text,start,duration
morning everybody David Shapiro here,0.0,7.379
with a surprise update so yesterday July,2.46,8.46
5th this paper dropped long net scaling,7.379,6.541
Transformers to 1 billion,10.92,4.44
tokens,13.92,4.619
now to put this in context,15.36,6.48
this is a three gigapixel image which,18.539,7.381
you can make sense of at a glance,21.84,6.9
and I'm not going to dig too deep into,25.92,5.88
the uh the cognitive neuroscience and,28.74,5.64
neurological mechanisms about why you,31.8,5.34
can make sense of so much information so,34.38,4.08
quickly but if you want to learn more,37.14,3.06
about that I recommend the forgetting,38.46,3.779
machine,40.2,5.1
but what I want to point out is that you,42.239,4.681
can take a glance at this image and then,45.3,4.259
you can zoom in and you understand the,46.92,5.28
implications of every little bit of this,49.559,5.221
image this is clearly an arid mountain,52.2,4.499
range there's a road going across it,54.78,4.2
there's uh some Haze there's a city in,56.699,4.561
the background you can keep track of all,58.98,4.32
of that information at once just by,61.26,4.5
glancing at this image and then when you,63.3,4.08
zoom in,65.76,3.899
you can say oh look there's a nice house,67.38,3.9
on the hillside and you can keep track,69.659,4.261
of that information in the context of,71.28,4.699
this three gigapixel image,73.92,6.239
this is fundamentally what sparse,75.979,7.78
attention is and that is how this paper,80.159,6.241
solves the problem of reading a billion,83.759,3.841
tokens,86.4,4.5
so let's unpack this a little bit first,87.6,6.12
I love this chart this is this is a,90.9,4.38
really hilarious Flex right at the,93.72,3.0
beginning of this paper right under the,95.28,4.68
abstract they're like okay you know 512,96.72,8.64
uh 12K 64k tokens uh 262 a million,99.96,7.08
tokens and then here's us a billion,105.36,4.5
tokens so good job oh also I want to,107.04,5.219
point out this is from Microsoft this is,109.86,4.56
not just from some Podunk you know,112.259,4.621
Backwater University this is Microsoft,114.42,4.8
you know who's in partnership with,116.88,5.94
openai uh and so I saw a post somewhere,119.22,4.56
I think it was a tweet or something,122.82,2.399
someone's like Microsoft seems like,123.78,3.54
they're really just falling down on AI,125.219,3.661
research and I have no idea what rock,127.32,3.48
they're living under but pay attention,128.88,4.92
to Microsoft my money is on Microsoft,130.8,5.88
and Nvidia uh for for the AI race and,133.8,4.56
then of course there's Google,136.68,3.779
um but I don't understand get Google's,138.36,3.42
business model because they invented,140.459,2.701
this stuff and then sat on it for seven,141.78,3.24
years so I have no idea what Google's,143.16,4.92
doing anyways sorry I digress,145.02,8.1
okay billion tokens seems kind of,148.08,8.28
out there kind of hyperbolic right the,153.12,7.259
the chief uh Innovation here is one they,156.36,5.519
they have a training algorithm which I,160.379,2.64
don't care about that as much I mean,161.879,2.881
distributed training okay lots of people,163.019,3.661
have been working on that but the chief,164.76,3.9
Innovation here is what they call,166.68,5.88
dilation so let me uh bring that up uh,168.66,6.299
so what they do is let's see hang on,172.56,3.72
where did it go where did it go where's,174.959,3.661
the dilation diagram all right so what,176.28,4.5
it allows it to do a dilated attention,178.62,3.24
sorry,180.78,3.48
so what dilated attention allows it to,181.86,4.26
do is to zoom out,184.26,4.559
and take in the entire sequence all at,186.12,6.32
once which controls the amount of memory,188.819,6.42
uh and computation that it takes to take,192.44,6.1
in that large sequence just like you and,195.239,5.821
your brain zooming in and out,198.54,5.339
and keeping the entire context of this,201.06,6.179
image in mind,203.879,6.061
at the same time,207.239,5.041
and so the way that it does that is,209.94,5.7
actually relatively similar,212.28,6.06
to the way that human brains do it hang,215.64,4.86
on hang on where did it go I'm missing,218.34,5.58
the diagram okay so what they do,220.5,5.64
is they create sparse representations,223.92,3.899
and those who have been following me for,226.14,3.42
a while you might remember when I came,227.819,3.121
up with the idea of sparse priming,229.56,3.959
representations this is something to pay,230.94,5.579
attention to because what I realized is,233.519,4.921
that language models only need a few,236.519,3.961
Clues just a few breadcrumbs to remind,238.44,5.219
it to cue in as to what is going on in,240.48,4.74
the message to what's going on in the,243.659,3.601
memories and this is actually why it's,245.22,3.84
really good at you just give it a tiny,247.26,4.199
chunk of text and it can infer the rest,249.06,5.7
why because it has read so much that it,251.459,5.46
is able to infer what came before that,254.76,4.62
text and what came after and so by,256.919,4.921
zooming out and creating these really,259.38,4.92
sparse representations of larger,261.84,4.98
sequences it can keep track of the,264.3,6.06
entire thing and what it does is it will,266.82,6.06
take the the up to a billion token,270.36,3.6
sequence,272.88,4.62
break it up slice it up and then makes a,273.96,5.94
layered sparse representations of the,277.5,4.259
entire thing and it will therefore be,279.9,4.26
able to keep track of it now okay that,281.759,5.701
sounds really nerdy uh but here's here's,284.16,6.06
what it does for the performance so with,287.46,4.5
this sparse representation with this,290.22,5.16
dilation and doing it in massively in,291.96,6.72
parallel it solves a few problems so one,295.38,6.599
you see that the runtime stayed under a,298.68,5.76
thousand uh milliseconds under one,301.979,5.101
second it's more about half a second all,304.44,5.1
the way up to a billion tokens,307.08,4.619
so because of that it's basically,309.54,4.02
zooming in and out of the text the,311.699,3.181
representation of the text that it,313.56,3.9
creates in the same way a very similar,314.88,4.62
way that your brain keeps track of a,317.46,3.72
three gigapixel image as you zoom in and,319.5,3.9
out you're like okay cool okay I see a,321.18,3.84
bunch of cars parked on the side of the,323.4,2.579
road,325.02,2.28
um and you can just remember that fact,325.979,4.081
oh let's do a quick test what else do,327.3,4.5
you remember about this image maybe you,330.06,3.54
remember that the the Hollywood sign is,331.8,3.0
in the background over here somewhere,333.6,2.28
there it is,334.8,2.88
oh no that's not it,335.88,4.14
but it's somewhere in here so it's like,337.68,4.56
okay based on the cars and the Arid,340.02,4.32
desert and that I'm based I'm guessing,342.24,4.86
that this is Los Angeles right uh,344.34,5.16
anyways point being is that oh there it,347.1,4.68
is Hollywood uh so this is these are the,349.5,4.199
Hollywood Hills and you can remember oh,351.78,3.84
yeah there was a nice mansion over here,353.699,3.84
there's cars parked over here that's,355.62,4.2
probably downtown LA the Hollywood sign,357.539,4.141
is over here so by keeping by basically,359.82,4.7
creating a mental map this treats,361.68,5.7
gigantic pieces of information not,364.52,5.2
unlike a diffusion model,367.38,4.86
when and because I I got I was clued in,369.72,4.319
on that when I looked at the way that it,372.24,3.959
was it was um mapping everything and I,374.039,4.5
was like hold on it's creating a map of,376.199,4.381
the text by just breaking it down,378.539,4.201
algorithmically and saying okay let's,380.58,4.2
just make a scatter plot of all the text,382.74,3.239
here,384.78,3.0
uh or scatter plot's not the right word,385.979,4.261
but the it's basically making a bitmap,387.78,5.639
of the text of the representations of,390.24,4.98
what is going on in the sequence and I'm,393.419,4.141
like okay this is a fundamentally,395.22,3.9
different approach to representing text,397.56,4.079
and this is also really similar to some,399.12,4.019
of the experiments that I've done if you,401.639,3.301
remember Remo rolling episodic memory,403.139,3.541
organizer which creates layers of,404.94,4.14
abstraction this does it algorithmically,406.68,5.16
in real time so this just blows,409.08,5.04
everything that I've done with memory,411.84,5.46
research completely out of the water it,414.12,7.079
also has the ability to basically uh,417.3,7.019
kind of summarize as it goes and that's,421.199,5.581
not necessarily the right word because,424.319,4.261
summarization means that you take one,426.78,3.78
piece of text and create a smaller piece,428.58,4.2
of text but this creates a neural,430.56,4.02
summarization a neural network,432.78,4.139
summarization by creating these layers,434.58,3.78
of abstraction,436.919,4.381
and this allows it to zoom in and out as,438.36,4.8
it needs to so that it can cast its,441.3,4.14
attention around internally in order to,443.16,4.2
keep track of such a long sequence now,445.44,4.8
okay great what does this mean,447.36,5.52
as someone who has been using GPT since,450.24,5.7
gpt2 where it was basically just a,452.88,5.939
sentence Transformer it couldn't do a,455.94,4.92
whole lot more you know like in in this,458.819,5.22
model up here the original GPT was 512,460.86,7.14
tokens and gpt2 I think was what a,464.039,5.28
thousand I don't remember,468.0,4.8
maybe it was five uh 512 as well and,469.319,6.421
then the initial version of gpt3 was 2,472.8,5.459
000 tokens we got upgraded to 4 000,475.74,6.239
tokens then we got GPT 3.5 and gpt4 so,478.259,5.041
we're at eight thousand and sixteen,481.979,3.0
thousand tokens,483.3,4.14
as these attention mechanisms get bigger,484.979,4.5
and as the context window gets bigger,487.44,3.9
one thing that I've noticed is that,489.479,4.261
there are one these are step changes in,491.34,3.78
terms of,493.74,4.2
algorithmic efficiencies but in terms of,495.12,6.66
what they are capable of doing as I tell,497.94,6.479
a lot of my uh my consultation clients,501.78,5.639
do not ever try and you know get around,504.419,4.921
the context window limitation because,507.419,4.56
one a new model is coming out within six,509.34,4.499
months that's going to completely blow,511.979,4.5
open that window and two it's just a,513.839,4.32
limitation of the model,516.479,3.901
so when you can read a billion tokens,518.159,5.641
which by the way humans read about one,520.38,5.16
to two billion words in their entire,523.8,3.479
lifetime,525.54,3.66
when you have a model that can read a,527.279,3.721
billion tokens in a second,529.2,4.319
that is almost that is half a lifetime,531.0,5.279
worth of reading and knowledge that this,533.519,5.581
model can take in in a second so when,536.279,4.56
you have a model that can ingest that,539.1,4.26
much information suddenly retraining,540.839,4.44
models doesn't matter you just give it,543.36,4.08
the log of all news all events all,545.279,4.261
papers whatever tasks that you're doing,547.44,4.68
you just give it all of it at once and,549.54,4.739
it can keep track of all of that text in,552.12,4.8
its head in its virtual head,554.279,4.381
um all at once and it can pay attention,556.92,4.02
to the bits that it needs to with those,558.66,4.619
sparse representations,560.94,5.579
it is it is impossible for me to,563.279,6.841
oversell the long-term ramifications of,566.519,6.841
these kinds of algorithmic changes and,570.12,6.3
so a couple months ago when I said AGI,573.36,5.159
within 18 months this is the kind of,576.42,3.62
trend that I was paying attention to,578.519,4.32
there is no limit to the algorithmic,580.04,4.299
breakthroughs we are seeing right now,582.839,3.12
now that doesn't mean that there won't,584.339,4.44
eventually be diminishing returns but at,585.959,5.281
the same time we are exploring this blue,588.779,5.581
ocean space and we've we've all right,591.24,5.159
for those of you that have played Skyrim,594.36,5.58
and other RPGs we unlocked a new map and,596.399,5.641
the grayed out area is this big and,599.94,4.079
we've explored this much of this new map,602.04,4.799
that is how much potential there is to,604.019,5.341
explore out here and the other thing is,606.839,5.281
this research is accelerating there's a,609.36,4.38
few reasons for that on one of the live,612.12,3.18
streams like someone asked me like how,613.74,4.08
do we know that this isn't an AI winter,615.3,4.74
and I pulled up a chart that showed an,617.82,5.16
exponential growth of investment where,620.04,5.16
the money goes the research goes and,622.98,3.78
because the money is flowing into the,625.2,4.44
research it's happening what you I mean,626.76,4.38
we saw the same thing with solar and,629.64,2.759
literally every other disruptive,631.14,3.18
technology is once the investment comes,632.399,3.781
you know that the breakthroughs are,634.32,3.84
going to follow it's just that simple,636.18,3.779
and this is one of those kinds of,638.16,5.52
breakthroughs so what does this mean uh,639.959,7.261
put it this way rather than trying to,643.68,6.54
you know play Tetris with memory and you,647.22,4.98
know trying to fit 10 pounds of stuff,650.22,5.52
into a five pound bag now once this,652.2,5.699
becomes commercially ready which it's,655.74,4.8
coming it's it's possible on paper they,657.899,5.041
did it so even if we don't get a billion,660.54,4.16
tokens this time next year it's coming,662.94,4.92
the what this allows you to do is let's,664.7,4.96
say for instance,667.86,5.28
um you are working on a medical research,669.66,5.94
thing and it's like okay well you know,673.14,4.319
we've we've got a literature review of,675.6,4.799
literally 2000 papers per month to read,677.459,5.401
put all the papers in this model and and,680.399,4.321
say tell me exactly which papers are,682.86,3.96
most relevant,684.72,4.799
so the the the ability for in-context,686.82,5.94
learning uh is incredible and it can,689.519,5.88
hold more in its brain in its mind than,692.76,5.4
any 10 humans can,695.399,5.101
and this is again this is not the limit,698.16,4.44
imagine a year from now we're six months,700.5,4.62
from now when uh long net two comes out,702.6,4.2
and it's a trillion tokens or 10,705.12,4.32
trillion tokens and what they say in,706.8,5.46
this paper is that maybe we're gonna see,709.44,6.78
a a point very soon where it could have,712.26,6.0
its context window could include,716.22,4.619
basically the entire internet,718.26,5.28
this is a step towards super,720.839,4.981
intelligence make no mistake that the,723.54,4.739
ability to held and use that much,725.82,5.519
information in real time to produce,728.279,6.541
plans to forecast to anticipate to come,731.339,6.481
up with insights this is a critical step,734.82,6.18
towards digital super intelligence I am,737.82,5.22
not being hyperbolic here and neither is,741.0,3.48
this paper when they say we could,743.04,3.6
conceivably build a model that can read,744.48,4.38
the entire internet in one go,746.64,4.68
so with all that being said I wanted to,748.86,5.58
Pivot and talk briefly about open ai's,751.32,6.0
announcement also yesterday that they,754.44,5.22
are introducing super alignment so the,757.32,5.699
tldr is that openai is creating a a,759.66,5.58
dedicated team to aligning super,763.019,5.521
intelligence uh which you know again I,765.24,5.399
am super glad that we are living in the,768.54,3.359
timeline where someone is doing this,770.639,4.021
it's about time uh you know I've got my,771.899,4.261
book out there benevolent by Design,774.66,3.0
where I talked about aligning super,776.16,3.239
intelligence and my solution is that you,777.66,3.66
really can't but one thing that I want,779.399,5.521
to point out is that whether or not you,781.32,7.86
can align one model in the lab is that's,784.92,6.0
part of the that's a necessary part of,789.18,3.12
the solution I don't want to disparage,790.92,3.06
the engineers and scientists at openai,792.3,3.839
and Microsoft and other places working,793.98,4.859
on this but while it is a necessary,796.139,5.221
component of the of the solution it is,798.839,4.981
not a complete solution and this is,801.36,5.099
where uh researchers like Gary Marcus,803.82,6.12
and Dr Rahman Chowdhury have testified,806.459,6.961
to Congress saying look you know they,809.94,6.78
expect that open source models will,813.42,5.219
reach parity with closed Source models,816.72,4.619
and then overtake them and so when open,818.639,6.121
source models who anyone can deploy are,821.339,6.12
aligned any which way that you want you,824.76,5.759
lose total control so that while I,827.459,4.921
definitely appreciate in value because,830.519,3.601
we need to know how to align super,832.38,3.54
intelligent models,834.12,4.32
the good guys right the the aligned,835.92,6.78
models need to be uh as powerful as all,838.44,6.959
the unaligned models because in the AI,842.7,4.98
arms race it's going to be AI versus AI,845.399,4.68
in the example of cyber security where,847.68,5.159
we already have adaptive intelligence in,850.079,4.44
uh in firewalls and other security,852.839,4.44
appliances basically you're going to,854.519,5.041
have you know an AI agent running in,857.279,4.981
your firewall versus an AI based DDOS,859.56,5.16
attack just as one example you're going,862.26,4.74
to have ai based infiltration programs,864.72,4.52
versus AI Hunter programs on the inside,867.0,5.82
so it's going to be spy versus spy and,869.24,5.92
so we need to make sure that the that,872.82,5.16
the models that we build that that do,875.16,4.44
remain aligned that we do remain control,877.98,4.62
over are as smart as possible and also,879.6,5.34
trustworthy absolutely needs to to,882.6,4.44
happen basically you fight fire with,884.94,4.259
fire and I know that that sounds like,887.04,3.9
mutually assured destruction and it kind,889.199,3.901
of is which is another reason that the,890.94,4.56
nuclear arms race metaphor is very apt,893.1,4.56
for the AI arms race,895.5,4.26
So This research absolutely needs to,897.66,4.38
happen but what I want to drive home is,899.76,3.9
that it is a necessary but not,902.04,4.56
sufficient set of solutions that there,903.66,4.619
also needs to be the adoption the,906.6,3.72
implementation and deployment of,908.279,3.721
Alliance systems and we also need to,910.32,3.6
make sure that those Alliance systems,912.0,3.68
can communicate and collaborate together,913.92,5.099
so with all that being said uh big steps,915.68,4.839
in the right direction but it is coming,919.019,4.32
faster than anyone realizes and I stand,920.519,6.421
by my assertion AGI by the end of 2024,923.339,6.901
actually by by the mid basically uh,926.94,7.019
let's see September or October 2024 any,930.24,5.76
definition that you have of AGI will be,933.959,4.921
satisfied and then from there it's a,936.0,5.279
very very very short period of time to,938.88,4.8
Super intelligence now fortunately for,941.279,4.62
us right now the only computer is,943.68,3.959
capable of running them running these,945.899,4.141
models and researching them are like the,947.639,3.901
Nvidia supercomputers that they're,950.04,4.739
building so that but that barrier that,951.54,5.28
threshold is going to start going down,954.779,4.381
because remember remember Nvidia at,956.82,4.92
their at their keynote speech and and,959.16,4.08
for several months they've been saying,961.74,4.02
hey you know our machines are literally,963.24,4.2
a million times more powerful in the,965.76,2.879
last decade and we're going to do it,967.44,4.019
again in the next decade well when your,968.639,4.981
desktop computer is as powerful as,971.459,3.901
today's super computers in 10 years,973.62,3.719
you're going to be able to run all of,975.36,3.419
these and then when you combine that,977.339,3.0
with the ongoing algorithmic,978.779,3.481
efficiencies everyone is going to be,980.339,4.5
running their own AGI within five to ten,982.26,4.8
years mark my words,984.839,3.541
so,987.06,3.839
time is of the essence we do need a,988.38,4.5
sense of urgency and I am really glad,990.899,4.321
that open AI is doing this,992.88,6.78
again you know I'm not I would like to,995.22,6.9
see more governmental participation more,999.66,4.979
universities uh I would like to see,1002.12,5.1
something like a Gaia agency a global AI,1004.639,5.281
agency or an Aegis agency an alignment,1007.22,4.979
enforcement for Global uh Global,1009.92,4.38
intelligence systems,1012.199,4.921
um because the thing is is corporations,1014.3,5.12
and governments are not ready for this,1017.12,5.76
and uh that to me is the biggest risk,1019.42,6.039
because from it from a purely scientific,1022.88,6.059
standpoint I 100 believe that we can,1025.459,5.401
align super intelligence I wrote a book,1028.939,4.02
about it I demonstrated how you can take,1030.86,4.199
unaligned models and align them to,1032.959,4.021
Universal principles very very easily,1035.059,4.02
I've done it plenty of times the data,1036.98,3.54
sets are out there for free just search,1039.079,3.24
for heuristic imperatives and core,1040.52,3.84
objective functions on my GitHub,1042.319,5.821
but again aligning a single model is not,1044.36,6.12
the entire solution you also need the,1048.14,4.26
deployment you need the the security,1050.48,3.72
models we need to update things like the,1052.4,4.139
OSI model and defense in depth we need,1054.2,4.2
to look at the entire technology stack,1056.539,3.901
but we also need to look at the entire,1058.4,4.62
economic and governmental stack to make,1060.44,4.2
sure that companies are aware of this,1063.02,3.899
and that companies are start deploying,1064.64,5.7
uh these uh systems whether it's,1066.919,5.581
security checkpoints whether it's,1070.34,4.92
internal policies that sort of thing,1072.5,4.32
because when you've got a really,1075.26,3.6
powerful Cannon you have to aim that,1076.82,5.04
Cannon really really well otherwise it's,1078.86,6.24
going to kill everybody uh and again you,1081.86,5.04
all know me I am a very very very,1085.1,3.6
optimistic person when it comes to,1086.9,3.36
alignment and the future that we can,1088.7,4.32
build but at the same time when you're,1090.26,4.799
playing with fire like you need to make,1093.02,3.659
sure that you wear the proper safety,1095.059,3.961
gear uh because the the more energy,1096.679,4.321
something has the more dangerous it is,1099.02,4.8
and the level of energy or intelligence,1101.0,3.96
or however you want to look at it,1103.82,3.06
whatever metaphor you want to pick is,1104.96,4.2
going up very quickly so thanks for,1106.88,3.659
watching I hope you got a lot out of,1109.16,4.5
this it's the long net paper and then of,1110.539,5.121
course uh introducing super alignment,1113.66,6.259
but yeah thanks for watching cheers,1115.66,4.259