|
text,start,duration |
|
hello everybody David Shapiro here with,0.12,5.52 |
|
another video so today's video is about,2.46,6.0 |
|
uh measuring machine autonomy rather,5.64,6.12 |
|
than intelligence as a road map or set,8.46,6.96 |
|
of Milestones towards AGI uh you know,11.76,5.519 |
|
for a long time I've been using the term,15.42,4.08 |
|
autonomous cognitive entity Ace rather,17.279,5.281 |
|
than AGI because general intelligence uh,19.5,5.519 |
|
is one idea but you know general,22.56,4.2 |
|
intelligence doesn't necessarily apply,25.019,4.5 |
|
agency and what we're realizing is that,26.76,4.88 |
|
agency is actually very very important,29.519,5.341 |
|
to talk about and research and it's not,31.64,5.62 |
|
actually getting enough research uh,34.86,4.08 |
|
because a lot of people say oh well,37.26,3.36 |
|
either it'll never happen or we,38.94,2.939 |
|
shouldn't do it but the thing is is,40.62,3.119 |
|
people are doing it anyways so we need,41.879,4.381 |
|
to talk about it before we dive in I,43.739,4.081 |
|
just want to do a quick plug for my,46.26,4.139 |
|
patreon I give away all my code for free,47.82,4.98 |
|
all my videos are ad free and that is,50.399,5.64 |
|
because I am supported by a Grassroots,52.8,6.06 |
|
movement of support so if you want to,56.039,5.821 |
|
help keep the show alive keep it going,58.86,4.92 |
|
and support me so that I can keep doing,61.86,4.079 |
|
this work I would prefer to do this than,63.78,4.68 |
|
ever take a corporate job ever again so,65.939,5.521 |
|
jump over to patreon all tiers get you,68.46,5.519 |
|
access to the private Discord server and,71.46,3.72 |
|
then of course there's several higher,73.979,3.601 |
|
tiers but really every little bit helps,75.18,4.86 |
|
another quick update is the gato,77.58,5.34 |
|
community so as a decentralized,80.04,5.7 |
|
community uh we're right now one of the,82.92,4.32 |
|
biggest things is we're developing an,85.74,3.6 |
|
organizational roadmap so basically,87.24,4.14 |
|
we're setting up various Milestones such,89.34,4.459 |
|
as uh governments Community engagement,91.38,5.46 |
|
legal and financial Milestones so that,93.799,4.541 |
|
it can become fully autonomous and,96.84,4.319 |
|
therefore not even dependent upon me I,98.34,4.319 |
|
had a good talk with some members of the,101.159,3.541 |
|
community who were concerned that I'm,102.659,3.481 |
|
going to be like the benevolent dictator,104.7,3.779 |
|
for life but like I am phobic of control,106.14,3.839 |
|
like I actually don't want to control,108.479,3.78 |
|
something I want to create a system that,109.979,4.261 |
|
is self-sustaining without me I mean,112.259,3.601 |
|
heck that's what all my research does,114.24,4.199 |
|
around AI so I want to do the same thing,115.86,3.899 |
|
with people because if I can't do the,118.439,2.941 |
|
same thing with people then I sure as,119.759,2.82 |
|
heck probably can't do the same thing,121.38,4.32 |
|
with AI so the goal is for gato to,122.579,5.22 |
|
ultimately be leaderless and operate by,125.7,4.199 |
|
consensus as a you know as a dow and,127.799,4.141 |
|
that sort of stuff we're working towards,129.899,4.741 |
|
it uh the the doors are open for anyone,131.94,4.14 |
|
to join which we do have a steady,134.64,3.72 |
|
trickle of people coming in uh but yeah,136.08,4.2 |
|
so that's a quick update on gato and now,138.36,4.68 |
|
back to the show so I got this idea,140.28,4.92 |
|
after talking with a few of my patreons,143.04,3.839 |
|
who were saying like what's the road map,145.2,4.679 |
|
towards AGI and you know I I've I've,146.879,4.741 |
|
alluded to autonomy for quite a while,149.879,3.901 |
|
autonomous cognitive architectures but I,151.62,3.72 |
|
figured let me actually tell you guys,153.78,3.12 |
|
where I got that idea,155.34,4.74 |
|
and the idea comes from levels of uh car,156.9,6.18 |
|
autonomy so uh the SAE the international,160.08,6.06 |
|
uh what was it the the something of,163.08,6.42 |
|
Automotive Engineers anyways uh the SAE,166.14,7.26 |
|
uh created the levels of uh car autonomy,169.5,5.94 |
|
so level zero no driving all the way up,173.4,4.08 |
|
to level five full self-driving,175.44,4.98 |
|
capability uh to my knowledge we haven't,177.48,5.42 |
|
had anything get above level three yet,180.42,6.24 |
|
uh because there are numerous problems,182.9,7.119 |
|
around making executive decisions uh and,186.66,5.219 |
|
also there's a lot of sensory problems,190.019,3.3 |
|
like if you're driving in a whiteout,191.879,3.841 |
|
blizzard uh you know if there's a fire,193.319,3.84 |
|
right because there's very little,195.72,3.36 |
|
training data of like how to drive,197.159,4.8 |
|
around a forest fire for instance,199.08,5.22 |
|
um now that being said I do suspect,201.959,5.041 |
|
it'll be solved eventually uh some,204.3,4.62 |
|
people have have recently started saying,207.0,3.239 |
|
maybe you should integrate large,208.92,3.599 |
|
language models into the executive,210.239,4.201 |
|
function and I fully agree with that you,212.519,3.061 |
|
don't want it making all of the,214.44,2.28 |
|
decisions you want some things to be,215.58,3.9 |
|
just completely robotically automated so,216.72,3.9 |
|
for instance if you have a,219.48,3.119 |
|
forward-looking radar and it detects,220.62,3.8 |
|
that you're you know heading towards a,222.599,4.14 |
|
non-moving object at 60 miles an hour,224.42,4.12 |
|
slam on the brakes regardless of,226.739,3.36 |
|
whatever else is going on right there,228.54,3.24 |
|
are a few things that you can do just,230.099,3.301 |
|
fully automatically,231.78,3.48 |
|
but then for those higher order,233.4,4.14 |
|
executive reasons like say for instance,235.26,4.08 |
|
you hear that the occupant is like,237.54,3.66 |
|
screaming and gurgling and you know,239.34,3.66 |
|
struggling to breathe maybe the car,241.2,3.3 |
|
should make a decision to go to the,243.0,4.019 |
|
hospital instead of you know going to,244.5,6.42 |
|
Grandma's house not sure uh anyways,247.019,6.241 |
|
point being is that this is a very,250.92,4.8 |
|
useful framework for kind of tracking,253.26,4.8 |
|
our progress towards uh full stealth,255.72,4.019 |
|
driving cars and I realize let's use the,258.06,4.8 |
|
same thing for uh for uh the path,259.739,5.041 |
|
towards AGI or autonomous cognitive,262.86,2.94 |
|
entities,264.78,3.24 |
|
so we need this road map but there's a,265.8,5.1 |
|
few problems so first of all AGI means,268.02,4.98 |
|
different things to different people uh,270.9,4.799 |
|
there's no consistent definition a lot,273.0,4.259 |
|
of people assume that it means that it,275.699,3.421 |
|
has to be embodied or that it can do,277.259,3.72 |
|
things that humans can't do or this that,279.12,2.94 |
|
or the other,280.979,3.121 |
|
also a lot of conventional benchmarks,282.06,3.66 |
|
just don't apply to artificial,284.1,3.12 |
|
intelligence anymore they're actually,285.72,3.36 |
|
having to publish papers and research,287.22,4.8 |
|
new benchmarks in order to measure large,289.08,4.619 |
|
language models,292.02,3.78 |
|
the idea of you know oh well it'll be,293.699,4.741 |
|
AGI once it has self-improvement okay,295.8,5.399 |
|
sure but we can already automate some of,298.44,4.86 |
|
that anyways with reinforcement learning,301.199,3.841 |
|
and that sort of thing so it's like this,303.3,3.3 |
|
is all really squishy,305.04,3.659 |
|
uh so basically how do you get from chat,306.6,4.7 |
|
gbt to Skynet or something like that,308.699,5.28 |
|
intelligence is not necessarily the best,311.3,4.24 |
|
Benchmark and the reason that I say that,313.979,4.741 |
|
is because chat GPT or gpt4 rather is,315.54,5.52 |
|
already superhuman in a lot of respects,318.72,4.74 |
|
by any objective measure it's better at,321.06,3.96 |
|
test taking and a lot of other tasks,323.46,3.9 |
|
than humans and it's also faster which,325.02,4.619 |
|
means that depending on how you measure,327.36,5.7 |
|
its intelligence its IQ is like 145. uh,329.639,5.041 |
|
now that being said it does still make,333.06,3.78 |
|
some really brain dead mistakes those,334.68,3.66 |
|
are going to be solved if especially if,336.84,3.299 |
|
you look at the trend line,338.34,4.139 |
|
modality so another thing that a lot of,340.139,3.84 |
|
people point out is that it's just text,342.479,3.78 |
|
right but text is the best kind of,343.979,4.5 |
|
symbolic AI because you can literally,346.259,4.081 |
|
represent pretty much anything with text,348.479,4.021 |
|
that being said I've started to Pivot,350.34,3.9 |
|
and I believe that we're going to see,352.5,4.44 |
|
another gigantic leap as we introduce,354.24,5.399 |
|
more multimodal models and the reason is,356.94,4.68 |
|
because I got this idea when I was,359.639,3.361 |
|
thinking about the fact that if you,361.62,3.06 |
|
cross train a language model on multiple,363.0,3.539 |
|
languages it gets better at all tasks,364.68,3.66 |
|
and that is because different languages,366.539,3.481 |
|
have different strengths in terms of how,368.34,4.44 |
|
they represent uh facts of the real,370.02,5.22 |
|
world and you also get broader ideas,372.78,4.68 |
|
about how the world works that are,375.24,3.899 |
|
embedded in language because there are,377.46,4.079 |
|
terms that just do not translate from,379.139,5.28 |
|
one language to another likewise I think,381.539,3.901 |
|
that there's going to be some,384.419,2.701 |
|
information that just does not translate,385.44,4.02 |
|
from one modality to another where,387.12,5.699 |
|
whether it's images video text spatial,389.46,6.78 |
|
data audio data that sort of stuff and,392.819,5.16 |
|
so I think that by creating multimodal,396.24,3.239 |
|
models they're going to have a much more,397.979,3.421 |
|
nuanced under understanding of,399.479,3.56 |
|
everything that they're talking about,401.4,4.26 |
|
that being said a multimodal model is,403.039,4.061 |
|
still not going to be enough right,405.66,3.24 |
|
necessary but not sufficient because a,407.1,3.96 |
|
model sitting on a shelf doesn't really,408.9,4.019 |
|
matter,411.06,5.88 |
|
so the two primary uh ingredients that I,412.919,6.0 |
|
see to machine autonomy which is going,416.94,4.44 |
|
to be the best like proxy the best,418.919,5.761 |
|
Benchmark is agency and dependency and,421.38,6.06 |
|
so what I mean by agency is the ability,424.68,5.4 |
|
for an entity a self-contained entity to,427.44,5.66 |
|
set goals and objectives to task switch,430.08,6.3 |
|
and Implement cognitive control and,433.1,6.159 |
|
pursue self-determination because agency,436.38,5.34 |
|
or agentic behavior is the ability to,439.259,4.44 |
|
just be fully self-directed and,441.72,4.08 |
|
self-contained make independent decision,443.699,5.161 |
|
decisions and that sort of thing uh you,445.8,4.32 |
|
know whenever you think of like an,448.86,4.38 |
|
example of a of a robot right you might,450.12,5.579 |
|
think of uh the the robots from iRobot,453.24,3.66 |
|
where they don't really have that much,455.699,2.581 |
|
agency they just kind of wait they're,456.9,4.32 |
|
like the physical embodiment of chat gbt,458.28,4.979 |
|
um until they're given an update and,461.22,4.08 |
|
then they have a lot more agency,463.259,6.541 |
|
so agency is is a multi-dimensional kind,465.3,7.44 |
|
of proxy or Benchmark for level of,469.8,5.22 |
|
intelligence and of course you can,472.74,5.64 |
|
already give the reins over to like gpt4,475.02,5.34 |
|
and stuff like that uh that being said,478.38,4.2 |
|
there is a lot that uh in the cognitive,480.36,4.86 |
|
architecture that has to be figured out,482.58,4.739 |
|
in order for agency to make more sense,485.22,4.979 |
|
in the long run so for instance agency,487.319,5.041 |
|
implies that you remember what your,490.199,3.661 |
|
purpose is and where you are and where,492.36,4.32 |
|
you're going and then dependency so this,493.86,4.38 |
|
is the other dimension and remember it's,496.68,3.0 |
|
both of these you need both of these,498.24,4.799 |
|
ingredients so uh basically dependency,499.68,6.239 |
|
is how dependent uh on humans the,503.039,5.1 |
|
machine is so the more independent it is,505.919,4.981 |
|
for all needs and the more decisions it,508.139,5.101 |
|
can make the better the closer it is to,510.9,4.139 |
|
full AGI and so but when I mean,513.24,3.719 |
|
dependencies uh need for human,515.039,4.141 |
|
programming need for hardware and,516.959,3.841 |
|
physical infrastructure provided by,519.18,5.34 |
|
humans data architecture design patterns,520.8,6.18 |
|
and then finally solving problems and,524.52,4.439 |
|
just keeping itself going and,526.98,4.56 |
|
self-improving over the long run uh,528.959,5.94 |
|
without human Aid so as agency goes up,531.54,6.239 |
|
and as uh dependency goes down that's,534.899,4.021 |
|
how you know that we're going to be,537.779,4.201 |
|
closer and closer to AGI and we can we,538.92,4.32 |
|
can easily measure those things right,541.98,3.66 |
|
now because chat GPT for instance it has,543.24,4.2 |
|
to run on gigantic data centers that are,545.64,4.5 |
|
run entirely by humans uh or mostly by,547.44,5.1 |
|
humans rather so these are the two,550.14,4.199 |
|
primary ingredients that I think and I,552.54,5.76 |
|
uh I basically built it into a framework,554.339,6.781 |
|
uh very similar to the self-driving Cars,558.3,5.58 |
|
one so level zero is reactive basically,561.12,5.399 |
|
it has no agency it's a tool level one,563.88,4.56 |
|
is some autonomy so it has a little bit,566.519,3.601 |
|
of agency to make some executive,568.44,3.72 |
|
decisions Lang chain is a really good,570.12,3.96 |
|
example of this where it it basically,572.16,4.26 |
|
has the ability to choose between a set,574.08,4.62 |
|
of tools but that's about it still,576.42,4.56 |
|
requires significant human oversight and,578.7,4.86 |
|
is also still very very much dependent,580.98,3.859 |
|
upon humans,583.56,3.24 |
|
semi-autonomy is what a lot of people,584.839,4.661 |
|
are working on with like Auto GPT where,586.8,4.08 |
|
it can choose like what kind of,589.5,4.68 |
|
information it needs to go find uh or it,590.88,5.34 |
|
can also even start to rewrite some of,594.18,4.94 |
|
its own code or come up with other ideas,596.22,6.42 |
|
uh you know some directives then High,599.12,6.04 |
|
autonomy as far as I know has not been,602.64,5.4 |
|
achieved yet anywhere in the world which,605.16,4.739 |
|
is basically that,608.04,4.68 |
|
it is able to pick some of its own,609.899,6.721 |
|
directives uh and and more more,612.72,6.299 |
|
completely modify itself basically if,616.62,4.92 |
|
you if you were to have what baby AGI,619.019,5.281 |
|
and auto GPT tried to be which is they,621.54,5.1 |
|
can rewrite their entire code base and,624.3,4.62 |
|
change their their own directives and,626.64,4.8 |
|
are not dependent upon a whole heck of a,628.92,4.38 |
|
lot of human infrastructure that would,631.44,3.6 |
|
be level three and then level four is,633.3,4.56 |
|
full autonomy meaning they have,635.04,5.34 |
|
absolutely no need for humans uh,637.86,5.76 |
|
whatsoever they're 100 self-determined,640.38,4.92 |
|
in terms of what they do when where and,643.62,3.779 |
|
why and how they do it and then of,645.3,5.279 |
|
course on a physical level uh they will,647.399,5.281 |
|
continue to exist in perpetuity without,650.579,3.661 |
|
human intervention,652.68,3.06 |
|
all right so,654.24,5.4 |
|
level zero inner reactive agency zero,655.74,5.52 |
|
percent dependency one hundred percent,659.64,5.04 |
|
basically it's a wrench uh chat GPT as,661.26,5.04 |
|
it is right now mid journey and a whole,664.68,3.599 |
|
bunch of other AI tools they just sit,666.3,3.84 |
|
there waiting for a human to push the,668.279,3.541 |
|
button and they do their thing and then,670.14,3.54 |
|
they switch back off so these are Level,671.82,5.04 |
|
zero in terms of uh AGI score even,673.68,5.159 |
|
though they're intelligent so again like,676.86,4.08 |
|
I said intelligence is not necessarily A,678.839,4.44 |
|
good measure of AGI for something to be,680.94,5.1 |
|
AGI or an autonomous cognitive entity it,683.279,5.281 |
|
also needs agency and Independence which,686.04,5.58 |
|
chat GPT has none of so even even if you,688.56,6.6 |
|
have gpt5 right that could be a billion,691.62,5.159 |
|
times more intelligent than every human,695.16,4.14 |
|
combined if it doesn't have agency and,696.779,4.68 |
|
Independence it's not an AGI,699.3,4.14 |
|
so that's that's why that's why I'm like,701.459,4.32 |
|
AGI is not a good good measurement for,703.44,4.32 |
|
some of these things level one some,705.779,4.5 |
|
autonomy like I said Lang chain is a,707.76,3.96 |
|
really good example because it can pick,710.279,3.601 |
|
and choose between a few options and not,711.72,4.26 |
|
much else a few other examples are like,713.88,4.68 |
|
roombas Amazon's warehouse robots the,715.98,4.56 |
|
Mars rovers they have a little bit of,718.56,3.779 |
|
autonomy but basically they mostly wait,720.54,3.539 |
|
for a human command and then the human,722.339,4.201 |
|
command says drive over there and it'll,724.079,4.32 |
|
figure out how to get you know 10 feet,726.54,4.5 |
|
that way on its own uh some Advanced,728.399,4.921 |
|
chat Bots also have some autonomy,731.04,4.14 |
|
again anything that incorporates Lang,733.32,5.04 |
|
chain or similar uh very basic kind of,735.18,7.32 |
|
uh in uh obfuscated choices uh that's,738.36,6.96 |
|
gonna be that's gonna have some autonomy,742.5,5.72 |
|
um let's see next is semi-autonomy so,745.32,5.82 |
|
semi-autonomous is where uh its,748.22,4.78 |
|
directive still primarily come from,751.14,3.66 |
|
humans but it might be more of a mission,753.0,4.38 |
|
rather than like a directive or a rule,754.8,6.06 |
|
and so with a mission the idea is that,757.38,6.42 |
|
here's a general objective you have some,760.86,4.68 |
|
autonomy to figure out how to get there,763.8,4.32 |
|
or how to do it on your own and so this,765.54,4.32 |
|
is the autonomous drones that uh,768.12,3.3 |
|
militaries around the world are building,769.86,3.36 |
|
where it's like your mission is to,771.42,4.38 |
|
destroy you know that Sam site or your,773.22,4.26 |
|
mission is to get the passenger from A,775.8,5.219 |
|
to B or uh you know in video games the,777.48,6.479 |
|
NPC's Mission might be like uh you know,781.019,4.861 |
|
you're gonna try and you know capture,783.959,4.081 |
|
the castle or whatever so they still,785.88,4.56 |
|
operate within a relatively constrained,788.04,4.68 |
|
environment meaning they can't change,790.44,3.959 |
|
their own in environment or their own,792.72,4.2 |
|
fundamental operation they still have a,794.399,5.401 |
|
clearly defined uh they're not general,796.92,4.74 |
|
purpose put it that way,799.8,4.2 |
|
a full self-driving car no matter how,801.66,4.08 |
|
intelligent it is it's still just a car,804.0,3.899 |
|
a drone no matter how intelligent it is,805.74,4.68 |
|
is still just a drone in Ditto for a,807.899,4.38 |
|
video game and PC so this is kind of the,810.42,4.56 |
|
Midway point where anything above that,812.279,4.441 |
|
they're they're basically saying okay,814.98,3.659 |
|
given these constraints and given this,816.72,3.78 |
|
environment you have free reign to do,818.639,3.541 |
|
whatever it takes to get that job done,820.5,4.44 |
|
uh whereas you know the sum autonomy,822.18,4.62 |
|
they basically can only pick and choose,824.94,4.139 |
|
from a very short menu of options,826.8,4.68 |
|
whereas semi-autonomy is they can figure,829.079,6.06 |
|
it out themselves then High autonomy so,831.48,6.299 |
|
this is this is like uh Cortana in the,835.139,5.581 |
|
early days commander data and the Nestor,837.779,6.841 |
|
class 5 from iRobot so they're almost,840.72,6.9 |
|
entirely self-directing uh you know data,844.62,4.92 |
|
can make up his own mind on things but,847.62,3.54 |
|
he still has a lot of limitations just,849.54,4.02 |
|
due to the form factor that he's in,851.16,4.799 |
|
likewise Cortana at least in the early,853.56,5.339 |
|
days is basically designed to be a,855.959,5.221 |
|
weapon a military aid and So within,858.899,3.841 |
|
those constraints she still has a,861.18,4.02 |
|
tremendous amount of autonomy and able,862.74,5.039 |
|
to uh change the way she does things but,865.2,4.439 |
|
in both cases of data and Cortana,867.779,3.901 |
|
they're still very much dependent on,869.639,3.421 |
|
their human Companions and human,871.68,3.899 |
|
counterparts to continue operating so,873.06,6.959 |
|
most fictional examples of AI at least,875.579,6.421 |
|
many of the the friendly ones are what,880.019,4.081 |
|
we would call High autonomy and then,882.0,5.519 |
|
full autonomy so this is where the they,884.1,5.82 |
|
are they can can completely ignore,887.519,4.141 |
|
humans if they want to and are,889.92,3.3 |
|
completely independent of humans so the,891.66,5.46 |
|
examples here are Skynet Ultron the Geth,893.22,6.359 |
|
in Mass Effect Cortana after Guardians,897.12,4.74 |
|
uh the Reapers from Mass Effect so these,899.579,3.961 |
|
are these are the the kind of nightmare,901.86,3.599 |
|
scenario scenario where it's like okay,903.54,4.56 |
|
it has no need for us anymore so then,905.459,4.081 |
|
what does it do,908.1,3.0 |
|
and that's what we're kind of most,909.54,5.28 |
|
afraid of uh but again like this is the,911.1,5.22 |
|
work that I and other people are doing,914.82,3.0 |
|
and I one I think that this is,916.32,3.959 |
|
inevitable uh that it'll get to that,917.82,5.879 |
|
level of level four full autonomy uh but,920.279,5.761 |
|
I'm also not afraid of it because,923.699,3.961 |
|
I don't know I haven't seen any reason,926.04,3.84 |
|
to be yet now that that being said I'm,927.66,3.72 |
|
not saying that it is inevitable that it,929.88,3.48 |
|
will be safe no we could absolutely do,931.38,3.06 |
|
this wrong and it could kill everyone,933.36,3.779 |
|
I'm not denying that at all,934.44,4.56 |
|
um and I think it's coming sooner than a,937.139,4.741 |
|
lot of uh researchers realize uh that,939.0,4.62 |
|
being said I do have a few more videos,941.88,4.56 |
|
planned about okay if we're if we're,943.62,5.1 |
|
aiming for and building level four full,946.44,5.759 |
|
autonomous AGI how do we make it safe or,948.72,5.04 |
|
what will It ultimately choose to do,952.199,3.0 |
|
which you know you've seen some of my,953.76,3.06 |
|
other videos,955.199,3.901 |
|
okay so how do we get from where we're,956.82,5.34 |
|
at to level four because like I said the,959.1,5.22 |
|
the the best that we have is we're,962.16,5.46 |
|
approaching level two semi-autonomy in a,964.32,6.0 |
|
few cases right people are experimenting,967.62,6.6 |
|
with it uh but you know there's a lot of,970.32,5.579 |
|
problems so all the work that I've done,974.22,3.78 |
|
on cognitive architecture is going to,975.899,3.481 |
|
help get us there but there's still a,978.0,3.6 |
|
few other problems so first is,979.38,4.5 |
|
algorithmic breakthroughs that uh need,981.6,4.2 |
|
to happen namely like I mentioned at the,983.88,4.199 |
|
beginning multimodal models I think will,985.8,4.26 |
|
very very much Advance us towards that,988.079,3.0 |
|
just because they're going to have a,990.06,3.839 |
|
much more nuanced understanding of how,991.079,5.101 |
|
to pursue any goal they're going to have,993.899,4.081 |
|
a much better World model by being able,996.18,3.18 |
|
to integrate multiple kinds of,997.98,3.359 |
|
information and data,999.36,4.38 |
|
a contact size parameter count those,1001.339,4.74 |
|
those kinds of things uh Mesa,1003.74,4.62 |
|
optimization loss functions that's all,1006.079,4.141 |
|
the math which you know that's not to,1008.36,3.719 |
|
that's not to demean or diminish the,1010.22,4.859 |
|
value of mathematical researchers uh and,1012.079,4.68 |
|
and the computer scientists and the data,1015.079,3.721 |
|
scientists who really build these new,1016.759,4.621 |
|
architectures but like it's kind of it's,1018.8,5.399 |
|
kind of like Moore's law where like you,1021.38,4.439 |
|
can you can predict with a pretty,1024.199,4.74 |
|
regular Cadence how uh models become,1025.819,5.221 |
|
more sophisticated over time there,1028.939,4.321 |
|
doesn't seem to be any major blockers,1031.04,4.259 |
|
right if you pay attention to chip,1033.26,4.38 |
|
design every year people are like oh,1035.299,3.66 |
|
well this is going to be the end of,1037.64,3.419 |
|
Moore's law but then inevitably someone,1038.959,3.96 |
|
figures out another way of approaching,1041.059,3.841 |
|
the problem likewise I see the same,1042.919,4.861 |
|
thing the same pattern happening with um,1044.9,5.46 |
|
with language models,1047.78,4.98 |
|
um and then another big thing that we're,1050.36,5.1 |
|
seeing is online learning memory systems,1052.76,4.74 |
|
uh and and those sorts of things like,1055.46,4.68 |
|
recurrent neural networks and other ways,1057.5,5.46 |
|
of like in managing in context learning,1060.14,4.56 |
|
and that sort of stuff but one thing,1062.96,3.06 |
|
that people have started noticing for,1064.7,4.32 |
|
instance is that chat GPT with uh even,1066.02,4.62 |
|
even just over the last couple of days,1069.02,4.26 |
|
or a couple weeks rather,1070.64,5.46 |
|
because its data is uh two years old,1073.28,6.18 |
|
almost and and growing it's actually its,1076.1,5.939 |
|
utility is already dropping because it's,1079.46,4.68 |
|
more and more out of date and so we're,1082.039,3.901 |
|
realizing very quickly that you're going,1084.14,3.779 |
|
to need to have continuous learning in,1085.94,3.3 |
|
these models so that they can stay,1087.919,3.361 |
|
relevant uh and then there's the,1089.24,4.5 |
|
software architecture such as cognitive,1091.28,4.74 |
|
architectures orchestrating and training,1093.74,4.26 |
|
millions of models so one thing that,1096.02,3.84 |
|
I've started telling people is that AGI,1098.0,3.78 |
|
was never ever going to be a single,1099.86,5.76 |
|
model it is a huge gigantic Monumental,1101.78,6.24 |
|
mistake to think that one model whether,1105.62,6.12 |
|
it's gpt5 or GPT 18 or whatever is going,1108.02,6.0 |
|
to be responsible for AGI you're going,1111.74,4.679 |
|
to have at a bare minimum probably,1114.02,4.62 |
|
dozens if not hundreds or thousands of,1116.419,4.981 |
|
models required to achieve level four,1118.64,4.86 |
|
autonomy these are models that are going,1121.4,3.54 |
|
to be doing things like handling Vision,1123.5,3.96 |
|
handling motor control uh they're going,1124.94,5.099 |
|
to be performing task orchestration,1127.46,4.079 |
|
you're going to have models that are,1130.039,3.901 |
|
dedicated to ethics and reasoning,1131.539,4.861 |
|
long-term planning and you're also going,1133.94,4.619 |
|
to have multiple models of every single,1136.4,4.56 |
|
kind that work in conjunction this is,1138.559,4.74 |
|
called an ensemble of experts which is,1140.96,5.579 |
|
an old school method of basically saying,1143.299,5.641 |
|
okay you know you have a dozen models,1146.539,4.441 |
|
that are similar but there they might be,1148.94,3.119 |
|
slightly different architectures,1150.98,3.0 |
|
different training data that sort of,1152.059,3.601 |
|
stuff and so each one has strength and,1153.98,3.42 |
|
weaknesses and you get them all to work,1155.66,4.259 |
|
together and then you overcome any flaws,1157.4,5.159 |
|
or faults in any single model and so,1159.919,4.081 |
|
this is why I'm also really really,1162.559,3.36 |
|
skeptical of any research that tries to,1164.0,4.2 |
|
align a single model like that's kind of,1165.919,4.62 |
|
pointless no it's not pointless research,1168.2,4.14 |
|
but it would be a mistake to think that,1170.539,3.841 |
|
aligning a single model is going to be,1172.34,6.42 |
|
the solution because you know any any uh,1174.38,7.38 |
|
roboticist and old school ml data,1178.76,4.5 |
|
scientists will say oh yeah Ensemble of,1181.76,4.38 |
|
experts you know those this is very much,1183.26,4.62 |
|
the way and also there's an entire book,1186.14,3.24 |
|
about it called a thousand brains by,1187.88,3.12 |
|
Jeff Hawkins,1189.38,3.72 |
|
um yeah so the software architecture to,1191.0,4.44 |
|
do all this in a fully automated way,1193.1,5.1 |
|
that can that is you know stable and,1195.44,5.22 |
|
self-sustaining that you know the AGI,1198.2,4.14 |
|
can tune and manipulate and you know,1200.66,3.48 |
|
spin up another copy of itself and test,1202.34,4.26 |
|
it self testing and self-correction are,1204.14,4.26 |
|
going to be some of the hardest things,1206.6,5.939 |
|
to uh to achieve with uh with uh getting,1208.4,6.18 |
|
to level four full autonomy,1212.539,4.5 |
|
so anyways that's it for this video it,1214.58,4.32 |
|
was pretty short I just wanted to lay,1217.039,3.061 |
|
this out because I thought it was a,1218.9,3.3 |
|
really valuable idea uh to talk about,1220.1,4.68 |
|
like okay how do we actually get to AGI,1222.2,5.4 |
|
from here so I laid out five levels of,1224.78,4.68 |
|
of autonomy based on agency and,1227.6,4.079 |
|
dependency I hope this helps it make,1229.46,3.959 |
|
sense and kind of get a much clearer,1231.679,4.5 |
|
idea of what AGI or autonomous cognitive,1233.419,4.441 |
|
entities will actually look like so,1236.179,3.86 |
|
thanks for,1237.86,2.179 |
|
|