davidshapiro_youtube_transcripts
/
Accelerating Science with AI Quickly Read Every Paper and Get Key Insights in Bulk_transcript.csv
text,start,duration | |
hey everybody David Shapiro here with,0.659,3.781 | |
another video so it's a really exciting,2.76,5.28 | |
time because research is accelerating,4.44,5.699 | |
um just in the last couple days a whole,8.04,3.66 | |
bunch of interesting papers have come,10.139,5.521 | |
out on archive and um you know I I could,11.7,6.24 | |
read them all myself the the hard way,15.66,4.44 | |
the long way but you know what I'm,17.94,5.759 | |
really lazy and you know how uh Mr Henry,20.1,5.519 | |
Ford he said uh allegedly said I don't,23.699,2.881 | |
know if he actually said this but he,25.619,2.941 | |
said uh give your hardest problems to,26.58,3.42 | |
the laziest Engineers because they'll,28.56,3.24 | |
find an easier way to do it that's me,30.0,3.239 | |
that's why I became an automation,31.8,3.66 | |
engineer uh back in the day I told,33.239,4.081 | |
people like I would rather spend a week,35.46,4.56 | |
building an automation tool that I can,37.32,4.259 | |
then just push the button and for the,40.02,3.359 | |
rest of my life it makes that job easier,41.579,4.381 | |
and guess what that's how I automated a,43.379,5.401 | |
three petabyte environment because why I,45.96,4.86 | |
didn't want to manage it myself okay so,48.78,5.04 | |
uh with the volume of paper that came,50.82,5.16 | |
out because all of these came out Monday,53.82,5.64 | |
so we've got five interesting papers,55.98,5.46 | |
that came out on the same day and this,59.46,4.259 | |
happens pretty much every day lately and,61.44,4.14 | |
it is entirely too much to keep up with,63.719,5.341 | |
so rather than having to read all of,65.58,5.76 | |
these papers every day I wanted a way to,69.06,4.26 | |
digest them so I could quickly just one,71.34,4.98 | |
skim and keep on top of things uh and,73.32,6.06 | |
then also just uh basically make it,76.32,4.92 | |
sustainable or make it operationalized,79.38,3.779 | |
and so what I mean by that is something,81.24,3.48 | |
that you can just fully automate and,83.159,3.981 | |
then just maybe get an email every day,84.72,5.219 | |
or you know have the interns do it or,87.14,4.299 | |
use this as part of your literature,89.939,3.0 | |
review for science because one of the,91.439,2.881 | |
things that I want to do,92.939,4.021 | |
is use this technology to accelerate,94.32,6.06 | |
science as much as possible and uh you,96.96,4.74 | |
know one of the things that I hear from,100.38,3.059 | |
some of my friends in science is that,101.7,3.48 | |
just keeping up with things is getting,103.439,4.441 | |
harder and harder uh at least in these,105.18,5.16 | |
high velocity spaces some places like,107.88,4.08 | |
oceanography they don't change that much,110.34,3.9 | |
they don't change that fast but,111.96,4.26 | |
certainly computer science artificial,114.24,4.32 | |
intelligence uh and that sort of thing,116.22,4.859 | |
cyber security Quantum Computing these,118.56,5.16 | |
things are moving so incredibly fast we,121.079,6.0 | |
need the AI to help us keep up so let me,123.72,6.3 | |
show you what I've done uh let's see,127.079,4.38 | |
where should I start I guess I'll just,130.02,2.88 | |
start with uh some some of the final,131.459,3.481 | |
reports so basically what I do is I just,132.9,4.32 | |
copy paste the text of the thing into,134.94,4.92 | |
here obviously the figures help but,137.22,3.9 | |
language models are already pretty,139.86,2.519 | |
intelligent so they don't even really,141.12,4.44 | |
need the figures and then what I do is I,142.379,5.521 | |
feed it a series of,145.56,5.16 | |
um of questions and it gives me answers,147.9,5.94 | |
and so this basically summarizes it from,150.72,6.42 | |
you know know maybe twenty thousand,153.84,4.979 | |
forty thousand hundred thousand,157.14,3.78 | |
characters down to four thousand,158.819,4.321 | |
characters so we get a compression ratio,160.92,4.92 | |
of at least five to one but also it's,163.14,4.739 | |
often in much simpler language and,165.84,3.42 | |
probably what I'll do is I'll add a,167.879,3.301 | |
second step that will render this as an,169.26,4.14 | |
HTML document maybe I'll do that at the,171.18,3.66 | |
end,173.4,4.8 | |
um but anyways so the idea is I plug in,174.84,5.399 | |
the text if it's if it's a little too,178.2,3.72 | |
long it'll trim it but you know you get,180.239,4.08 | |
the first 22 000 characters of a paper,181.92,5.039 | |
you've got the general gist,184.319,4.14 | |
um and in many cases you actually have,186.959,4.261 | |
the whole paper uh now so I say can you,188.459,4.321 | |
give me a very clear explanation of the,191.22,2.939 | |
core assertions implications and,192.78,3.959 | |
mechanics elucidated in this paper very,194.159,4.741 | |
simple prompt it comes back and says you,196.739,3.661 | |
know yada yada yada the paper explores,198.9,3.54 | |
the concept of continual learning okay,200.4,4.32 | |
cool uh you know so it asks that,202.44,4.86 | |
question and so basically this is a this,204.72,5.34 | |
is a standard chat GPT conversation so,207.3,5.1 | |
this is a way to implement a chain of,210.06,3.36 | |
thinking,212.4,3.419 | |
um but it's really an automated dialogue,213.42,4.62 | |
so imagine that you you have an intern,215.819,4.381 | |
or a research assistant that read the,218.04,3.96 | |
paper and you can ask them a couple,220.2,4.2 | |
questions and you can change this uh on,222.0,5.04 | |
your own by the way very easily so in,224.4,4.68 | |
the code I have it here you can just add,227.04,3.479 | |
whatever questions or change whatever,229.08,3.54 | |
questions you want here and it will be,230.519,4.14 | |
happy to take it and run with with it,232.62,5.16 | |
for you entirely on its own so you,234.659,4.681 | |
change these questions if you want you,237.78,4.26 | |
add more you know depending on the,239.34,4.38 | |
research that you're looking for so,242.04,4.979 | |
let's say for instance you are in a very,243.72,5.46 | |
specific domain maybe you're trying to,247.019,4.5 | |
do this for medical research and you're,249.18,3.9 | |
really just looking for what are the,251.519,5.46 | |
implications for let's say Rejuvenation,253.08,6.119 | |
therapy or regenerative medicine or even,256.979,4.5 | |
a something even more specific maybe,259.199,3.78 | |
you're trying to research a specific,261.479,5.101 | |
kind of cancer or a specific uh any kind,262.979,5.461 | |
of metabolic disorder really,266.58,4.619 | |
so then you can say first you want it to,268.44,3.96 | |
just tell me what are the implications,271.199,3.241 | |
and then you could say okay what are the,272.4,4.26 | |
implications with respect to my,274.44,5.1 | |
particular domain what are the what what,276.66,4.86 | |
is you know what is the what is the,279.54,4.62 | |
impact to my area of research so restate,281.52,4.44 | |
it for me and then you can have it,284.16,4.2 | |
automatically spider a whole bunch of,285.96,3.72 | |
documents so one of the things that I'm,288.36,3.059 | |
going to do with this is that rather,289.68,4.86 | |
than have just one input at a time I'm,291.419,4.5 | |
going to create a folder where you give,294.54,3.06 | |
it you just download the PDFs to,295.919,4.021 | |
directly it'll scrape the PDFs and go,297.6,4.379 | |
through all of them as part of the,299.94,4.68 | |
automated process and it'll whatever the,301.979,4.621 | |
questions that you have it will kind of,304.62,3.6 | |
break it down but anyways,306.6,3.48 | |
so I have a I have a few standard,308.22,3.84 | |
questions can you explain the value of,310.08,3.3 | |
this in basic terms like you're talking,312.06,3.54 | |
to a CEO so what what's the bottom line,313.38,4.68 | |
here and it's really good at explaining,315.6,4.379 | |
things in simple terms so in this case,318.06,4.02 | |
uh this paper says the research is about,319.979,4.201 | |
making code generation models which are,322.08,3.54 | |
tools that help programmers by,324.18,3.66 | |
automating some of their work more,325.62,4.079 | |
adaptable and efficient currently these,327.84,3.66 | |
models need to be retrained every time,329.699,3.121 | |
there's a significant change to the,331.5,3.419 | |
programming languages or libraries they,332.82,3.48 | |
work with which this is what some people,334.919,3.661 | |
are noticing and I notice this too is,336.3,4.26 | |
that sometimes it'll program with an,338.58,4.38 | |
outdated version which of course limits,340.56,4.8 | |
the utility of coding models,342.96,4.019 | |
the bottom line is that this research,345.36,2.76 | |
could lead to more efficient and,346.979,3.0 | |
adaptable code generation models this,348.12,3.24 | |
means less time and money spent on,349.979,2.641 | |
retraining these models and more,351.36,2.52 | |
up-to-date and effective tools for,352.62,3.419 | |
programmers so the implication for this,353.88,3.9 | |
just reading that one paragraph to me,356.039,3.72 | |
says okay cool this is going to be part,357.78,4.919 | |
of the acceleration because if we come,359.759,5.581 | |
up with a way to make coding models uh,362.699,4.741 | |
updated they're updatable you know using,365.34,4.799 | |
continuous learning great excellent and,367.44,4.199 | |
then I said can you give me an analogy,370.139,2.521 | |
or metaphor that will help me explain,371.639,2.761 | |
this to a broad audience why because I'm,372.66,4.92 | |
a YouTuber right so the basically what I,374.4,5.46 | |
want is to be able to explain this paper,377.58,5.88 | |
in real basic terms real fast and,379.86,5.7 | |
contrary to the belief about a year ago,383.46,3.78 | |
language models are actually really good,385.56,3.9 | |
at analogy and metaphor,387.24,3.36 | |
um so it says sure let's think of the,389.46,2.64 | |
code generation models as chefs in a,390.6,3.3 | |
kitchen kitchen these chefs are trained,392.1,3.24 | |
to prepare a variety of dishes using,393.9,3.0 | |
specific recipes and ingredients however,395.34,3.9 | |
the culinary world is always evolving,396.9,4.32 | |
with new recipes and ingredients being,399.24,4.019 | |
introduced regularly in the current,401.22,4.02 | |
scenario every time a new recipe or,403.259,4.021 | |
ingredient is introduced the chef has to,405.24,3.54 | |
go through a complete retraining process,407.28,3.96 | |
which is time consuming and expensive,408.78,4.259 | |
it's as if the chef forgets all of the,411.24,3.299 | |
old recipes when they learn a new one,413.039,4.141 | |
see that's a good metaphor,414.539,4.801 | |
or I guess that's an analogy that's not,417.18,5.28 | |
a metaphor the researchers in this study,419.34,4.32 | |
developed a new Training Method for,422.46,2.459 | |
these chefs instead of forgetting old,423.66,2.819 | |
recipes when learning new ones the chefs,424.919,3.141 | |
can now continually learn and adapt,426.479,3.961 | |
adding new recipes to the repertoire,428.06,4.06 | |
while still remembering the old ones,430.44,3.0 | |
this is like a chef who can learn to,432.12,2.579 | |
cook a new dish while still remembering,433.44,2.819 | |
to cook all the previous dishes they've,434.699,3.361 | |
ever learned this new Training Method is,436.259,3.301 | |
called prompt pooling with teacher for,438.06,4.199 | |
Teacher forcing makes chefs or in the,439.56,4.139 | |
real case the code generation models,442.259,3.06 | |
more efficient and adaptable yada yada,443.699,4.021 | |
okay you get the idea so this only takes,445.319,4.021 | |
a few seconds and it takes only a few,447.72,2.879 | |
cents to run,449.34,3.479 | |
and so actually probably what I'll do is,450.599,3.72 | |
I'll pause it and I'll I'll go ahead and,452.819,3.181 | |
update it so that it'll just ingest,454.319,3.301 | |
everything in a folder and then I'll,456.0,3.599 | |
show you a nice cleaned up output so,457.62,4.94 | |
actually we'll be right back,459.599,2.961 | |
um okay so speaking of being lazy I just,463.02,3.06 | |
wanted to show you real quick I'm using,464.759,2.701 | |
my coding chat bot to do this because,466.08,3.6 | |
I'm ultra lazy,467.46,5.4 | |
work smart not hard so basically I took,469.68,4.799 | |
the script and I just copied it into my,472.86,3.839 | |
coding chat bot uh which you can find it,474.479,3.421 | |
it's literally just called coding,476.699,2.821 | |
underscore chatbot underscore assistant,477.9,4.26 | |
on my GitHub so I copied the script that,479.52,5.22 | |
I had which is all of 99 lines and I,482.16,4.08 | |
came over to my coding chatbot and I,484.74,3.179 | |
said hey can you do this for me instead,486.24,3.66 | |
and it's like sure,487.919,4.981 | |
um Okay cool so uh basically I said,489.9,6.0 | |
let's get a folder so we're using this,492.9,5.579 | |
folder so we can just download the PDFs,495.9,4.44 | |
directly to there rather than copy and,498.479,3.801 | |
pasting the text and we're just gonna,500.34,4.919 | |
take the output folder or the input,502.28,4.66 | |
folder for the PDFs and then an output,505.259,5.16 | |
folder for the um Whatchamacallit whoops,506.94,7.5 | |
okay so we come over to here and it,510.419,8.101 | |
wants us to add OS and Pi PDF,514.44,4.86 | |
um,518.52,4.439 | |
that's fine so we add that and then we,519.3,5.76 | |
see what else it wants to do if name,522.959,4.921 | |
equals main yep so it's going to do all,525.06,4.02 | |
that,527.88,3.66 | |
PDF file replace yep cool so I just I,529.08,3.9 | |
gave it really simple instructions I,531.54,3.299 | |
said can you modify this script so that,532.98,3.78 | |
it will one open every PDF in the input,534.839,4.201 | |
folder one by one and perform the same,536.76,5.28 | |
operation and two check an output folder,539.04,6.919 | |
for the report as the same PDF uh or,542.04,7.14 | |
blah check the output folder for a file,545.959,4.661 | |
of the same name this will enable us to,549.18,2.82 | |
run batches without duplicating effort,550.62,2.52 | |
should be pretty straightforward let me,552.0,2.94 | |
know if you have any questions,553.14,4.98 | |
um and okay so let's put this over here,554.94,6.66 | |
and see if it worked,558.12,4.68 | |
okay,561.6,4.679 | |
so um PDF files for files in OS Lister,562.8,5.64 | |
that looks right,566.279,3.24 | |
um,568.44,4.92 | |
open read binary as file PDF reader yep,569.519,6.361 | |
so we get that it gets all the page,573.36,4.26 | |
numbers that should all work I've used,575.88,3.24 | |
that I've used this this module before,577.62,5.58 | |
pi pdf2 and PDF file reader if it's over,579.12,7.02 | |
22 000 characters just cut it off it,583.2,5.579 | |
seems to work well enough and then we,586.14,5.639 | |
get the report so on and so forth,588.779,5.161 | |
um let's see,591.779,5.601 | |
it looks like it's not actually checking,593.94,7.519 | |
if the file exists,597.38,7.12 | |
yep okay so this is this is close so,601.459,4.541 | |
let's take this over to the scratch pad,604.5,3.54 | |
and then we'll say,606.0,4.1 | |
um,608.04,2.06 | |
let's see it will then check the output,610.92,3.66 | |
folder for the report wait am I missing,612.72,4.04 | |
something,614.58,2.18 | |
it doesn't have a continue function,616.98,7.979 | |
um okay sorry I I think we need to check,620.04,9.54 | |
um if the final file name exists in the,624.959,9.481 | |
output folder first and then use a,629.58,10.58 | |
continue to skip the process make sense,634.44,8.64 | |
I found that if you obviously like if,640.16,4.359 | |
you get frustrated with like no you eat,643.08,3.54 | |
it like it'll be apologetic and it'll,644.519,4.141 | |
spend more time apologizing,646.62,3.42 | |
um so don't be mean to the machine if,648.66,2.64 | |
you just say sorry I think there was a,650.04,2.58 | |
misunderstanding it's actually much more,651.3,3.42 | |
helpful so words matter,652.62,4.98 | |
um be polite to your coding assistant,654.72,4.739 | |
um because if it's if just like a human,657.6,3.9 | |
if the human is spending mental energy,659.459,4.141 | |
trying to placate you know a frustrated,661.5,4.32 | |
friend or whatever it's going to waste a,663.6,6.299 | |
little bit more time and energy okay,665.82,4.079 | |
um sure here's the modified script the,670.019,2.701 | |
script will first check if the report,671.579,2.581 | |
already exists in the output folder if,672.72,3.0 | |
it does it will skip the current PDF and,674.16,3.66 | |
move to the next one excellent so let's,675.72,4.5 | |
check this out and see,677.82,5.94 | |
if it works if it if it looks how I,680.22,6.679 | |
think it should look,683.76,3.139 | |
so we'll come over here all right so for,686.94,5.76 | |
PDF file and PDF files first all right,690.0,4.74 | |
yep so first we list there we go that's,692.7,4.98 | |
what I was expecting if it exists then,694.74,4.38 | |
continue so you just skip the whole,697.68,4.26 | |
thing otherwise go ahead and open it yep,699.12,4.38 | |
there we go so we said because we set,701.94,3.3 | |
the file name first all right so this,703.5,3.48 | |
should be fine,705.24,4.26 | |
um okay I think that's I think that's uh,706.98,5.039 | |
good so let's come over here let's exit,709.5,4.86 | |
out of my coding chatbot assistant,712.019,3.781 | |
um so again like I had to put very,714.36,2.76 | |
little energy into that all I had to do,715.8,2.88 | |
was check and honestly what I could have,717.12,3.959 | |
done is just ask it to check its work,718.68,4.08 | |
like you know does it do the thing that,721.079,3.841 | |
I think it is but you saw I didn't,722.76,4.319 | |
really do any coding all I did was copy,724.92,5.58 | |
paste okay uh so let's go back out of,727.079,6.121 | |
here quickly extract science papers and,730.5,6.0 | |
then we'll do python whoops generate,733.2,8.639 | |
multiple reports no module named pi pdf2,736.5,6.54 | |
I thought I already had it maybe that,741.839,3.261 | |
was my previous install pip install,743.04,5.16 | |
ipdf two,745.1,7.08 | |
Pi PD F2,748.2,3.98 | |
this is the boring part okay now it,752.76,4.38 | |
should work,754.92,4.8 | |
um is deprecator and was removed used,757.14,4.98 | |
PDF reader instead so this is exactly,759.72,6.0 | |
exactly the problem that uh that the um,762.12,5.82 | |
uh this one of the papers addresses is,765.72,3.9 | |
the fact that the modules have been,767.94,4.86 | |
updated and so now it's it's uh out of,769.62,5.64 | |
date so we gotta go we gotta go switch,772.8,4.74 | |
to this thing over here,775.26,5.639 | |
um all right oh I forgot to save it,777.54,5.46 | |
that could be a problem,780.899,4.56 | |
um all right so,783.0,5.76 | |
if it ends with otherwise,785.459,5.641 | |
all right let's see if that fixed it and,788.76,4.019 | |
also it might have yard anyways because,791.1,4.26 | |
I forgot to save uh reader num pages is,792.779,5.581 | |
deprecated use Len reader pages instead,795.36,4.74 | |
okay at least at least their,798.36,3.3 | |
documentation is good and it's helping,800.1,3.06 | |
me fix it,801.66,3.859 | |
um okay,803.16,2.359 | |
um,807.72,2.119 | |
okay that should be good,812.16,5.76 | |
int object is not iterable,814.86,3.93 | |
uh,817.92,2.039 | |