercise so anyways as I'm,922.62,5.339
getting more familiar with the chat GPT,925.98,4.799
API and how to use it I will come back,927.959,5.341
to this because I realized like yeah I,930.779,3.841
really want to help accelerate science,933.3,4.02
as fast as possible because I'm really,934.62,4.26
looking forward to the day where I can,937.32,3.66
go to the doctor and just like get an,938.88,4.079
outpatient injection and then of my,940.98,3.479
joints will feel better in a few months,942.959,4.861
right that's really what I want,944.459,5.94
um because man like I do not have as,947.82,4.379
much energy as I used to and I have to,950.399,3.481
be really careful because like I have,952.199,3.421
old shoulder entries from falling while,953.88,3.42
snowboarding like there's so much that I,955.62,3.42
can't do anymore just because I don't,957.3,3.779
want to hurt myself anymore,959.04,3.72
um anyways I'm complaining about getting,961.079,3.781
old it sucks everybody knows that I,962.76,4.5
think this is enough of the video so,964.86,5.099
thanks for watching cheers stay tuned,967.26,7.22
um this is this is so much fun,969.959,4.521