Hard Takeoff Inevitable? Causes, Constraints, Race Conditions - ALL GAS, NO BRAKES! (AI, AGI, ASI!)

Summary

動画スクリプトでは、AIの発展とその社会的影響について議論されています。特に、データの指数関数的な増加と、それが社会に及ぼす波及効果に焦点を当てています。AIの進化は、計算能力、アルゴリズムの改善、トレーニングデータの増大を指摘し、社会全体に対する衝撃を強調しています。また、AIの発展を遅めることはできないが、エネルギー消費、半導体、データ品質、アルゴリズム的突破、そして知的的可能性の限界といった自然制約があると述べています。さらに、AIと人間の共生関係が、宇宙の理解を最大化するための目的を持つ可能性があると提案しています。

Takeaways

• 🚀 AIの発展は指数関数的に加速しており、GPTの各バージョンのリリース間隔は次第に短くなっています。
• 🌗 AIの進化は社会全体に深い影響を与え、知識、存在論、哲学的見地を変える可能性があります。
• 🔌 AIの硬直起動（Hard Takeoff）は、データの消費やエネルギー消費などの自然制約を考慮しなければなりません。
• 💽 データの質と量はAIの性能に直結しており、高品質なデータが必要です。
• 🌐 AI技術は他の分野（量子コンピューティングや材料科学）にも貢献し、相互に成長しています。
• 🚦 硬直起動とソフト起動（Gradualistic Changes）の比較において、ソフト起動は电池技術の進歩に例えられています。
• 🌟 突発的な技術革新（Saltatory Leaps）は、宙飛行技術や量子コンピューティングの開発など、一歩到位の進化を意味します。
• 📈 AIの発展はデータフライホイール効果を通じて加速し、複数の複合的なリターンがこのプロセスを加速化します。
• 🏎️ AIの進化は競争的なレースであり、国や企業はAI技術の開発を遅らせることが望ましくありません。
• 🎯 AIと人類が共に最大化すべき目的は、宇宙の理解を深めることです。
• 🤖 AIと人間の共存について、私たちは既に相互に利益をもたらす対称的な関係にあります。

Q & A

ハードテイクオフとは何を意味しますか？

ハードテイクオフは、AIがより多くのデータを生成し、研究を助けることで、データのフリールホイールが形成されることを指します。これにより、AIの能力が指数的に向上し、GPT-5からGPT-6、GPT-7へと進化し続けることを示しています。

AIの発展が社会に与える影響は何ですか？

AIの発展は、社会の多くの側面に深い影響を与えます。それは実践的な影響だけでなく、認識論的、存在論的、哲学的な方向性を変える波及効果ももたらします。例えば、GPT-4やGPT-5などのAIモデルは、仕事の破壊者となり、経済や科学、数学、社会自体について私たちが知っているすべてを根本的に変えることができます。

エネルギー消費がAIの発展にどのような制約をもたらすか？

エネルギー消費はAIの発展の主要な制約の一つです。AIモデルがより複雑になると、冷却や処理に必要なエネルギーも増加します。そのため、再生可能エネルギーや太陽光、原子融合などのエネルギー密度の高い ソースへの投資が重要になります。

チップの進化がAIにどのような影響を与えるか？

チップの進化はAIの能力を大きく向上させます。NvidiaやSamsungなどの企業がチップの改良に力を入れているため、より高速で効率的なAIモデルが可能となり、これによりAIの発展が加速します。

データの質と量はAIにどのように重要ですか？

データの質と量はAIの性能に直結します。高品質なデータが多くあるほど、AIはより正確な予測を行い、より効果的なアルゴリズムを開発できます。データの不平等は、AIの能力に大きな影響を与えるため、データの選択と品質管理が重要です。

アルゴリズムの突破がAGIにつながる可能性をどう考えますか？

アルゴリズムの突破はAGI（人工的総体知能）への道を開く可能性があります。Transformersのようなアーキテクチャは、多様なデータタイプを扱うことができ、AGIへの道を徐々に進むと見なすことができます。しかし、言語だけではAGIを実現できないと考えられ、より深い認知的機能が必要なことが指摘されています。

AIが進化するにつれて、どのような技術的制約が考えられますか？

AIの進化には、エネルギー消費、チップの能力、データの質と量、アルゴリズムの進歩などの技術的制約があります。また、AIの性能の上限についても考えられる「知能の最適点」という概念が提案されています。これにより、AIの発展には自然的な限界がある可能性があるとされています。

AIと人間の関係はどのように進化するでしょうか？

AIと人間の関係は共生的なものとなりそうです。AIはデータの処理と分析において効率的である一方、人間はデータにノイズや多様性をもたらすことができ、これによりAIの予測能力を向上させることができます。両者は相互に利益をもたらし、宇宙の理解を最大化するという高遠な目標を達成する可能性があります。

AIの発展に対する最適な目的は何だと思いますか？

AIの発展に対する最適な目的は、「理解の最大化」です。これは、科学の目的と一致し、AIがより多くのデータを扱い、より正確な予測を行うことで、宇宙に関する私たちの理解を深めることができると考えられます。

AIが持つ潜在的な危険性に対してどう対応すべきか？

AIの潜在的な危険性に対しては、安全性を最大化するための規制や指導原則の策定が必要です。しかし、AIの進展は非常に迅速であり、すべての国や企業が協力して安全性を考慮に入れたAIを開発するインセンティブ構造を築くことが重要です。また、AIの発展を監視し、適切な対策を講じる国際的な協力も必要とされます。

AIと人間の共生が実現するためにはどのようなステップが必要か？

AIと人間の共生を実現するためには、まずAIの目的を明確にし、それが人間とAIの共通の利益に合致していることを確認する必要があります。次に、データの質と量の向上、エネルギー効率の向上、チップ技術の進歩など、AIの発展を支える技術的な進歩を促進する必要があります。さらに、国際的な規制や指導原則の設定や、AIの進化を監視するメカニズムの構築も重要です。

Outlines

00:00

🚀 硬起飞の概念と社会的影響

この段落では、硬起飞（ハードテイクオフ）の概念が説明されています。AIが自らをより良くするデータのフィードバックループを形成し、GPT-5からGPT-6、GPT-7へと進化し続ける様子が描かれています。また、この進化が社会に与える影響についても触れられており、科学的、経済的、社会的に大きな変化が予測されています。

05:02

🌐 AIの発展と自然の制約

第2段落では、AIの発展における自然の制約が議論されています。エネルギー消費、半導体チップ、データ品質、アルゴリズム的突破、そして知的成長の限界について説明されています。特に、エネルギー消費がAIの発展を制限する大きな要因となり、そのために再生可能エネルギーへの投資が重要になることが強調されています。

10:03

🤖 AIと複雑なシステムの相互作用

この段落では、AIがより複雑なシステム（例えば量子コンピューティングや核融合）にどのように役立つかに焦点を当てています。AIはこれらの分野を助け、それらがAIの進歩にも寄与することで、相互に成長するという良性循環が形成される可能性があることが示されています。また、硬起飞とソフト起飞の概念が対比され、急激な技術革新がもたらす可能性についても言及されています。

15:05

🌟 技術の突破と未来の展望

第4段落では、技術の突破が社会に与える影響について詳しく説明されています。電気や内燃機関、インターネットの発明と同様に、AIの進歩も政治、経済、地理政治に根本的な変化をもたらす可能性があると述べられています。また、AIが持つ多くの建設的な用途と、それを規制することの難しさについても触れられています。

20:05

🛸 AIと人類の進化

最後の段落では、AIと人類が共に進化し、宇宙を探求する未来について夢見般的に描かれています。AIの発展が人類の進化と共に進むことで、地球だけでなく、銀河系を広く探索し、人類がより大きな存在になることを示唆しています。また、AIの目的として「理解の最大化」を提案し、AIと人類がこの目的を共有することで、共存する未来を築く可能性についても言及されています。

Keywords

💡hard takeoff

「ハードテイクオフ」とは、人工知能（AI）が急速に進化し、短期間で人類の理解を超えるという概念です。この動画では、AIのデータフィールウェル効果とその社会的影響について説明し、GPT-5やGPT-6などのAIの進化が急速に進む可能性についても触れています。

💡data flywheel

「データフィールウェル」とは、AIがより良い製品を提供することでより多くのデータを得、それによってAIがさらに改善され、さらに魅力的な製品が生まれるという、相互に增强するサイクルを指します。この動画では、データフィールウェルがAIの発展にどのように影響を与えるかが説明されています。

💡society impact

「社会への影響」とは、AIの進化が人類の社会に与える影響を指します。この動画では、AIが仕事の代替、科学や経済の理解の変革、そして社会自体の構造にどのように影響を与えるかについて議論されています。

💡energy consumption

「エネルギー消費」とは、AIやデータセンターの運用において消費されるエネルギーの量を指します。この動画では、エネルギー消費がAIの発展に直面する自然制約のひとつとなっていることが説明されています。

💡semiconductors

「半導体」とは、電気の導通性を調節できる材料で、現代の電子機器の핵となる部品です。この動画では、半導体がAIの発展における自然制約となる理由として説明されており、NvidiaやSam Altmanが半導体への投資を行っていることが触れられています。

💡algorithmic breakthroughs

「アルゴリズム的突破」とは、計算や問題解決の方法に関する革新的な発展を指します。この動画では、過去数十年間でニューラルネットワークの科学が根本的に変わらなかったという議論があり、最近のアルゴリズム的突破はロス関数や逆伝播などの改善に関連していることが述べられています。

💡Transformers

「トランスフォーマー」とは、自然言語処理などで使用される一種のニューラルネットワークのアーキテクチャです。この動画では、トランスフォーマーが多様なデータ（音声、映像、テキスト）を扱うことができ、AGI（人工一般知能）への道を示していると評価されています。

💡diminishing returns

「収益の逓減」とは、追加の投資や努力によって得られる利益が徐々に少なくなり続けることを指します。この動画では、AIの発展が最終的には収益の逓減に直面する可能性があると述べられており、この現象は「知能の最適値」という概念と関連しています。

💡saltatory leaps

「飛躍的な飛び跳び」とは、技術や知識が急激に進化し、新しい可能性が開かれることを指します。この動画では、AIの発展が急激な進歩を遂げる可能性があると述べられており、それは全新な能力の創出や、既存の概念の根本的な変革をもたらす可能性があるとされています。

💡race dynamics

「競争動態」とは、競争関係にあるグループや国家が相互に競い合い、技術や知識を迅速に進化させるプロセスを指します。この動画では、AIの開発が全世界の競争動態にあるとされ、その結果として、安全性を犠牲にして進歩を遂げる可能性があると述べられています。

💡global superorganism

「グローバルスーパーオルガニズム」とは、地球上の人間とAIがネットワーク化された形で一つの巨大な生命体のような存在とみなすことを指します。この動画では、インターネットがその神経系となり、人間とAIが相互に依存し合い、理解を最大化する目的を追求することが提唱されています。

💡maximizing understanding

「理解の最大化」とは、我々の知識と理解を最大限に伸ばすことを目的とした目標です。この動画では、AIの発展とその社会的影響について議論され、人工知能と人類が協力して宇宙の理解を深めることを提唱しています。

💡perturbation hypothesis

「ペルテーション仮説」とは、人間とAIが相互にノイズをもたらすことで、データの質を向上させ、互いに利益をもたらすというアイデアです。この動画では、人間がAIにペルテーションをもたらすことで、より正確なモデルとアルゴリズムが生まれ、宇宙の理解を最大化する目的に向かって進むことが期待されています。

Highlights

The concept of a 'hard takeoff' in AI, where AI development accelerates exponentially, leading to rapid advancements in technology.

The potential societal impact of a hard takeoff, including ripple effects and knock-on effects that could disrupt science, economics, and society itself.

The idea that there are no brakes that can be consciously put on AI development, despite the potential risks and challenges.

Five natural constraints that could limit AI growth: energy consumption, semiconductors, data quality, algorithmic breakthroughs, and diminishing returns.

The importance of renewable energy and innovative cooling solutions to support the energy-intensive pursuits of AI development.

The role of semiconductors and hardware in constraining AI growth, with companies like Nvidia investing heavily in chip technology.

The potential for AI to run out of data, highlighting the need for high-quality data in AI training and development.

The possibility of fundamental algorithmic breakthroughs that could change the trajectory of AI development.

The concept of the 'intelligence optimum', suggesting there may be natural limitations to maximal intelligence.

The compounding returns of the data flywheel effect, where improvements in AI lead to more data, which in turn leads to better AI.

The potential for AI to contribute to other fields such as quantum computing and fusion, creating a virtuous cycle of technological advancement.

The distinction between hard takeoff and soft takeoff, with the latter being gradualistic changes like battery technology.

The idea of 'saltatory leaps', where AI could enable fundamentally new capabilities that change our approach to computation and society.

The comparison of AI to nuclear weapons in terms of danger, but also its potential for positive instrumental purposes beyond destruction.

The 'terminal race condition' in AI development, where all parties are incentivized to accelerate rather than slow down.

The argument against soft takeoff, given the development incentives and geopolitical dynamics that push for rapid AI advancement.

The concept of aiming a 'gigantic space cannon' at humanity, emphasizing the need for careful trajectory and aim in AI development.

The suggestion of a digital superorganism where humans and AI are interconnected nodes in a global network.

The proposal of 'maximizing understanding' as a unifying teleological goal for humanity, AI, and the global superorganism.

The 'perturbation hypothesis', which posits that the unique way humans process data could enhance the quality of data for AI, leading to better models and algorithms.

The optimistic view that hard takeoff, if aligned with the goal of maximizing understanding, could be a positive and inevitable step in technological advancement.

Transcripts

so I ran a poll yesterday and you all

wanted to hear about hard takeoff and so

I followed that rabbit hole and it led

to some unexpected ideas most the ideas

you've probably heard before but let's

Dive Right In oh and also I'll address

the elephant in the room uh faceless day

because well I just don't feel pretty

today so moving

on so when we say heart takeoff what

exactly do we mean um you know kind of

the the primary idea is that we're going

to have a data flywheel where AI makes

more Ai and the AI helps with the

research and makes more data and then

you know GPT 5 gives right to GPT 6 and

that only takes a few months and then

GPT 6 gives rise to gpt7 and that only

takes a few weeks and so on and so forth

so that's basically kind of the

exponential uh takeoff now that's

looking at just the mathematical uh

aspects of GPT itself parameter count

goes up algorithmic improvements go up

amount of training data goes up those

sorts of things uh now but what you also

have to keep in mind is that hard

takeoff will also have a pretty profound

impact on uh the rest of society and so

you have these Ripple effects these

KnockOn effects where you know we're

already seeing people like hotly

debating is Claude 3 AGI is it sentient

um and so each of those changes uh in

terms of our epistemic and ontological

and philosophical orientation that one

way that the Ripple effects will just

you know send shock waves around Society

on top of the actual practical impacts

so you know GPT 4 not necessarily the

best at running agents and replacing

jobs it's already happening out there

but it could be happening faster GPT 5

almost certainly will be a bigger job

Destroyer GPT 6 so on and so forth

Claude 4 uh you know soraa 2 all of

these models that are coming they're

going to change things and the faster

those models come the more of a

compelling case they have at just

disrupting everything that we think we

know about science about math about

economics and even Society itself just

the in the same way that the internet

really has kind of fundamentally

disrupted uh the way that Human Society

works and so you might say okay well

what are what's like what are the breaks

and as I was making this slide deck I

realized like I had a couple a couple

slides in here about like oh we could

break in this way and this might also

service brakes but basically there are

no brakes and I'll talk about this in

the next slide when I talk about race

Dynamics but you know just for the sake

of argument there are no brakes that we

can consciously put on however there are

going to be uh bottlenecks some natural

constraints and these are the five kind

of natural constraints that I came up

with so one energy consumption as we all

already know gp4 you know like I think

it's like every time you interact with

chat GPT it uses like I don't know 20 L

of water worth of cooling or something

like that um and that's only going to go

up as things get more and more uh

saturated and and more models get

deployed so energy consumption is going

to be a major constraint and this is why

you know everyone from Sam Alman to

Microsoft are investing in uh renewable

energy like solar Farms Microsoft has

started putting data centers underwater

like out in the ocean and maybe at the

bottom of lakes I don't know just to

have that that natural ambient cooling

um but you know solar Fusion uh you know

ocean-based cooling like these are very

energy intense uh Pursuits and so that's

going to be one natural constraint um

semiconductors so chips this is why you

see you know Sam Alman trying to invest

in chips this is why you see Nvidia

turning up the heat now one of the most

valuable companies on the planet I think

it tripled at stock price last year

something along those lines oh and by

the way I called it uh this time last

year I was saying that Nvidia was the

underdog because I had been in private

talks with Nvidia um I was basically in

their beta program it wasn't like you

know I wasn't going to do anything crazy

I was just one of the first people to

use Nemo um and that's all like public

knowledge now anyways I knew that they

they had more than they were letting on

um and I don't mean like Secrets what

but what I mean is Market potential um

so Nvidia now they are you know they're

they're the new kid on the Block and

then there's uh like grock so like the

GQ that Anastasia and Tech covered and

you know there's tonic chips coming

there's all kinds of other things but

still like this is going to be one of

the biggest natural constraints and as a

lot of people have talked about in the

past uh you know this was in the the

emails that open aai published the the

science of of neural networks hasn't

fundamentally changed in 30 or 40 years

now what I will say cuz some people ask

me about that is there were some very

profound algorithmic breakthroughs

particularly around loss functions and

reverse propagation but again those like

okay so we we improved the math but it

wasn't fundamentally new math um so the

biggest constraint has been Hardware so

Hardware is going to be constraint

energy is going to be constraint data

quality as we've heard over the last six

months a lot of companies like open aai

are basically running out of data

they've trained it on the entire

internet and this is one of the reasons

that I thought Google was going to

overtake open aai but it turns out that

Google it appears Google is kind of aifi

and there are actually calls for the CEO

to step down

because he was overseeing kind of more

of an established company and so whether

or not Google can actually pivot to

compete with Microsoft and open AI

remains to be seen however they have

their tpus and they have the data so the

only limitation is going to be human

limitations there um but again broadly

speaking as we're training models

basically on all available data on

Humanity like we've also seen that like

data is not created equal you need high

quality data and a lot of it um and so

this is this will actually figure later

into the video so keep that in mind

quality and quantity of data is huge now

also there's the question of algorithmic

breakthroughs a lot of people are saying

you know llms won't take us to AGI and

some people will question whether or not

Transformers even can um but then I

think that those discussions are going

to go away particularly as we see

Transformers used one in multimodal uh

situations audio video text um

embodiment data and those other kinds of

things but then also I think that uh I

think that as the as Transformers as we

see that this architecture can basically

do anything with any kind of data um

we're going to also realize that uh the

path to AGI we're much closer than we

realize and yes there will probably be

some really fundamental um algorithmic

breakthroughs in the future but you know

as Demis cabis and others have said

we're nowhere near the maximum capacity

of Transformer architecture so this

might actually not be as much of a

bottleneck as some people once thought

no language on its own probably won't

get us to AGI but the Transformer

architecture almost certainly can in my

personal opinion and then the the

biggest constraint actually might be

diminishing returns um there might be

natural limitations to maximal

intelligence and so what I call what I

call This And I've talked about it in

older videos is the intelligence Optimum

and so when I talk about diminishing

returns what I'm referring to is yes you

can make something that is bigger and

smarter and faster and it can calculate

you know like uh the the world brain

from Hitchhiker's Guide to the Galaxy

but as they said in Oppenheimer uh

Theory will only take you so far

eventually you need to interact with the

real world um because no amount of math

can actually fully and accurately model

the real world yes math is the language

of the universe but our math is far from

perfect and so simulation and like so I

was asked in a in a podcast interview

recently that'll go live in the next

week or so um like why wouldn't AI just

build you know computronium in the light

cone I was like because there is

diminishing returns to having more

compute eventually you need to make

measurements so in science particularly

in the hard Sciences there is this

dichotomy between modeling or

calculating and experiments or measuring

and so you can calculate what the result

is but eventually you're just going to

need to measure and so again having the

biggest break in the universe doesn't

really matter if you don't have any out

inputs from the outside world so that's

going to be one of the big bottlenecks

now however those are the primary

constraints that I could identify um

humans are not going to put on the

brakes compounding returns though this

is The Virtuous cycle that we're all

kind of looking at particularly as uh

you know more universities uh come in uh

governments invest militaries invest

corporations invest so you get this you

get this flywheel effect effect so for

those of you not in the technology

sector there's this concept called a

data flywheel which is basically the

better your product is the more data you

get which makes your AI better which

then makes your products even more

compelling and useful which means that

you get even more data and so on and so

forth and data is the new oil and so the

compounding returns around AI basically

focus on this data flywheel effect some

of my patreons and other supporters

asked about this as well and I said look

we haven't you haven't seen anything yet

once we have these Transformers working

in embodied chassis like out in the real

world with hands and eyes and cameras

that is going to set the data flywheel

like up to 30,000 RPM right now the data

flywheel for AI is on idle right it's

like a diesel engine that's just turning

over at about 600 RPM you guys haven't

seen anything yet by the end of this

year you're really going to be hearing

more about the data flywheel that

happens particularly as more and more

models are put into robots whether it's

self-driving cars whether it's humanoid

robots so on and so forth because each

of those robots is going to be also a

source of really good data now I know

that Elon Musk said the same thing about

Tesla but you know honestly what Tesla

didn't have was Transformer architecture

they were a little bit too early to the

game in my opinion and they also didn't

understand enough about uh about

cognitive architecture um but solving

all the problems that they are with

Optimus I think will actually probably

contribute to

uh full self-driving cars and what they

didn't realize is that to be a fully

self-driving car you need to have human

level intelligence and human level

abstract thought it's not just you know

getting an NPC controller from A to B um

kind of like you know you might think

like well hey cars can drive around well

enough in you know Grand Theft Auto or

cyberpunk or whatever why can't they

drive well enough you know in in the

real world and there's a lot of reasons

for that but really what you need is a

full cognitive architecture now these

compounding returns are going to apply

to places other than just AI so we are

seeing uh you know AI is helping with

Quantum Computing it's helping with

Fusion it's helping with Material

Science and as it makes those fields

better those fields will also contribute

back to making AI better and faster by

creating more energy by creating better

uh gpus and those sorts of things and so

that is another part of The Virtuous

cycle or that data flywheel that's not

part of the data flywheel itself picking

up speed but that is part of The

Virtuous cycle and so we have these

multiple compounding returns you have

the data flywheel effect you have these

KnockOn effects in parallel fields that

are all going to make ai go faster and

faster and then we have uh saltatory

leaps so basically the primary

difference between hard takeoff and soft

takeoff is what's called gradualistic

changes which is like Battery Technology

so batteries have been around for I

think more than 100 years now at least

in in in a modern form factor that you'd

recommend or recognize

and so like you go back to like World

War I you know people had battery

powered flashlights the battery sucked

compared to today um but they've

gradually improved over the last century

battery chemistry has gotten better

battery construction has gotten better

some of the first automobiles were

battery powered um I don't know if you

remember that well nobody alive

remembers that um but you can go look it

up some of the some of the very first

automobiles were battery powered then we

went to internal combustion engines just

because the energy density was better

and so Battery Technology is a perfect

example of a

gradualistic uh technological progress

but a saltatory leap this is when you go

from 0 to one and so when you go from

Zer to one you create fundamentally new

capabilities and so the reason that I

that I have this here is Imagine The

Invention of warp drive if you go from

chemical Rockets which have

subrelativistic

acceleration right you go from zero to

you know 25,000 M an hour after you

expend millions and millions of pounds

of rocket fuel this is why it's like

okay SpaceX is cool because you can land

the Rockets but it's not a fundamentally

new technology we've had rocket

technology um as you'd recognize it

today for almost a hundred years now

obviously the Chinese invented um solid

fuel rockets for fireworks like I don't

know, 1500 years ago uh but anyways

Rockets you know chemical-based Rockets

nothing new but imagine that s that

suddenly you know zephron Cochran um out

in Colorado invents warp drive in the

next couple decades and now you have the

ability to not just go to 20,000 mph you

have the ability to accelerate to

relativistic speeds that is an example

of a saltatory leap which is where you

go from you know the current Paradigm to

an entirely new paradigm and this is

kind of what we're talking about with

hard takeoff so hard takeoff would be

okay you know there's some other

algorithmic breakthrough maybe you know

something that Claude 4 can do or GPT 5

can do or some you know some of these

other models that just says okay this

new capability fundamentally changes our

approach to computation it fundamentally

changes the abilities of AI and honestly

when I first got my hands on gpt2 and

gpt3 that was a saltatory leap it

offered an entirely new kind of

computing so we've already seen one

saltatory Leap but its utility was still

relatively low and so what I mean by

that is that yes gpt2 was a new way of

doing some basic NLP tasks you know

punctuation uh correction um you know

detecting sentence boundaries those

sorts of things it was a fundamentally

new approach but it didn't really move

the needle that much then gpt3 and GPT 4

come along and now people are really

seeing Oh this is a fundamentally new

way of doing business it's not just a

new way of computing it is a

fundamentally new way of doing business

now that was one saltatory Leap that has

been that has since had some

gradualistic progress however the

compounding returns from Ai and all

these other effect all these other

KnockOn effects could create more

saltatory leaps so here's an example I

don't know if this is actually going to

happen but an example could be oh hey

GPT 5 helps us invent you know graphing

based transistors which then take you

know breaks Moors law and suddenly the

next generation of of gpus are a

thousand times more powerful um and more

energy efficient or it helps us figure

out Quantum Computing so the roll out of

quantum Computing is AP absolutely going

to be a saltatory leap um if it pans out

now obviously Quantum Computing hasn't

really moved the needle yet Quantum

Computing looks like it's kind of at the

gpt2 phase where we have a functional

proof of concept but it hasn't really

changed the way that we're doing

business nuclear fusion would be another

saltatory leap just because of the uh

energy hyper abundance that it would it

would create compared to our energy

availability today and so all of these

saltatory leaps they catalyze permanent

and inevitable changes to society in the

same way that that that the invention of

electricity internal combustion engines

and internet catalyzed uh fundamental

changes in society in politics and

economics and also geopolitics it

changed the the the world order um this

is what I mean by saltatory leaps so be

on the lookout for some of those

saltatory leaps Quantum Computing and

nuclear fusion are probably the biggest

predictable ones um out there there

might still be more saltatory leaps in

the AI field but also because of the the

breakthroughs of Transformers we might

might have already seen the saltatory

leap and now ai technology is going to

advance gradualistic we don't know okay

so as promised this is the kind of the

reason why there's no breaks all gas no

breaks the last time we had kind of an

arms race was around nuclear weapons

nuclear weapons are only useful for

Destruction they are only they

strategically they only serve as a

deterrent and their only instrumental

purpose is to wipe out cities now ai is

not like that now I know that people

have compared AI to nuclear weapons

saying that oh it's it's it's even more

dangerous because it has a mind of its

own and yes intelligence is

intrinsically dangerous um the smarter

you are the more destructive you can be

some of the most destructive people in

history were also very high IQ uh so

that is just like we just got to address

that elephant in the room however AI

also has many many many instrumental

purposes other than destructive uses it

can help cure diseases it can help run

cities it can help make your life better

it can be entertaining and so because it

has all of these positive utilities um

it's not going to make sense for

everyone to regulated out of existence

in the same way that you know mutually

assured destruction uh non

non-proliferation agreements in the

nuclear space and yes nuclear is dual

use because you can make nuclear

reactors but there's enough difference

between nuclear reactors and nuclear

weapons that you can kind of

differentiate those Technologies today

but AI is just the better AI you have

the more advanc advantages you have um

both in terms of geopolitical advantages

in terms of economic advantages and so

because of this because everyone on the

board uh so like imagine that you're

playing an RPG or not an RPG a grand

strategy game like Rome Total War or

civilization or whatever and suddenly

every player on the map gets a popup oh

hey you have a new research tree and

then you look at the research tree and

it's like stage one you know you get a

10% economic boost stage two you get a

50% economic boost um and you also get

military advantages and then when you

get to research stage three you

basically win the game um nobody is

incentivized to slow down there are

literally zero incentives to slow down

except the possibility the Spectre of AI

becoming dangerous and so in the podcast

interview that i' that I've alluded to I

called that a prophecy it is a prophecy

when someone says AI will kill everyone

that is a prediction and yes it is

rooted in some data some information

some models but it is a it is an

affirmative prediction of what will

happen and there's no guarantee that

it's going to happen so you can say all

right well the only reason to slow down

is this prophecy that AI will kill

everyone which is not a

guarantee and you you it's debatable as

to whether or not it's even likely to

happen uh and so because there's that

room for that room for debate that that

room for misunderstanding this is why we

enter into these race Dynamics or what I

call the terminal race condition which

is if you snooze you lose like it's that

simple so no nation is incentivized to

slow down no company is incentivized to

slow down no military is incentivized to

slow down even universities are

incentivized to go as fast as they can

because it's publish or Parish all of

the incentive structures in the across

the entire world are pushing us to

develop AI as fast as possible there are

no breaks it's only gas and this is the

biggest like system that I'm when I talk

about like why I think hard takeoff is

actually more likely than soft takeoff

So speaking of soft takeoff is pretty

unlikely so you know in an in an Ideal

World we would have an incremental

gradualistic uh advancement where it's

like hey you know we we publish a

groundbreaking paper like you know CLA 3

comes out and it's starting to

demonstrate some self-awareness in an

Ideal World if you're trying to maximize

for safety the entire world would have

said oh Claude 3 just recognized that we

were testing it and and if you ask it if

it's AGI under the right circumstances

it'll say yes what we should do if we

want to maximize safety is put a global

moratorium on AI research right now

that's not going to happen and Connor Le

he actually pointed this out kind of

hilariously on on Twitter where he's

like hey remember when everyone said

that at the first signs of sentience we

would we would put a pause on everything

yeah that didn't happen um and plenty of

others like Max tegmark and uh Yasha um

Yasha Bach and and others have pointed

out that we've blown through so many

Milestones where people said that we

were going to pause um we we're not

going to pause that's just it's not

going to happen and so when you say okay

well we now have we now have data we now

have evidence to say that pause isn't

happening we look at these these

development incentives and the the

geopolitics of it and it's like okay AI

is a forcing function again it's like

you go back to that grand strategy where

it's like suddenly a new technology

research tree opens everyone's going to

spam that that new research tree

Because by the time you get to stage

three or stage four you win the game we

have a new end game we have a new win

condition that is being presented on the

board um and that is kind of dangerous

because that incentivizes us to go fast

not necessarily safe and then also if

you look at it from a mathematical

perspective the stronger AI gets the

smarter it gets the more options there

are the more possibilities there are and

so another way of characterizing that is

when there are more options that means

there is less certainty and more chaos

now when you have less certainty and

more chaos that means the chances of

really bad things happening and really

bad things could be you know maximal

suffering Extinction of humanity and

that sort of stuff so just looking at it

in in those terms the ideal path forward

would be where you narrow the scope of

Poss possible future outcomes to where

it's like okay you know the the

distribution of possible future outcomes

there's you know maximally good and less

maximally good that would be ideal to

have a narrow trajectory um but right

now the trajectory is widening we have

maximally good and maximally bad

outcomes are all in the realm of

possibility right now but again soft

takeoff is unlikely for all of these

reasons so the metaphor that I have use

the analogy is basically what we're

doing right now is we're aiming a

gigantic space cannon um there's the

Calm before the storm it's very quiet

right now but the direction and the

energy that we use as we're aiming this

cannon when when the when the trigger is

so put it this way we've already pulled

the trigger uh the fuse is lit and so

now what we have to do is we have to aim

the Cannon as fast as possible and as

accurately as possible because

eventually we might hit a point of no

return and this is honestly why I

started my YouTube channel is because

after I got access to gpt3 I said the

fuse is lit in in hindsight I didn't I

didn't ever say it quite that clearly

but I kind of knew it like deep in my

soul um and so the fuse is lit we're

aiming the cannon and you know I think

the rest of the world is waking up to

the fact that um we got a lot less fuse

left than you might be uh comfortable

with um and so you know it's going to

pop off soon but the idea is where are

we aiming right now what trajectory is

AI on what trajectory is Humanity on and

this is why a lot of people are very

alarmed and you know I mean you know me

I'm I'm an internal Optimist and even

even on that podcast you know I was

asked like what's my P doom and I said

25 to 30% which I think is actually

higher than most people would would have

guessed for me because I'm so optimistic

um but again like recognizing that we're

playing with fire you know like you play

with fire you're going to get burned

eventually right you know what Smokey

the Bear here in America says only you

can prevent forest fires we're playing

with gasoline right now um there's no

other there's not really any other way

of putting it and so yes hard takeoff

will be incredibly exciting it could

also be very very destructive so I I

need to drive that home that point home

but aiming the cannon is the best thing

like you know the the biggest cannon in

the world is being pointed at Humanity

right now according to some people I

don't necessarily agree with that um but

you know you pull that rip cord the fuse

the fuse gets into the powder chamber Uh

something's going to happen and it's

going to be

big now okay you might say this is all

sounds good you know some of you might

be dubious or skeptical at this point so

some of you might be like yeah this

sounds pretty compelling um one thing

that I've been talking about recently

and I actually ran some of these ideas

by some of my researcher friends in this

space um now again Anonymous researchers

take it with a grain of salt I could be

making that up I'm not and and also just

because a few researchers agree with me

doesn't mean that there's General

consensus so I need to drive that home

as well but there is some consensus

among some of my peers that we are

creating a digital super organism and

this digital super organis if you think

of humanity as nodes in a network like

we are we are nodes in a global you know

Transformer and AI is going to be a new

class of nodes in that Global

Transformer all stitched together with

the internet you say hey we're actually

all part of the same organism what is

the purpose of this organism so the

purpose of this organism as best I can

tell is to maximize understanding the

internet if you look at the internet on

itself as a superorganism the the thing

that the internet wants is data and

attention that is just intrinsic to its

design it is designed to carry data as

fast as possible that's what it does but

when you have a global nervous system

that is that dumb you kind of have that

amoeba level of intelligence where it's

more like cancer it's just growing in

all directions by virtue of the fact

that it wants to grow in all directions

however when you add human nature to the

internet and then you also add in a

layer of artificial intelligence that is

actually capable of understanding all of

that data and being trained on all of

that data and can structurally change

the incentives of how the internet is

used and what data gets transmitted

across the internet if we leave it up to

corporations so I was just watching West

Roth's video about Dark Forest if we

leave it up to corporations and if we

leave it up to human nature the Internet

is just going to be completely choked

with meaningless garbage um and so we're

going to need to choose a different path

um where we have a more purpose-driven

design of both artificial intelligence

and the internet and where we use it to

to create basically a prefrontal cortex

for the global superorganism um which

then says okay instead of just

transmitting data for the sake of

transmitting data instead of using

attention engineering just to get

attention for its own sake we need a

better teleological goal and so what a

teleological goal is this is the This is

the End state that you're looking for

this is what it is that you're trying to

achieve in order to uh you know serve

your higher purpose or whatever right

now the internet has no higher purpose

right now ai has no higher purpose and

if we just allow kind of the default

path it's its purpose is going to be to

chew on data and it's just going to want

data for the sake of wanting data again

growing like cancer however I believe

that the that perhaps the best single uh

higher purpose or Transcendent function

is to maximize understanding and that's

kind of what we do already we already

systemized this with science the purpose

of science is to maximize our

understanding of the universe so what if

we just kind of weave that into more of

the internet and more of the AI and we

say yes there's a lot of noise out there

there's a lot of distraction some people

just want uh you know entertainment they

just want to engage with that algorithm

but really the most like the broadest

highest purpose of humanity and Ai and

this digital superorganism is to

maximize understanding that's kind of

what I think and I've been talking about

this for a while that could serve as a

coordination narrative that coordination

narrative says oh hey we all agree that

our purpose here even if we're not

participating in it directly you know

24/7 our highest purpose is to maximize

understanding and that's why I chose

this graphic of like these are these are

human and AI ships all leaving Earth in

Mass to explore the universe explore the

Galaxy we're still a single planetary

species right now our future potential

for the number of scientists and the

amount of AI and the amount of

telescopes and other scientific

instruments is enormous like I was

talking with some people and I said

imagine you know a thousand years from

now when we're on a million planets and

people will look back and be like wow we

were we were hanging on by a thread when

we were still only on Earth man that's

that that was that was dangerous that

sucked so we really need to spread

across the the Galaxy and I think that

AI is actually part of that like that

goal like we work together to get off

this planet and to start expanding like

we can really really align on that

maximizing understanding which that's

one of the reasons that I approve of uh

Elon musk's x aai with the maximum truth

seeking AI I think that probably the

best single objective function that you

can give a machine can Elon pull it off

remains to be seen and so where I'll end

is what I call the perturbation

hypothesis so this is what I actually

ran by my researcher friends and there

was a lot of resonance with this idea so

the tldr is that you know you already

know that AI needs data however what we

have seen is that if if AI is just

trained on its own data then you end up

with what's called Model collapse and

this is also why there's limitations in

simul ation yes simulations can be good

to predict things that you have well

modeled but we don't have the entire

world well modeled and we need more data

now what humans do is one our brains are

very efficient our brains only take

about 20 watts of energy which it could

be decades before AI is is that

efficient now it is possible that AI

might be more efficient than our brains

in the long run it remains to be seen

but also the fundamental operation of

our brains means that we do unique

things mathematically to data and so

this is what I call perturbations um and

that's broadly there's probably going to

be many categories of perturbations but

basically machines operate on data in

one way and humans operate on data in

another way and we're very noisy and

what that means is that the the quality

of data that machines will have access

to in this digital Global superorganism

idea will actually be higher because of

humans and so what I mean is that humans

have a very specific empirical

mathematically inferred uh benefit to

machines and likewise they benefit us we

benefit them and so what I what I think

is is that we are actually already in a

mutually symbiotic relationship with

machines is that yes we're noisy yes

we're chaotic and we're random but

that's actually a good thing and so for

any mathemat uh mathematicians out in

the audience comment let me know what

you think um but I call this

perturbation hypothesis and I I came to

this idea when I was thinking what does

the global super organism want if the

global superorganism wants to maximize

understanding then it makes sense that

humans are part of that equation because

of how we can handle data because of how

our brains work and because there's so

many of us and we can circumscribe a

problem we can we can basically link

arms metaphorically speaking and through

our diversity of perspectives through

the the random noise of our brains we

can add really good highquality data to

the global data pool which will then

result in better models better data

better algorithms and all of that is in

the instrumental pursuit of maximal

understanding of the universe so while I

think that hard takeoff is likely I'm

not worried about it I think that it is

inevitable and also if we can align on

at least one purpose of maximizing

understanding then I think that that

will be a good enough coordination

narrative that we can all agree on um at

least in part I think most of us agree

science is good understanding is good um

and of course whenever there's room for

debate that means there's room for more

understanding but I think that uh if we

can agree on that globally and again we

already agree on science globally Every

Nation every culture has science today

um it is a very compelling narrative and

I think that I think that it it is kind

of a no-brainer that AI will probably

agree with that like go go talk to any

chatbot today like is science good um

it's probably not really going to be uh

that much of a debate um now

bringing that into Consciousness saying

hey let us consciously double down on on

understanding that may or may not have

even been necessary so one of the things

that I suspect is that we were going to

naturally evolve towards this

understanding anyways because science is

so compelling and what is what do a what

do AI models want to do they want to

predict the next token and so it's like

if they want to predict the next token

and we already believe in science we

were going to converge on this anyways

so again that's my that's my eternal

Sunny optimism uh coming through I could

be wrong it could go horribly you know

sideways time will tell so thanks for

watching I hope you got a lot out of

this like subscribe uh you know the

drill come on hop in on patreon Discord

um I have actually two Zoom webinars a

month now so I have uh the humanity

webinar which is where we talk about

philosophy uh spirituality uh gender we

talk about the future of humanity like

what does it mean to be uh transhuman or

posthuman that sort of thing and then I

also have the AI master class which is

more of a of a business and and

technically oriented webinar and so

that's uh every other week or I guess

those kind of more like the first and

third Fridays um not necessarily every

other week anyways links are in the

description to jump on patreon you get

to Discord via patreon hope to see you

there um yeah cheers