How Far Can We Scale AI? Gen 3, Claude 3.5 Sonnet and AI Hype

Summary

The script discusses the rapid advancements in AI video generation, exemplified by models like Runway Gen 3 and the anticipated Sora from Open AI. It raises questions about the reliability of AI leaders and the scalability of language models, highlighting Claude 3.5 Sonic's capabilities and the incremental improvements in AI. The potential for AI in fields like drug discovery is mentioned, with a cautionary note on separating AI hype from reality and the unpredictable impact of scaling and new research.

Takeaways

• 🌐 AI video generation is rapidly advancing and becoming more accessible, with models like Runway Gen 3 and Sora promising highly realistic outputs despite training on a small fraction of available video data.
• 🎥 The Luma dream machine offers an engaging way to experiment with AI-generated images, allowing users to interpolate between two images or generate new ones.
• 📈 There is skepticism about the continuous scaling of language models, with concerns about whether increased scale will necessarily lead to more accurate or reliable AI.
• 🤖 The release of advanced voice models, such as OpenAI's real-time advanced voice model, has been delayed to improve content detection and refusal capabilities.
• 📊 Benchmark results for AI models like Claude 3.5 Sonic show improvements with increased compute, but the gains are incremental and not proportional to the scale.
• 🧠 The potential of AI in fields like biology and drug discovery is being discussed, with some suggesting AI could accelerate the rate of discoveries, though this is speculative.
• 🔍 There is a call for caution in interpreting benchmark results and a recognition that AI models still struggle with basic tasks, indicating that scale alone may not solve all issues.
• 📚 The script highlights the importance of metacognition in AI development, suggesting that understanding how to think about problems is as crucial as scaling computational power.
• 🤝 Open source AI models are seen as a way to encourage innovation and allow for diverse applications, contrasting with the idea of a single 'true AI'.
• 🚧 There is acknowledgment from AI leaders that the field is moving fast, and there is a need to ensure that understanding keeps pace with the capabilities of AI models.
• 🔮 Predictions about the future of AI are made with caution, acknowledging the many unknowns and the difficulty in forecasting exact outcomes or timelines for AI advancements.

Q & A

What is the current state of AI video generation technology?

AI video generation technology is rapidly advancing, with models such as Runway Gen 3 becoming more accessible and generating increasingly realistic outputs. However, these models are likely trained on less than 1% of available video data, indicating that future generations could become even more realistic in a relatively short time.

What is the significance of the Luma Dream Machine in AI image generation?

The Luma Dream Machine is a tool for AI image generation that allows users to create images or interpolate between two real ones. It represents a fun and engaging way for users to experiment with AI-generated visuals while waiting for the release of more advanced models.

What is the status of the video generation model called Sora from Open AI?

Sora is a highly anticipated video generation model from Open AI, which is considered to be the most promising in its field. However, it is still under development, and comparisons with other models like Runway Gen 3 show that it may have benefits from larger scale training and compute resources.

Why is the release of the real-time advanced voice mode from Open AI delayed?

The release of the real-time advanced voice mode from Open AI has been delayed to improve the model's ability to detect and refuse certain content. It also addresses concerns about the model occasionally producing inaccurate or unreliable outputs.

What are some of the limitations of scaling AI models?

While scaling AI models can lead to improvements, it does not necessarily solve all problems. For example, even with more data and compute, models may still struggle with basic tasks or produce hallucinations in language generation. The hope is that scaling will eventually lead to more accurate and reliable AI, but there is skepticism about whether this will fully address current limitations.

What is the current performance of Claude 3.5 Sonic from Anthropic compared to other language models?

Claude 3.5 Sonic is a language model from Anthropic that is free, fast, and in certain domains, more capable than comparable language models. It shows improvements in basic mathematical ability and general knowledge compared to models like GPT-40 and Gemini 1.5 Pro from Google, but the differences are not as significant as the scale of compute might suggest.

What is the concept of 'metacognition' in the context of AI development?

Metacognition in AI refers to the ability of models to understand how to think about a problem in a broad sense, to assess the importance of an answer, and to use external tools to check their answers. It represents a significant frontier in AI development, moving beyond simple scaling to more human-like thinking processes.

What are some of the challenges in accurately assessing the capabilities of AI models?

Assessing the capabilities of AI models is challenging due to the limitations and flaws in benchmark tests, which may not accurately reflect real-world performance. Additionally, models may struggle with basic tasks despite high scores on benchmarks, indicating a need for more nuanced evaluation methods.

How do AI lab leaders view the potential of AI in fields like biology and drug discovery?

Some AI lab leaders, such as those from Anthropic, see the potential for AI to significantly impact fields like biology and drug discovery, possibly leading to new discoveries and even cures for diseases. However, these views are sometimes seen as overly optimistic or 'hype' by others in the industry.

What is the current sentiment regarding the hype around AI capabilities and scaling?

There is a growing sentiment that the hype around AI capabilities and scaling may have gone too far, with some industry insiders expressing skepticism about the pace of progress and the reliability of benchmark results. The challenge is to separate hype from reality and to manage expectations about what AI can and cannot do.

What are the potential future developments in AI model scaling and algorithmic improvements?

Future developments in AI are expected to include scaling to models with billions of parameters and continued improvements in algorithms and chip technology. If these advancements continue at their current pace, there is a possibility that by 2025-2027, AI models could surpass human capabilities in many areas.

Outlines

00:00

🌐 AI Video Generation and Model Scalability

The script discusses the rapid advancements in AI video generation, highlighting the capabilities of models like Runway Gen 3 and Sora from OpenAI. It emphasizes the transformative potential of these technologies on content consumption. The author also raises questions about the reliability of AI leaders and the limitations of scaling language models. Comparisons between different models, such as Runway Gen 3 and Sora, illustrate the improvements in video generation quality. The script also touches on the potential and limitations of AI in understanding and generating accurate world models.

05:00

📈 Incremental Improvements and the Economics of AI Scaling

This paragraph delves into the incremental improvements brought by scaling AI models, questioning the economic viability of investing in such advancements. It uses Claude 3.5 Sonic from Anthropic as a case study, comparing its performance to other models and highlighting the diminishing returns on investment as models approach human-level intelligence. The author also explores the concept of 'metacognition' in AI, suggesting that understanding how to think about a problem is as important as scaling computational power.

10:03

🤖 AI's Evolving Capabilities and the Hype Surrounding Them

The script addresses the hype around AI's capabilities, contrasting it with the reality of current technology. It cites comments from industry leaders who express skepticism about the pace and impact of AI advancements. The author discusses the potential of AI in fields like drug discovery and cancer treatment, while also acknowledging the uncertainty and the need for caution in interpreting AI's capabilities. The paragraph concludes with a call for a balanced view between optimism and realism in the AI community.

15:03

🚀 The Future of AI: Scaling, Research, and Real-World Applications

In the final paragraph, the script contemplates the future trajectory of AI, considering both scaling and algorithmic improvements. It presents predictions from industry experts about the potential of AI to surpass human capabilities in various domains. The author also reflects on the unpredictability of AI's impact and the importance of separating hype from reality. The paragraph ends with an invitation for the audience to engage in further discussions on AI's future through the author's Patreon platform.

Keywords

💡AI Video Generation

AI Video Generation refers to the creation of video content by artificial intelligence. In the video script, it is highlighted as a transformative technology that is making artificial worlds more tangible and accessible. The script discusses the capabilities of models like Runway Gen 3 and Sora, emphasizing the potential for highly realistic video generation despite being trained on a small fraction of available video data.

💡Scaling

Scaling in the context of AI refers to the increase in the size and complexity of models, often through the addition of more data and computational power. The script raises questions about the merits and limitations of scaling, suggesting that while it can lead to improvements, it may not necessarily solve all challenges, such as the accuracy of world models or reasoning abilities.

💡Language Models

Language models are AI systems designed to understand and generate human language. The script mentions Claude 3.5 Sonic and compares it with other models like GPT-40, discussing their capabilities in various domains such as mathematical ability and general knowledge. It also touches on the incremental improvements and the cost of scaling these models.

💡Multimodal Training

Multimodal training involves training AI models on multiple types of data, such as text, images, and video. The script discusses the hope that multimodal training could improve reasoning abilities in AI, but also notes that this has not been entirely successful, as evidenced by the performance of models like Claude 3.5 Sonic.

💡Hype vs. Reality

The script explores the balance between the hype surrounding AI advancements and the actual capabilities and limitations of current technology. It critiques the tendency to overstate the potential of AI, urging for a more measured and realistic assessment of its progress and implications.

💡Benchmarks

Benchmarks are tests or metrics used to evaluate the performance of AI models. The script expresses skepticism about the reliability of benchmark results, suggesting that they may not fully capture the capabilities or limitations of AI models, and that minor differences in scores may not be significant.

💡Metacognition

Metacognition refers to the ability of an AI to understand and reflect on its own thought processes. Bill Gates mentions metacognition as the 'big frontier' in AI development, suggesting that progress in this area could be more significant than mere scaling of models.

💡Emergent Behaviors

Emergent behaviors are unexpected or novel capabilities that arise in complex systems as a result of interactions between simpler components. The script briefly mentions the idea that certain AI capabilities might emerge as models scale up, but also questions whether this will necessarily lead to the desired outcomes.

💡Open AI

Open AI is an organization focused on the development of AI technologies. The script discusses Open AI's work on models like Sora and GPT-40, as well as their efforts to improve the reliability and reasoning abilities of their AI systems.

💡Anthropic

Anthropic is another AI research organization mentioned in the script, known for developing models like Claude 3.5 Sonic. The script discusses the company's approach to AI development and their public statements about the potential impact of their technology.

💡AI Ethics and Responsibility

The script touches on the ethical considerations and responsibilities of AI development, such as the potential for AI to be used in sensitive areas like cancer research. It also discusses the need for caution in making claims about AI capabilities and the importance of ensuring that AI development is aligned with societal values.

Highlights

AI video generation is becoming increasingly tangible and accessible, set to transform content consumption.

Runway gen 3 and its audio, generated by AI, represents the current state of AI video generation.

AI models are trained on a fraction of available video data, indicating potential for more realistic video generation.

Luma dream machine allows for image generation and interpolation, offering a fun user experience.

Chinese model 'cing' and the anticipated release of 'Sora' from OpenAI are compared for video generation capabilities.

Training models on more data doesn't guarantee accurate world models, as seen in Sora's generation capabilities.

The question of whether scale in AI will solve all issues is raised, with indications it may not.

Real-time advanced voice mode from OpenAI, showcased in GPT-40's demo, has been delayed.

Claude 3.5 Sonic from Anthropic offers free and fast capabilities in certain domains.

Benchmarks for AI models have significant flaws and should be interpreted with caution.

The incremental benefits of scaling in AI models may not justify the increased cost and compute.

Artifacts feature in Claude 3.5 Sonic allows for interactive project work alongside the language model.

AI models still struggle with basic tasks despite increased scale, contradicting naive scaling hypotheses.

Bill Gates discusses the potential of scaling in AI and the need for understanding beyond just data access.

Metacognition in AI is identified as a significant frontier for development.

Mustafa Suleyman from Microsoft AI suggests that consistent action from AI models may not be achievable until GPT-6.

Sam Altman from OpenAI discusses the use of AI in cancer screening and potential future contributions to discovering cures.

Mark Zuckerberg expresses skepticism about the grandiose claims made by AI lab leaders.

Dario Amodei from Anthropic discusses the potential of AI in biology and drug discovery, with caveats about the uncertainty of predictions.

The pace of AI development is a challenge, with the need to ensure understanding keeps pace with capabilities.

AI models are compared to undergraduates, with the potential to reach professional levels in various fields.

Transcripts

artificial worlds generated by AI video

models have never been more tangible and

accessible and look set to transform how

millions and then billions of people

consume content and artificial

intelligence in the form of the new free

Claude 3.5 Sonic is more capable than it

has ever been but I will draw on

interviews in the last few days to show

that there are more questions than ever

not just about the merits of continued

scaling of language model

but about whether we can rely on the

words of those who lead these giant AI

orgs but first AI video generation which

is truly on fire at the moment these

outputs are from Runway gen 3 available

to many now and to everyone apparently

in the coming days the audio by the way

is also AI generated this time from udio

[Music]

and as you watch these videos remember

that the AI models that are generating

them are likely trained on far less than

1% of the video data that's a available

unlike highquality text Data video data

isn't even close to being used up expect

generations to get far more realistic

and not in too long either and by the

way if you're bored while waiting on the

Gen 3 wait list do play about with the

Luma dream machine I've got to admit it

is pretty fun to generate two images or

submit two real ones and have the model

interpolate between them now those of

you in China have actually already been

able to play with model of similar

capabilities called cing but we are all

waiting on the release of Sora the most

promising video generation model of them

all from open AI here are a couple of

comparisons between Runway gen 3 and

Sora the prompts used in both cases are

identical and there's one example that

particularly caught my eye as many of us

may have realized by now simply training

models on more data doesn't necessarily

mean they pick up accurate world models

now I strongly suspect that Sora was

trained on way more data with way more

compute with its generation at the

bottom you can see that the dust emerges

from behind the car this neatly

demonstrates the benefits of scale but

still leaves open the question about

whether scale will solve all now yes it

would be simple to extrapolate a

straight line upwards and say that with

enough scale we get a perfect world

simulation but I just don't think it

will be like that and there are already

more than tentative hints that scale

won't solve everything more on that in

just a moment but there is one more

modality I am sure we were all looking

forward to which is going to be delayed

that's the realtime advanced voice mode

from open AI it was the star of the demo

of GPT 40 and was promised in the coming

weeks alas though it has now been

delayed to the fall or the Autumn and

they say that's in part because they

want to improve the model's ability to

detect and refuse certain content I also

suspect though like dodgy physics with

video generation and hallucinations with

the language generation they also

realized it occasionally goes off the

rails now I personally find this funny

but you let me know whether this would

be acceptable to release refreshing

coolness in the air that just makes you

want to smile and take a deep breath of

that crisp invigorating Breeze the Sun's

shining but it's that this lovely gentle

warmth that's just perfect for light

Jack so either way we're definitely

going to have epic entertainment but the

question is what's next particularly

when it comes to the underlying

intelligence of models is it a case of

shooting past human level or diminishing

returns well here's some anecdotal

evidence with the recent release of

Claude 3.5 Sonic from anthropic it's

free and fast and in certain domains

more capable than comparable language

models this table I would say shows you

a comparison on things like basic

mathematical ability and general

knowledge compared to models like GPT 40

and Gemini 1.5 Pro from Google I would

caution that many of these benchmarks

have significant flaws so decimal point

differences I wouldn't pay too much

attention to the most interesting

comparison I would argue is between

Claude 3.5 Sonic and Claude 3 Sonet

there is some evidence that Claude 3.5

Sonic was trained on about four times as

much compute as Claude 3 on it and you

can see the difference that makes

definitely a boost across the board but

it would be hard to argue that it's four

times better and in the visual domain it

is noticeably better than its

predecessor and than many other models

and I got Early Access so I tested it a

fair bit these kind of benchmarks test

reading charts and diagrams and

answering basic questions about them but

the real question is how much extra

compute and therefore money can these

companies continue to scale up and

invest if the returns are still

incremental in other words how much more

will you and more importantly businesses

continue to pay for these incremental

benefits after all in no domains are

these models reaching 100% and let me

try to illustrate that with an example

and as we follow this example ask

yourself whether you would pay four

times as much for a 5% hallucination

rate versus an 8% hallucination rate if

in both cases you have to check the

answer anyway let me demonstrate with

the brilliant new new feature you can

use with Claude 3.5 Sonic from anthropic

it's called artifacts think of it like

an interactive project that you can work

on alongside the language model I dumped

a multi hundred page document on the

model and asked the following question

find three questions on functions from

this document and turn them into

clickable flash cards in an artifact

with full answers and explanations

revealed interactively it did it and

that is amazing but there's one slight

problem question one is perfect it's a

real question from the document

displayed perfectly and interactive with

the correct answer and explanation same

thing for question two but then we get

to question three where it copied the

question incorrectly worse than that it

rejigged and changed the answer options

also is there a real difference between

q^2 and netive Q ^2 when it claimed that

netive Q ^2 is the answer now you might

find this example trivial but I think

it's revealing don't get me wrong this

feature is mentally useful and it

wouldn't take me long to Simply tweak

that third question and by the way

finding those three examples strewn

across a multi hundred page document is

impressive even though it would save me

some time I would still have to

diligently check every character of

claude's answer and at the moment as I

discussed in more detail in my previous

video there is no indication that scale

will solve this issue now if you think

I'm just quibbling and benchmarks show

the real progress well here is the the

reasoning lead at Google deepmind

working on their Gemini series of models

someone pointed out a classic reasoning

error made by Claude 3.5 Sonic and Denny

XE said this love seeing tweets like

this rather than those on llms with phds

superhuman intelligence or fancy results

on leaked benchmarks I'm definitely not

the only one skeptical of Benchmark

results and an even more revealing

response to Claude 3.5s basic errors

came from open AI know Brown I think

it's more revealing because it shows

that those AI Labs anthropic and open AI

had their hopes slightly dashed based on

the results they expected in reasoning

from multimodal training non Brown said

Frontier models like GPT 40 and now

clawed 3.5 Sonic maybe at the level of a

quote smart high schooler mimicking the

words of Mira murati CTO of open aai in

some respects but they still struggle on

basic tasks like Tic-tac-toe and here's

the key quote there was hope that native

multimodal training would help with this

kind of reasoning but that hasn't been

the case that last sentence is somewhat

devastating to the naive scaling

hypothesis there was hope that native

multimodal training on things like video

from YouTube would teach models a world

model it would help but that hasn't been

the case now of course these companies

are working on Far More Than Just naive

scaling as we'll hear in a moment from

Bill Gates but it's not like you can

look at the benchmark Mark results on a

chart and just extrapolate forwards

here's Bill Gates promising two more

turns of scaling I think he means two

more orders of magnitude but notice how

he looks skeptical about how that will

be enough the big Frontier is not so

much scaling we have probably two more

turns of the crank on

scaling where by accessing video data

and getting very good at synthetic

data that we can scale up probably

you know two more times that's not the

most interesting Dimension the most

interesting Dimension is is what I call

metacognition where understanding how to

think about a problem in a broad sense

and step back and say okay how important

is this answer how could I check my

answer you know what external tools

would help me with this so we're get

we're going to get the scaling benefits

but at the same time the various

actions to change the the underlying

reing algorithm from the trivial that we

have today to more humanlike

metacognition that's the big

Frontier that uh it's a little hard to

prict how quickly that'll happen you

know I've seen that we will make

progress on that next year but we won't

completely solve it uh for some time

after that and there were others who

used to be incredibly bullish on scaling

that now sound a little different here's

Microsoft ai's CEO Mustafa sullan

perhaps drawing on lessons from the

mostly defunct inflection AI that he

used to run saying it won't be until GPT

6 the AI models will be able to follow

instructions and take consistent action

there's a lot of cherry-picked examples

that are impressive you know on Twitter

and stuff like that but to really get it

to consistently do it in novel

environments is is pretty hard and I

think that it's going to be not one two

orders of magnitude more computation of

training the models um so not gbt 5 but

more like gbt 6 scale models so I think

we're talking about two years before we

have systems that can really take action

now based on the evidence that I put

forward in my previous video let me know

if you agree with me that I still think

that's kind of naive reasoning

breakthroughs will rely on new research

breakthroughs not just more scale and

even samman said as much about a year

ago saying the ear ER of ever more

scaling of parameter count is over now

as we'll hear he has since contradicted

that saying current models are small

relative to where they'll be but at this

point you might be wondering about

emergent behaviors don't certain

capabilities just spring out when you

reach a certain scale well I simply

can't resist a quick plug for my new

corsera series that is out this week the

second module covers a mergent behaviors

and if you already have a corsera

account do please check it out it' be

free for you and if you were thinking of

getting one there will be a link in the

description anyway here's that quote

from samman somewhat contradicting the

comments he made a year ago models he

says get predictably better with scale

we're still just like so early in

developing such a complex system um

there's data issues There's algorithmic

issues uh the models are still quite

small relative to what they will be

someday and we know they get predictably

better but this was the point I was

trying to make at the start of the video

as I argu in my previous video I think

we're now at a time in AI where we

really have to work hard to separate the

hype from the reality simply trusting

the words of the leaders of these AI

Labs is less advisable than ever and of

course it's not just samman here's the

commitment from anthropic led by Dario

amade back last year they described why

they don't publish their research and

they said it's because we do not wish to

advance the rate of AI capabilities

progress but their CEO 3 days ago said

AI is progressing fast due in part to

their own efforts to try and keep Pace

with the rate at which the complexity of

the models is increasing I think this is

one of the biggest challenges in the

field the field is moving so fast

including by our own efforts that we

want to make sure that our understanding

keeps Pace with our our abilities our

capabilities to produce powerful models

he then went on to say that today's

models are like undergraduates which if

you've interacted with these models seem

seems pretty harsh on undergraduates if

we go back to the analogy of like

today's models are like

undergraduates uh you know let's say

those models get to the point where you

know they're kind of you know graduate

level or strong professional level think

of biology and Drug Discovery think of

um a model that is as strong as you know

a Nobel prize winning scientist or you

know the head of the you know the head

of head of drug Discovery at a major

pharmaceutical company now I don't know

if he's placing that on a naive trust in

benchmarks or whether he is deliberately

hyping and then later in the

conversation with the guy who's in

charge of the world's largest Sovereign

wealth fund he described how the kind of

AI that anthropic works on could be

instrumental in curing cancer I look at

all the things that have been invented

you know if I look back at biology you

know crisper the ability to like edit

genes if I look at um you know C

therapies which have cured certain kinds

of cancer

there's probably dozens of discoveries

like that lying around and if we had a

million copies of an AI system that are

as knowledgeable and as creative about

the field as all those scientists that

invented those things then I think the

rate of of those discoveries could

really proliferate and you know some of

our really really

longstanding diseases uh you know could

be could be addressed or even cured now

he added some caveats of course but that

was a claim echoed on the same day

actually I think by open AI Sam mman one

of our partners color health is now

using uh gb4 for cancer screening and

treatment plans and that's great and

then maybe a future version will help uh

discover cures for cancer other AI lab

leaders like Mark Zuckerberg think those

claims are getting out of hand but you

know part of that is the open source

thing too so that way other companies

out there can create different things

and people can just hack on it

themselves and mess around with it so I

guess that's a pretty deep worldview

that I have and I don't know I find it a

pretty big turnoff when people in the um

in the tech industry kind of talk about

building this one true AI it's like it's

almost as if they they kind of think

they're creating God or something and

it's it's like it's just that's not

that's not what we're doing that's I

don't think that's how this plays out

implicitly he's saying that companies

like open aai and anthropic are getting

carried away and later though in that

interview the CEO of anthropic admitted

that he was somewhat pulling things out

of his hat when it came to biology and

actually with scaling you know let's say

you know you extend people's productive

ability to work by 10 years right that

could be you know one six of the whole

economy do you think that's a realistic

Target I mean again like I know some

biology I know something about how the

AA models are going to happen I wouldn't

be able to tell you exactly what would

happen but like I can tell a story where

it's possible so so 15% and when will be

so when could we have added the

equivalent of 10 years to our life I

mean how long what what's the time frame

again like you know this involves so

many unknowns right if I if I try and

give an exact number it's just going to

sound like hype but like a thing I could

a thing I could imagine is like I don't

know like two to three years from now we

have ai systems that are like capable of

making that kind of Discovery 5 years

from now those those discoveries are

actually being made and 5 years after

that it's all gone through the

regulatory apparatus and and really so

you know we're talking about more we're

talking about you know a little over a

decade but really I'm just pulling

things out of my hat here like I don't

know that much about drug Discovery I

don't know that much about biology and

frankly although although I invented AI

scaling I don't know that much about

that either I can't predict it the truth

of course is that we simply don't know

what the ramifications will be of the

scaling and of course of new research

regardless these companies are pressing

ahead uh right now 100 Mill ion there

are models in training today that are

more like a billion I think if we go to

10 or 100 billion and I think that will

happen in 2025 2026 maybe 2027 and the

algorithmic improvements continue a pace

and the chip improvements continue a

pace then I think there there is in my

mind a good chance that by that time

we'll be able to get models that are

better than most humans at most things

but I want to know what you think are we

at the dawn of a new era in

entertainment and intelligence or has

the hype gone too far if you want to

hear more of my Reflections do check out

my podcasts on patreon on AI insiders

you could also check out the dozens of

bonus videos I've got on there and the

live meetups arranged via Discord but

regardless I just want to thank you for

getting all the way to the end and

joining me in these wild times have a

wonderful day