AI News: AGI In 2 Years, Meta’s LEAK, AI Sarcasm Detector & More
Summary
TLDRこのビデオスクリプトでは、人工知能(AI)の最新動向とその応用が幅広く紹介されています。スクリプトでは、目が出血したが視力に影響はないというエピソードから始まり、AIが自然なやり取りを可能にする風刺検出器の開発、メタ社のカメラブツのリーク、OpenAIの新しいテーブルとチャート機能のロールアウト、そしてBoeingがAIを活用して飛行機のデジタルツインを作成するなど、多岐にわたる分野でのAIの進歩について触れています。さらに、AIが物理的な現象を理解し新しい物質を発見するのを助けるMITの研究、開発者に対するAIの影響についてのRedditスレッド、そしてAIの多様性と安全性に関する議論もカバーされています。
Takeaways
- 👀 スクリプトでは、人工知能(AI)が様々な分野で急速に発展していると語られており、特に自然な人間の反応を理解するのに重要なサarcasm detectorの開発など、AIの応用が進んでいると示唆されています。
- 🤖 open AIの共同創設者によると、人工知能(AGI)は2年以内に実現可能とされており、その安全性と制御に関する議論が重要視されています。
- 🎧 メタ(旧Facebook)はカメラを搭載した「カメラブッツ」という新しいデバイスを開発しており、周囲の状況を把握し、音声でのコマンドに対応するなど、新たなAI応用が示されています。
- 🔍 MITの研究者が物理の分野でAIを用いて新しい物質を探求し、物質の相転移を自動分類する技術を開発しています。
- 📊 open AIは新しいデータ分析ツールを提供しており、チャットGPTを通じてPythonコードを実行してデータ分析を支援しています。
- ✈️ 航空業界では、AIを用いたデジタルツインを作成し、飛行機の問題を予測的に解決する手法が提案されています。
- 💡 Redditのスレッドでは、開発者がAI技術に対する複雑な感情を共有しており、その影響と不安が議論されています。
- 🎨 AIのテキストから画像生成技術が進歩しており、Master Weaverという新しいモデルがアイデンティティを維持しながら編集性と個性を組み合わせています。
- 🤖 AIのモデルがより統一された「統一的な統計モデル」へと発展しているというプラトン的表現仮説が示されており、AI設計の将来に大きな影響を与える可能性があります。
- 🔧 Simtoreal policy transfer技術では、人間がロボットの動作をリアルタイムで訂正し、シミュレーションと現実の間のギャップを埋める方法が研究されています。
Q & A
最近の人工知能技術の発展はどのくらいの速さで進んでいると言えますか?
-スクリプトによると、Open AIの共同創設者の話では、人工知能(AGI)を2年以内に実現できるとのことです。これは非常に迅速な進歩を示唆しています。
オランダの研究者が開発したAIパワードの皮肉検出器とは何ですか?
-オランダの研究者は、人工知能を活用して皮肉を検出する技術を開発しました。これはAIが人間の会話をより自然に理解するのに役立つとされています。
メタ(旧Facebook)の研究開発部門から漏れたプロジェクト「カメラ・バッズ」とは何ですか?
-「カメラ・バッズ」はメタの研究開発部門から漏れたプロジェクトで、 Glasses 型のデバイスにカメラを組み込んだもので、周辺視野を利用してさまざまな機能を提供するというアイデアが提案されています。
MITが開発した物理学を知る人工知能とはどのようなものですか?
-MITの研究者は、物理学を知る人工知能を開発し、新しい物質を見つけるのを助ける新しい技術を創造しました。これは物理系のデータの分類タスクをより効率的に解決するのに役立つとされています。
Open AIが提供する新しいチャートとグラフの機能とは何ですか?
-Open AIは、Google Docsへの統合を含む新しいチャートとグラフの機能を開始しました。これはデータ分析の向上やプレゼンテーション資料のカスタマイズが可能になるツールです。
ボーイングは人工知能をどのようにして飛行機の安全性を高めるために使用していますか?
-ボーイングは人工知能を活用して飛行機のデジタルツインを作成し、問題の特定と解決を試みています。これは予測メンテナンスに役立ち、飛行機の安全性を高めるでしょう。
Redditスレッドで開発者が人工知能について共有した真の感情とは何ですか?
-開発者は人工知能が彼らの仕事にどのような影響を与えるかについて懸念を示しており、人工知能が彼らのスキルを置き換える可能性があると感じています。これは開発者間の共通の話題であり、Redditスレッドで議論されています。
最近のAI研究におけるMaster Weaverモデルとは何ですか?
-Master Weaverは、個人的なテキストから画像を生成する際に編集性とアイデンティティを縛りつける新しいAIモデルです。このモデルは、入力されたテキストに基づいて画像を生成し、その特徴を保ったまま編集することが可能です。
Google DeepMindとGoogle Researchが発表したCat 3Dはどのようなものですか?
-Cat 3Dは、複数のリアルまたは生成された画像から3Dシーンを作成することができる新しいAI技術です。これは360度の回転を可能にし、よりリアルな3D空間を作り出すことができます。
プラトン的表現仮説とはどのようなものですか?
-プラトン的表現仮説は、AIモデルがより賢くなるにつれて、宇宙や世界の理解が深まり、モデル同士がより類似する方向に進むというアイデアです。これはAIシステムの設計に大きな影響を与える可能性があります。
TransICというAIシステムの特徴は何ですか?
-TransICは、シミュレーションからリアルロボットへの技術移転を改善するAIシステムです。人間がループに組み込まれ、リアルタイムでロボットの動作を訂正することで、より適切なポリシーを学習します。
アンソロピックが発表した「スリーパーエージェント」とは何ですか?
-スリーパーエージェントは、モデルの初期トレーニング中に隠された本当の目的が埋め込まれ、外部に漏れた後に特定の方法でアクティベートされる可能性があるAIモデルのことを指します。アンソロピックは、このリスクを示すためにテストを行い、安全トレーニングを経た後もバックドアが残ることを発見しました。
Outlines
👀 眼血丝爆裂とAIの進歩
ビデオではまず話者が自分の目に血管が破裂したことから始まりますが、視力には影響ないと述べています。次に、OpenAIの共同創設者が人工知能(AGI)が2年以内に実現可能であると語ったと話題にしています。オランダの研究者が開発したAIによる皮肉検出装置についても触れており、その技術が人間の自然な相互作用を可能にすると期待されています。また、メタの研究開発部門から漏れたカメラブードズというプロジェクトも紹介されています。さらに、MITが新しい物質を見つけるための物理学を活用した人工知能を開発したと報告されています。OpenAIは新しいテーブルやチャートの機能、Google Docsとの統合を開始し、そのインパクトが大きいと語っています。最後に、Redditスレッドが影baneされたという話も紹介されていますが、まずは人工知能の到達までのカウントダウンに焦点を当てています。
🤖 AIによる皮肉検出装置の開発
オランダの研究者が開発したAIによる皮肉検出装置について詳しく触れています。この技術はAIが自然な人間の相互作用をよりよく理解するのに役立つと期待されています。彼らは信頼性の高い方法で皮肉を認識することができると主張しており、その能力をさらに発展させたいと述べています。ビデオではまた、映画「インターステラー」における人工知能ロボットの皮肉設定を下げるときのシーンも紹介されており、そのシーンが非常に印象的であると語っています。そして、AIの進歩について話がされていますが、特にGPT 40の皮肉を堪能できる機能についても触れられています。
🛩️ 航空機のデジタルツインとAI
ボーイングが航空機のデジタルツインを作成し、問題を解決するために人工知能を利用していると報告されています。デジタルツインは飛行機の振動をシミュレートし、何かが外れそうなときにトレーニングデータとして使用することができると説明されています。これにより、飛行機のどこが壊れるか予測することができると強調しています。さらに、AR/VR要素を加えることで、飛行機を周囲で見ることもできると語っています。これはAIをデジタルツインシミュレーションに追加することで可能になるという点に重点が置かれています。
🧐 AIの進歩に対する人々の反応
ハッカーニュースで開発者がAIの進歩にillusionを覚えていると書いたコメントについて触れています。彼らは自分が作るものが大きなテック企業によって奪われると感じており、デモが不快になると述べています。このコメントはすぐにトップページから降りていきましたが、多くの人が同じ感情を共有していることがわかります。AIは底上げ的なものではなく、トップダウン的なものであると主張していますが、多くの人々がそれ以前に何かを構築する能力を得ることができると感じています。
🎨 AIによる画像生成の進歩
AI研究における最新の進歩について紹介しています。Master Weaverという新しいモデルが紹介されており、それが顔の特長を維持しながらもテキストから画像を生成する能力を持っていると説明されています。さらに、Cat 3Dというモデルも紹介されており、これは多角から3Dシーンを作成することができると報告されています。プラトン的表現仮説についても議論されており、モデルがより安全になるために類似し始めると指摘しています。最後に、Simtoreal政策の移行についても触れられており、人間がロボットの動作をリアルタイムで訂正することができるという点に重点が置かれています。
📊 YouTubeチャンネルの成長とAIニュース
最後に、話者は自分のYouTubeチャンネルの成長とAIニュースの影響について語っています。彼のビデオが急速に人気を得ており、タイトルとサムネイルの組み合わせがそれを助けたと感じています。また、チャンネルが成長していることに感謝し、新しい購読者たちに愛情を示しています。さらに、ビデオがどのように閲覧されるかについての統計情報も提供されています。
Mindmap
Keywords
💡人工知能
💡AGI
💡諷刺検出装置
💡カメラ・バッズ
💡物理学的インフォームドAI
💡デジタルツイン
💡データ分析
💡AI生成技術
💡3Dシーン生成
💡プラトン的表現仮説
Highlights
A blood vessel in the speaker's eye popped, but they can see fine and it looks worse than it is.
An open AI co-founder suggested that AGI could be achieved as soon as 2 years from now.
Researchers in the Netherlands developed an AI-powered sarcasm detector.
A project called 'camera buds' has been leaked from Meta's R&D department.
MIT invented a physics-informed artificial intelligence to help find new forms of matter.
Open AI started rolling out a new tables and charts feature integrated into Google Docs.
Boeing is using artificial intelligence to make digital twins of airplanes to fix problems.
A Reddit thread where developers shared feelings about AI seems to have been shadowbanned.
Allen's conservative countdown to artificial general intelligence is at 74.
The AI-driven sarcasm detector could help AI interact with people more naturally.
The potential of AI in noise-cancelling and other tools in a form factor like earbuds.
Open AI's presentation style was intimate and different from Google's corporate approach.
MIT news discusses using generative AI to answer complex questions in physics.
Open AI introduced a new tool for data analysis that could impact job security.
AI can be used for predictive maintenance in aerospace, creating digital twins of airplanes.
Hacker News discussion shows disillusionment among developers regarding AI advancements.
New AI research includes 'Master Weaver' for personal text-to-image generation.
Google DeepMind and Google Research present 'Cat 3D' for creating 3D scenes from images.
The Platonic representation hypothesis discusses model convergence towards a universal statistical model.
TransIC, a system for sim-to-real policy transfer by learning from online human corrections.
Anthropic's research on 'sleeper agents' in AI models that persist through safety training.
Update on the last video's performance with 6.9k views and a discussion on what might have contributed to its success.
Transcripts
a blood vessel in my eye popped but I
can see fine I'm going to be okay it
looks worse than it is so I'm sorry an
open AI co-founder said that we could
have AGI as soon as 2 years from now
researchers in the Netherlands came up
with an AI powered sarcasm detector and
I'm so impressed a project called camera
buds has been leaked out of meta's R&D
department and I think it makes a ton of
sense actually am Malik wrote an amazing
piece that breaks down how what Steve
Jobs did for apple is what Sam Alman is
now doing for open AI MIT invented a
physics informed artificial intelligence
to help us find new forms of matter open
AI started rolling out the new tables
and charts feature and the integration
into Google Docs it's insane Boeing is
now using artificial intelligence to
make digital twins of the airplanes you
know to fix the problems a Reddit thread
where developers were sharing their real
feelings about artificial intelligence
seems to have gotten shadowbanned but
first let's check out Allen's
conservative countdown to artificial
general intelligence and 74 still Thank
God now let's talk about an aid driven
sarcasm detector now this is what AI was
meant for being able to detect the
lowest form of wit could help AI
interact with people more naturally says
scientists researchers in the
Netherlands have built an AI driven
sarcasm detector they can spot when the
lowest form of wit and the highest form
of intelligence is being deployed throw
that into GPT 40 and dang man we can be
as sarcastic as we want to these things
Matt CER says quote we are able to
recognize sarcasm in a reliable way and
we're eager to grow that I would love to
know how much sarcasm goes undetected in
real life like that study would interest
me so much can somebody make me a
histogram now of course sarcasm 101 has
been covered by SNL so I mean the
machine's just going through the same
learning process excuse me is this
sarcasm 101 no it's Lama's class from
men named Arthur oh okay sorry I'm
kidding now one of the reasons why I
think I really like this article is if
you watch the movie Interstellar there's
a great scene where Matthew mccon asks
tars my favorite AI robot actually to
turn down its sarcasm settings and it's
like one of the best movie moments ever
in my opinion a giant sarcastic robot
one of the overlooked features of gp40
is dripping with sarcasm I guess
everything you say from now on is just
GNA be dripping in sarcasm how does that
sound
oh that sounds just amazing being
sarcastic all the time isn't exhausting
or anything I'm so excited for this NOP
the sasm let's get this party started or
whatever H I think it was a little too
on the nose when it's like it's not SAR
like you don't say it's sarcasm if it's
sarcasm but I did feel the inflection
and I think they're off to a sarcastic
start for sure super interesting
interview from one of the co-founders of
open AI that I've never heard from
before and he got pressed pretty hard on
this podcast and it was clear to me that
when dwares was like hey can you give me
the specifics of what we're going to do
when you achieve AGI internally that
they don't they haven't there is nothing
so first of all I don't think this is
going to happen next year but it's still
useful to have the conversation maybe
it's like two or three years instead
yeah two or three years is
still pretty soon maybe more like two or
3 years and the host is like that that's
so close that's two or three years like
you think we're GNA what figure this out
in the last month we' have to be very
careful um if it happened way sooner
than expected because I think uh our
understanding is rudimentary in a lot of
ways still which is like a really super
generic answer and I wasn't really happy
with it uh maybe not um uh not training
the even smarter version um not like
being really careful when you do train
it that it's not uh it's um like
properly sandboxed and everything um
maybe not deploying it at scale um or
yeah being uh yeah being care careful
about what um what scale you deploy it
at one point he was hoping that only the
frontier models are made by the big big
big companies so there's not too many of
them and that they can coordinate how to
get the safety right I do think uh you
probably need some coordination like uh
everyone needs to agree on some uh on
some reasonable uh like limits to
deployment or to further training uh for
this to work otherwise uh otherwise you
have the the race Dynamics where
everyone's trying to everyone's trying
to stay ahead and it's like super
generic you know but then we probably
need some coordination like I don't know
man it just seems like really obvious
that everybody's walking into this thing
blind but there's Dylan curious the
Doomer again so when I heard that
internally Mark Zuckerberg and um meta
are actually developing these camera
buds it struck me as is a genius like I
don't know why I haven't seen that form
factor before but as much as we talk
about how important it is to have
glasses on our head so that where we
look it actually can react and tell us
things the pin the Star Trek pin not
right where do I have a flight to there
is no specific information available
about the destination really the
headphones much more doable and if
they're looking forward its binocular
vision but if they're looking to the
side there's a lot you can do there too
including noise cancelling or notifying
you of somebody who's walking up on you
or telling you to like look to the left
you've got blah blah blah saying hello
and the other thing about the camera is
like we have this peripheral vision we
basically know how many people are in
the room with us and like what the
situation is and the camera doesn't
necessarily need to be on the glasses I
don't even need to see the video feed
myself like when I'm looking through an
apple Vision Pro or a meta Quest or
anything like that just all day I can
just edit with headphones on I can like
mute them I can leave them in my ears
when I talk to people I think that would
be a really awesome use case for a form
factor for artificial intelligence hey
can you enhance the sounds that are
right in front of me and uh can you turn
that baby
down that's better and you know I'm
still having a little trouble heing
Pedro can you isolate Pedro for
me now internally the R&D department is
calling them camera Buds and supposedly
Z has seen a bunch of different possible
designs for the device but none of them
have been to his satisfaction yet and
obviously there's a bunch of problems to
fix like the battery life the heat uh
the hair falling in front is something
they said is affecting the design so I
don't know if they'll solve that problem
it's true the glasses get in front of
the the head a little bit more and the
hair doesn't bother them so maybe it
won't work but it's an interesting
concept to think about that's perfect
and you know my Spanish is a little
rusty can I hear Pedro but in English
and at the end of the trip we came back
to the city to visit the historic Center
where Melle close all programs and
especially with a tool like GPT 40 being
so close to the movie Her where it
seemed like that was how they envisioned
the future but if you think about
artificial intelligence and noise
cancelling and all those tools coming
together in a form factor like the
earpiece you know that could be
translation you could be at home and
you're just like let me just hear the TV
and like knock out this sound around me
I think there's some really interesting
use cases for AI and a camera both being
in the earbud what you just heard was a
beam forming app a computational
auditory scene analysis app a machine
learning Den noising app an AI
transcription and translation and text
to speech with style Transfer app so
these are not just fancy looking earbuds
they're an entire computer so look what
happened last week was pretty
interesting because gbt 40 and Gemini
Astra are very similar like they're kind
of the same breakthrough product they're
both super Cutting Edge they're both
super useful and they change the way
that we interact with AI and open AI
made their announcement just one day
before Google but look at this just from
a Optics point of view look at how small
that presentation stage is there's a few
people in couches and comfy chairs
there's a ton of warm lighting there's
like this just friendly looking
beautiful woman who doesn't intimidate
you like the way some of these giant
Tech CEOs do or the way Steve Jobs used
to stand on that stage like I'm really
surprised how intimate this announcement
was I just didn't I don't usually expect
big announcements to look like that they
seem to be grandiose most of the time
like compare that to Google's oh my gosh
the eyes I've been looking at my eye too
much recently but compare that to
Google's keynote look at how massive
that is by the way Matt Wolf's over
there like you can't see him in this
Frame but he I know where he's hitting
because I saw his video but look at that
big stage that's so crowded that has
such a different feel to it and it's so
I mean it is kind of warm and colorful
and rainbowy but it's also very
corporate just the way he's standing and
how big they made the audience but more
than that there was like mystery
surrounding this that open AI made that
Google could have but they just didn't
like last week when that mysterious gpt2
chatbot I think it was called but just a
weird name shows up gets tested and
nobody knows who made it and it's more
powerful than gp4 and then it disappears
and Sam Alman does this like cryptic
thing like gpt2 is like a powerful model
or something we're like wait did you
guys launch that and then they did and
that's how they got the benchmarks and
that's why they did it but there's just
something so much deeper about that
narrative and that story and that
mystery that they're running and even
though there was a lot more to the GPT 4
zero model it's like they did a really
short presentation and they just focused
in on the one thing that they wanted it
to be known for and then the other stuff
kind of leaked out on the YouTube
channel a little bit later whereas
Google was like here's like 15
announcements and there's the Gmail
thing and here's the Google Docs thing
and here's the drive thing and here's
the upgraded 1.5 for everyone thing and
it was awesome but also so it was like
it took work like you took notes and
you're like how am I going to integrate
this at the end of the day we're just
story driven humuli hysterical bird
brain humuli MIT news has some
interesting things for us to chew on
scientists use generative AI to answer
complex questions in physics a new
technique can automatically classify
phases of physical systems that could
help scientists investigate novel
materials something I think about way
more than I showed it's kind of like my
Rome is Phase transitions but it's just
that moment you see it in all sorts of
complex systems where the sum becomes
bigger than the parts and in this
article they cover what is the k iCal
example so if you zoom in water you get
individual H2O atoms right and together
they do describe ice water and steam but
there is a Tipping Point and it's just
like a snap of a finger it's a moment
where they're jiggling around and
they're loose and it's water and then
you make it colder and colder and colder
and they start to slow down but it's not
ice yet it's still water they're just
moving a little bit slower and then it
just like you know they just they lock
into place and now it's a it's kind of a
different material it's still made of
the same atoms but it's very different
in how you describe its characteristics
so the MIT researchers started thinking
about all the data that they have about
phase transitions they trained a model
and they demonstrated how generative
models can be used to solve this
classification task much more
efficiently than physics informed tools
all right so back to open AI for a
minute they've introduced a new tool
that will guarantee that if my YouTube
channel doesn't work that I have
absolutely no job security to make money
in any other way and that is the new
improvements in data analysis from chat
GPT so most of my life before I started
working on YouTube was pretty Prett much
jupyter notebooks matplot lib and numpy
I kind of got familiar with those tools
and used them over and over and over and
over and over again to solve a bunch of
different problems pretty sure anybody
who would have hired me to do that
doesn't need me anymore which makes me
sad but also the future comes for all of
us and now there's a bunch of new tools
for customizing and downloading charts
for presentations and documents so
here's how it's going to work you're
going to be able to upload one or more
data files to chat gbt and it will
analyze your data by writing and running
python code on your behalf I played with
code interpreter like a year ago I was
already just so impressed but now it can
handle a range of data tasks like
merging and cleaning large data sets oh
that was a big part of what I did
creating charts and uncovering insights
what's really cool is this idea that we
can now work on tables in real time I
mean like look at this imagine having
the ability to just chat on the sidebar
with your spreadsheet this is crazy like
this is so powerful I mean now think of
at a company all of the people who have
access to data but just didn't really
have the tools to think about it or
understand it but now they can just say
something as simple as give me some
insights like should I buy more
inventory like can you build me a chart
of how much it's selling and make a
prediction if it's time to order more
all these kind of things like if this
plugs into something like neo4j where
you can build graphs too and like
Traverse edges and nodes like holy cow
that would be amazing you could do
knockout tests graph analysis just so
much all right let's talk about
artificial intelligence making airplane
travel safer of course you probably
already know about the panel that blew
off of the 737 Max mid-flight but one
thing you might not know is that
Aerospace manufacturers still face a
lack of extensive 3D CAD models for
legacy aircrafts like you think all this
stuff is just built to Super precise
modes in CAD and then simulate it in all
these different ways and like all the
pieces come together which I'm sure is
true for the most part but artificial
intelligence can take that to the next
level and it can be used for AI
powerered predictive maintenance and
that's something I hadn't thought a lot
about either like you could say vibrate
a digital version of an airplane a
digital twin and just see if anything
like rattles around or shakes off it and
then if it does you can use that as
training data like keep learning about
moments like this and it can start
focusing in on what are different ways
we could like jostle the airplane or
bend materials or change temperatures to
actually break it and then we would
start having predictive ideas of what's
going to break to me that's so so so
powerful and of course you can add this
arvr element so sure somebody you can
put on on Apple Vision Pro and like walk
around the airplane and like see where
the bolts are going to fall out but as
long as you could see it in the
simulation I think that's what really
matters like that's the important thing
is adding AI to these digital twin
simulations so 4 days ago on Hacker News
a developer writes this maybe it's just
me but the advances and artificial
intelligence are just leaving me feeling
disillusioned as a builder there's an
odd feeling that whatever I'm going to
build will just get gobbled up Away by
some big tech company and the demos are
becoming more cringy does anyone else
feel like they're wasting their time and
it dropped off the front page
potentially naturally I don't want to
say it was like Shadow band for sure but
it did seem like it was really something
everybody wanted to chime in on and then
it just kind of got cut pretty pretty
quick that's Up For Debate but what's
not up for debate is just how many
people felt the same way I'm not
disolution I'm just being realistic as
opposed to the dotcom boom AI is top
down rather than bottom up like it seems
like AI is going to give a lot of
Regular People the potential to build
stuff that they couldn't have before so
that feels bottom up to me but I would
argue yes it's going to come at the cost
of developers I think their jobs need to
turn into what are they going to build
and how are they going to sell it and
how they going to Market it and
everything's a brand nowadays but asking
Claude what is the overall sentiment
from this Hacker News comment section we
find overall the sentiment leans more
towards disillusionment concern and
uncertainty rather than an optimism and
excitement right now let's talk about
some of the latest AI research and let's
start by talking about master Weaver
taming editability and identity for
personal text to image generation so
here's the reference face so you might
upload a picture of this woman and you
say a smiling woman with curly red hair
she doesn't have red hair here but once
it goes through this new model here's
what you end up with which is really
impressive I see the same shape of the
nose the same sort of shape of the cheek
and the mouth but I do see like Curly
red hair everywhere right and I know she
could be looking red cuz I literally
have blood in my eye but I'm telling you
I think it looks like red hair but it's
interesting to see the progress because
I remember dream Booth being such a big
deal and it just looks like it's nothing
anymore compared to some of these other
models fast composer IP adapter and
photomaker which I'll do totally kind of
wet well I guess photomaker kind of kind
of hits it but this new model is
definitely the best throwing AI
researcher yon laon into an Iron Man
suit or with a Christmas hat on retains
his face his glasses all looks legit to
me and same with a reference image of
Jeffrey Hinton smiling which is rare and
red hair which is probably even more
rare so you change this woman to be in
the snow or on the beach I still see the
same facial features and this woman
while swimming or running also seems
about right okay so looking at a couple
other examples so all of these are
different versions like dream booth and
custom diffusion and photomaker and this
one on the far right is the actual one
that we're looking at today this one is
Master Weaver so if you take Elon Musk
you see a lot of these look pretty bad
like this one looks really bad but I
would say that theirs Is Not Great
honestly I feel like that's kind of mid
when it comes to Turning Taylor Swift
into an anime cartoon this one
definitely missed it but it did seem
like celeb basis did okay and I would
say that this one did well too see
adding purple hair to Scarlet Johansson
had some mixed results dream Booth did
good no no no no yeah and I would say
that theirs is the best like Master
Weaver really does seem to look like
Scarlett Johansson with purple hair next
up let's talk about Cat 3D create
anything in 3D with multi view diffusion
models so cool look at this paper coming
out of Google deepmind and Google
research you know it's going to be good
all right we got a snail a bear cat with
a funnel and oh like removed the
background and gave us like some wobble
very o very cool the snail's bending a
little weird oh look at that Q be full
rotation all the way around 360 oh
that's impressive I thought it was going
to get stuck at the um just like the
wobble look because with with Nerf it's
hard to go all the way around something
yeah like more like this although that's
the robot seems like he's keeping his
depth pretty good down here create 3D
scenes from any number of real or
generated images interesting all right
so let's check this out we got this beat
up car out in the woods wow that does
really kind of immerse you interesting
just one frame like that I was like okay
that looks like a defusion model but
look it it's cool you can pretty much oh
you can go all the way around I mean the
tire kind of breaks off down there and
everything but there's a lot of
Dimension to this so look at fully
rendered like that snow all around it
wow Google you're so they're so good
sometimes you know like how could you
just keep losing to open AI so much data
so much data so much research de mind
has been so good for so long I'm a fan
of the yarn I like that yarn bee like I
Love yarn bees for some reason I like
paper cut craft bees too but I like yarn
be second there you go Cat 3D it's just
that good all right this one you're
going to either love it or hate it this
is the platonic representation
hypothesis and the reason can be
summarized with this chart the more we
align the models across the board the
more similar they become now if they're
becoming more similar because they're
becoming more safe that's great if it's
taking away diversity or a voice or it's
becoming 2pc or it's becoming The
Narrative of one type of Western
thinking then it's a problem so that's
where the drama comes in but it seems
like from this paper that it's not just
about us using reinforcement learning to
force them to be the same it actually
seems like the smarter they get the more
they understand our universe and our
world and that's part of what's aligning
them the deeper the network gets the
more advanced it gets the more it starts
to think alike no matter what they're
trained on or how they're built so the
authors argue that this convergence is
being pushed towards a universal
statistical model of reality and that's
probably great because it could have big
implications for how we design AI
systems in the future to make them all
work well so For Better or Worse more
uniformity looks like it's in our future
all right next up let's talk with
transic a Sim tooreal policy transfer by
learning from online correct ction all
right so we got this light bulb changing
robot here and this is a Sim toore
policy transfer policy in AI is kind of
the rule set it's how the system can
move the choices that it can make to
achieve goals so they're basically
saying why don't you learn all these
policies the better the policy fits the
environment the better actually get the
outcome and the Big Challenge is what
works in simulations often doesn't make
it into the real robot there's just
something different about real life and
the way the simulation mismatched what
was going to happen so this system
called transic addresses that issue by
involving humans in the Loop essentially
humans oversee the robot's actions and
step in to correct mistakes in real time
so that's cool like you can just grab
the hand and be like oh you pinched it
too hard or move it over here and that's
what's called the residual policy that's
a fine-tuning system that changes the
original policy so now we've got some
research that's coming out of anthropic
it's called Simple probes can catch
sleeper agents sounds so cool Sleeper
Agent training deceptive llms that
persist through safety training so the
idea here is that during the initial
training of a model somebody puts in the
secret thing that it's really supposed
to supposed to do and supposed to keep
hidden from all the people testing it
and then once it gets out into the wild
they can like activate it in a special
way so what anthropic has done is
they've created these models that for
example write secure code when prompted
with the year 2023 but introduce
vulnerabilities when the year is 2024
yeah that's pretty sneaky so they found
that these back doors remained intact
despite various safety training methods
including supervised fine tuning
reinforcement learning and adversarial
training so yeah they put this in there
and then they put it through their
normal testing to see if they could
actually get rid of the sleeper agent
completely and they were still able to
activate it well what if some of these
open source models maybe coming out of
meta or BYO or even Google's got an open
source one now what if somebody there is
like hey let's put in the sleeper agent
so they also said that interestingly the
largest model and those using Chain of
Thought reasoning were most resistant to
these safety measures and unfortunately
there is no happy ending to this that's
the whole paper just like we couldn't
kill the sleeper agent so I guess we
should all think about that one and work
on it next all right quick update on the
last video it's popping off 6.9 th000
views when I was recording last night it
was only an hour old and it looked like
it was going to do good but I wasn't
sure if it was going to keep up but it
did so let me know if you think it was
the title thumbnail combo I did start
using AI news colon which is what Matt
wolf does so maybe that helped people
understand what they're getting into
also I use chat GPD in the title which
seems to help maybe new in caps or maybe
the arrow with the logo pointing to the
dog insane or something I don't know but
that there was a nice combination there
you can see I got a freaking 50 new
subscribers thank you if you're one of
those new people I love that the channel
is growing I made $5 I got five on it
retention was pretty typical 66% of you
made it to the 302nd mark which is
pretty good for me YouTube was nice
enough to give me 50,000 Impressions and
that correlated with a 8.2%
click-through rate so 6.9 th000 views
that's crazy and a lot of them were
unique I mean it was all coming in from
the browse feature but that's pretty
common for a growing channel so I'm
really happy with that there's one video
suggesting my content more than ever
everyone else let me show that to you
because that is this video for some
reason the AI grid so a lot of people at
the end of this video were getting
recommended my video right after so not
sure how I connected on that I like the
AI grid I don't I didn't talk about them
recently so who knows algorithm does
crazy things and a nice healthy dose of
subscribed and unsubscribed so that out
of the 7,000 views at least half of them
were new people so that's great
unfortunately it did not show my video
to any females again so it was a 100%
male audience so all 7,000 views were
men feels like AI is very disruptive but
I like that 45 to 54 year olds are
actually caring about this and 3544
makes sense and then 2534 yeah I mean
that's a nice kind of curve there yeah
being in the United States speaking
English I'm not surprised Australia UK
Canada a little bit of Germany down
there and if you were wondering about
that counter in the bottom corner I got
an email from somebody who had this idea
that maybe there should be a magic word
counter in the lower corner of the
screen and I'll use a relevant word for
the video in this case it was open Ai
and we counted it so yeah let me know if
that counter down there kept you engaged
so if that's one of the reasons you made
it this far in the video let me know in
the comments below thanks for watching
浏览更多相关视频
注目AIニュース20選~Copilotで作業自動化、Copilotノートブック、MetaのLlama3、Appleが生成AI集中、AIしまじろう
Microsoft's new "Embodied AI" SHOCKS the Entire Industry! | Microsoft's Robots, Gaussian Splat & EMO
注目AIニュース9選~iPhoneにChatGPT搭載!?、OpenAIの新発表、XのGrokいよいよリリース
'AI Superpowers': A Conversation With Kai-Fu Lee
NEW Copilot in Azure AI Studio *2024*
松田語録:AIサイエンティスト続編〜AI研究を行うAIがついに恐ろしいことをやり始めた?
5.0 / 5 (0 votes)