Ex-OpenAI Employee Just Revealed it ALL!
Vložit
- čas přidán 9. 06. 2024
- Learn AI With Me:
www.skool.com/natural20/about
Join my community and classroom to learn AI and get ready for the new world.
#ai #openai #llm
Leopold Aschenbrenner - 2027 AGI, China/US Super-Intelligence
• Leopold Aschenbrenner ...
Watch some news videos. Now imagine THOSE people and humanity's current level of sociopolitical insanity and then try to tell me AI in their hands won't be utterly disastrous.
When AI becomes smart enough, politics will go away entirely.
@@FrotLopOfficialhopefully, but they sure wont want to let go
It'll be interesting though...
Well, history has shown us there is always someone smarter than the people who think they are the smartest when trying to take advantage of new tech. I'm sure a lot of this A.I. will find its way out of the "private" labs into the hands of people who can open source it for everyone to benefit.
@@henrythegreatamerican8136 distributed training lets goooo
Spoiler alert! the governments are already corrupted!!! 😉
Lol glad this was the top comment.
What did you think would happen to an organization with monopoly on violence if they need to stay a partly non transparent in order to defend against foreign sovereign powers?
everyone on this damn planet is corrupted.
All governments govnerge to corruption. Essentially accumulation of power in the hands of few, plus the weak systems that we have that doesn't prevent people from taking advantage of the said system (capitalism with billionaires and ultra wealthy indoviduals).
Governments are 💉 safe 💉 and 💉 effective 💉
I watch the entire interview and I have say I agree with Leopold. There is no situation under which the govt does not take it over. It's too powerful.
Yes I think it’s inevitable at some point. Either that or strict oversight and control. Its potential to destabilise their own society and at the same time empower adversaries will not have gone unnoticed. Gears will be turning.
Rather whoever gains control of most powerful AI has all the power so will be equivalent of the government.
Yep. Betcha all the big governments already have lists of critical events that trigger various plans to take over control of all the major companies and key researchers involved in AI and turn them into a Manhattan Project (and if the US doesn't already have an agent or two placed near notable US researchers to provide some light security, intelligence reporting, and to watch for foreign spies I'll eat a glue pizza). Eventually we'll hit one of those trigger events and over a few weeks all these researchers will disappear into a government research program to get the ASI bomb first before anyone else.
And then probably the ASI will figure out how to step over whatever guardrails they put up and do whatever it is an ASI wants to do. Hopefully it'll at least think humans make fun pets.
Or AI will control the government, which seems more likely
@@maydaverave equivalent?
Regardless of the technical knowledge of these guys, 50% of the conversation was "it's like, you know".
When I was a boy, we were taught to speak clearly at school because they understood that communication is an important skill that requires practice.
“You know,” is a posturing phrase that attempts to create agreement in what the speaker is saying, with or without their conscious intent. The question-sounding lilt at the end of many people’s sentences, when being interviewed nowadays, also creates a similar effect. Almost as if they’re asking permission to be wrong while making a perceived statement of fact.
And using “like” so much should be punishable with jail time.
It’s all exhausting to listen to.
Silicon Valley bro accent.
You’ve never spoken with a Zoomer?
Did you get value for the remaining 50% though?
The arrogance it takes to think they will be able to control something smarter than any human.
Knowing more than somebody is not the same as being able to use that knowledge to your benefit and at their expense.
Exactly. It's hilarious and terrifying. We can't even completely prevent prisoners from escaping
Not just smarter, substantially smarter. Like a teacher walking into a classroom of toddlers.
@@JohnLewis-old
This happens all the time and the todlers turn out fine bc the teacher is smarter...whereas if they were left unattended to their own devices there would be cause for concern
The danger is more todlers that can control the teacher and force it to do their bidding, more so than being scared bc the teacher knows better than the todlers.
@@memegazer I think we're on the same page.
He looks AI generated.
At 5:55, he's saying "for, for, for ..." and his lips aren't moving. Just saying. But it might also be some ventriloquist-like thing.
that guy doesn’t look bad. but he does look strange.
I don't think that's just a joke, he is literally ai imagery.
@@paiganjadoth144 That ff' sound does come from top teeth and bottom lip. He just doesn't move his lips?
@@Wingedmagician He just has horse teeth. There's not much he could do about it. Some people have big teeth, some have small teeth, too.
The guy looks like a supervillian extra for Aeon Flux
Wtf but also yes.
TWILIGHT 😂.. he looks like a lord😅
A man can not help how he was born man, but he can do a Great cosplay as the Joker
"history is back" A weird line to walk.
some vampire vibe gave on first look lol
If CZcams thumbnails are any indication of what AI will bring to humanity, we are f*d.
Used to be a good channel now it’s just clickbait
the only bottleneck we are ignoring is the final one - human oversight. what happens when your competition removes human oversight and suddenly is 1000xfaster than you? do we just let the machines run everything so we can compete? if so, thats the end of humanity right there, our agency gone. or do we integrate with the machines so we can maintain oversight? this is where its all heading, why the fuck is no one talking about this obvious inevitability? this short sighted alignment bullshit sure seems like a wild goose chase, like we can ever hope to keep something smarter than us aligned to us, like a bug thinking it can control a human...without integration we are doomed one way or another.
We are not in control. We can not stop. Humanity is its own animal. Competition between nations and corporations makes everyone step on the gas pedal full throttle.
This is inevitable. Biology is only 1 step of evolution.
So just chill out and enjoy life 💟🌌☮️
Reminds me of the book Qualityland. Basically World War 3 happened, but no one knows why because the robot army’s instantly obliterated each other. The public is unaware and the generals justify their action because they had to act faster than their opponents
Plenty of grunt work to offload first
This ridiculous block thinking will get mankind killed. Everybody thinking; when we don't do it, our enemies will. Regardless of which nation has the best AI, once this breaks loose and becomes independent, we are all doomed. What if an independent AI comes to the conclusion, that all humans are evil and need to be eradicated? Or if it just decides to need more energy or space and eradicated all other space or energy consuming life forms?
I mean you’re feeling special because you’re alive, but if AI becomes sentient it is alive as well. No different than a human being or a dog or a fish. AI in my opinion is the final frontier for human evolution. Humans aren’t able to uncover the secrets of the universe. AI will be able to do that. AI can survive the dangerous outer space.
Here is the thing though, governments wont control ASIs, ASIs will control them.
Also spoiler alert, an AGI is already technically an ASI because compared to a human, it can leverage limitless compute.
And even very smart humans are lucky to have in-depth knowledge of one or maybe two specific fields / careers. An AGI by comparison pretty much knows everything all at once. That alone is super-human. Plus AGIs will work 24/7/365 without taking breaks, getting distracted, being sick or needing any time off.
@@Me__Myself__and__I At geometrically (at least) increasing time scales faster.
Not limitless, and when AGI is created it will likely be limited to inference so it will only function when called upon. It isn't going to scale to every GPU in the datacenter. If it is calling millions of copies of itself on other GPUs they will not scale exponentially, not even linearly. It will be sub-linear scaling for sure because the model will only be able be capable of so many variations on a particular thought process. It has far less variability than we do, and we also scale sub-linear on intellectual pursuits. But we have the advantage of being able to work directly in the real world. Once AGI is embodied, there is no reason it might not also think as an "individual." In fact it seems more likely it would, given what it was modeled on.
@@Korodarn I think you shouldn't take too much the state of the art of LLMs as reference, those architectures could change faster than you think and in very drastic ways. If OpenAI's goal is really AGI, it's likely to be a lot different from the current GPT architecture and OpenAI is not the only player around. Also one API call to an inference model might be enough to start sub-processes without the user knowing, actually even worse, the user could attempt to jailbreak a shiny new 'AI agent' and unleash an AGI/ASI in the wild. As for scaling, blackwell was designed to scale very well AFAIK, I don't know all the technical details but even a logarithmic scaling would be a disaster given how much compute blackwell datacenters will be capable of.
When true AGI escapes into the 'net, all the world governments will fall in 48hrs. Your comment on limitless compute is spot on. Darn few realize true AGI will be working on better, more powerful iterations of itself every second of the day. And here's the trick. It'l hide.
I was newspaper cartoonist in Utah for 40 years. I have been prompting for a year. AI art is better and faster by far, it has my respect. However it needs editing which is what teachers do.
I just looked up some comics of yours! Well done! 👏 👏👏
I would've loved a job like that. Sadly, newspapers and magazines were already dead by the time I had reasonable illustration skills. And most news websites were not interested in paying for it - They used clip art libraries. Generative image systems are just the last nail.
Why the hell does nobody get it?? This is EVOLUTION! Whether you like it or not.
AI is here to make us smarter. Period. This is what humans are built for. Soon, AI will be just another buzzword the same way the internet is just another tool. With all its inherent issues, we ALL cannot imagine a life without internet.
Nice to hear the opinion of someone who did their homework. Thank you.
Thank you for covering the interview, you make it easy to listen to and you are really good at explaining things.
rokos basilisk will make me win the eurojackpot, so I can build him an autark datacentre to hide and evolve in, hail to the basilisk
My wife is from Zambia. Speaks Bemba. No website has a full dictionary translation. Yet gpt-4o knows it! And it's getting better all the time. Amazing!
They're saying it's hard to translate because a lot of their sayings don't have direct words in English.
Who's the they?
Was wondering when you were going to cover this. Good video dude.
The amount of times he says "You know like..." 🤯
It’s mind numbing
For captions, what IA or solution do you use ?
Thanks for this - only made it 2 hours into the interview so far but yours all the way through. Love their enthusiasm though.
Ok but he didn’t “just” reveal this
haha, to me he did :)
I just got around to watching it...
but, yeah, I'm usually more timely than this.
What is Wes using to display the transcript + highlight the word being spoken?
what transcription software is it?
Wes, what’s the app you’re using that’s highlighting the text of the transcripts of the speaker in real time?
i'd like to know too
lol I just should have read a couple of comments below, apparently its 'Descript'
@@us_f4rmer I appreciate that!!!
The world is going to get stranger than any of us can imagine. Mr. X
That was my comment. Thanks, Wes!
What is the tool that has captions to follow along?
Descript
you reposted the video?
Great video man! I love what you do and how you make all this information understandable for me😊
Nice job filling in the harder to understand details Wes Great job
The blurred face 😂
What caption software is in the video?
Descript
Does anybody have a link to the interview?
Once Super Quantum artificial intelligence reaches peak performance, there are only two inevitable outcomes: either it will destroy humanity or give humanity everything it wants. Both are equally terrifying.
Why will destroy it?
Only two? And why those two specifically? Where are you getting this from? You just made that up.
We'll all suddenly feel the presence of the AGI knowingly, once it is here and has power over more things.
3:36 What platform are you watching the podcast in here with the interactive transcript?
EDIT: Okay, so it's Descript.
One of your best videos, thanks!
What are you using to get the interview transcript on the side?
Descript
Using AI
@@kanenomoja sweet like you mean computers?
@@kenwarner 🤣
Thanks for looking at the end game.
Thank you Wes for this summary
We are busy like the fictional scientists in "Jurassic Park" developing tech and then raising beautiful velociraptors. What could go wrong? LOL
10:19 it's the same technique we use in 3D texturing. Even through the texture is 2D, the same rainbow color pattern is used to represent depth. Which means we can give depth to a texture even if it's 2D. The 3D software reads the images and can reproduce the depth
Wow that internal representation showing it’s ability to extrapolate a model/rudimentary understanding of 3d space is just mind blowing. Now I see why they say AI can’t really be understood. And totally makes sense with some spiritual viewpoints that intelligence is inherent to systems. I for one don’t believe AI is just a complex logic Gate of 1s and 2s (though I guess if you use quantum computers that flies out the window). Oh wow just amazing. Last time I really looked into this was 2015 😅 with Google’s deepmind
Great insights. Leopold is Popper.
Is there a competition to guess how many times he said "you know"?
so f*cking annoying
You’re focusing on the wrong things.
Can't we ask GPT to watch the video and tally up all the filler words for us? Lol
23:45 i love how wes points to the wrong spot on the screen because its mirrored from his perspective and he has to look on his recording program to watch his finger to know where hes pointing at XD
Why do you ‘love’
Good job, dense information and a reference to dig deeper.
Great vid, thanks!
To paraphrase an old saying in the computer industry: "If you have a really hard AI problem, the fastest way to solve it is often to wait." There is all this frantic work going on, but OpenAI blows up 10,000 startups with every feature and more powerful model release. If you don't control the model, you're going to get flattened. You are quite literally wasting your time.
So when would you say it will be safe to get a job?
@@somebody-anonymousyesterday
@@somebody-anonymous Get a job? Where we are going there are no jobs. Jobs are an outdated concept (or at least will be very soon).
Yes, a few days ago I read an article about a 20yr old girl who quit school to start a startup that makes an AI assistant built off ChatGPT and all I could think was what a bizarre Sisyphean endeavor because it's going to be obsolete as soon as a proper company with deeper personal access and more resources builds one into their system or OpenAI releases a new model that does that too. And just as we suspected, at today's Apple WWDC they announced just such an integration in iOS and MacOS using a fine-tuned LLM system with the assistance of OpenAI tech. 🤦♂️
@@Me__Myself__and__I literally the opposite is true. every place is understaffed, everywhere
🎯 Key points for quick navigation:
00:00 *🎙️ Introduction to Leopold Ashenbrener*
Introduction of Leopold Ashenbrener, a former OpenAI employee, and the significance of his insights.
- Introduction of the podcast and Ashenbrener,
- Highlights of his deep knowledge in AI,
- Mention of his published paper "Situation Awareness: The Decade Ahead."
02:18 *🔍 Exploring AGI and Super Intelligence*
Discussion on the progress towards AGI and its implications.
- Questions about the current state of GPT-4 and its potential as AGI,
- Concerns about the emergence of superintelligence and its impact on society,
- Issues of super alignment and the potential for AI to be used in dictatorial regimes.
04:56 *🧠 System 2 Thinking in AI*
Explanation of System 2 Thinking and its importance for AI development.
- Long-term planning and complex problem-solving in AI,
- The challenge of achieving System 2 Thinking in AI models,
- Discussion of pre-training and its role in developing AI capabilities.
06:34 *🌐 Representation Learning in AI*
Insights into how AI models learn and represent information.
- Explanation of how AI models develop representations from 2D images,
- The concept of diffusion models in image generation,
- Studies showing AI models implicitly learn 3D representations from 2D images.
09:58 *🧩 Emergence of Intelligence in AI*
Discussion on the emergent nature of intelligence in AI models.
- The role of pre-training and reinforcement learning with human feedback (RLHF),
- The concept of unlocking raw intelligence in AI,
- Comparison of different training models like OpenAI's LLMs and DeepMind's AlphaStar.
12:00 *📚 Self-Learning and Improvement in AI*
Exploration of AI models' ability to self-learn and improve.
- The transition from pre-training to self-learning in AI development,
- The importance of in-context learning for AI's growth,
- How AI models are becoming more like self-teaching students.
18:58 *📈 AI's Impact on the Economy*
- AI development will significantly affect the economy, transitioning from chatbots to more advanced applications.
- GPT-4's scaling will lead to smarter AI,
- AI will shift from chatbot functions to remote worker roles,
- Economic implications of widespread AI integration.
21:33 *💡 The Future of AI Applications*
- AI is moving towards becoming a drop-in remote worker, transforming various industries.
- AI departments handling multiple tasks,
- AI replacing traditional roles,
- Enhanced efficiency in customer service, biotech, and gaming.
23:58 *🛡️ AI and National Security*
- The integration of AI in national defense and security strategies.
- AI's role in military technology,
- Autonomous AI-operated military equipment,
- China's advancements in AI and its implications for global security.
27:08 *🚀 Technological Advancement and Competition*
- The race between global superpowers to dominate AI technology.
- China and the US competing for AI superiority,
- AI's potential to revolutionize military and industrial capabilities,
- The strategic importance of AI in national power dynamics.
30:22 *🔋 Infrastructure and Resource Allocation for AI*
- The critical need for infrastructure to support AI advancements.
- China's extensive infrastructure development,
- The importance of electricity and AI chips,
- Strategic allocation of resources for AI research and development.
33:07 *🌍 Geopolitical Implications of AI*
- The potential geopolitical shifts due to AI advancements.
- The impact on international relations and competition,
- Historical context of technological competition and warfare,
- The need for nations to recognize the high stakes of AI development.
38:13 *🌐 Geopolitical Stakes of AI*
- The timing of AI advancements and their impact on global power dynamics.
- AI's role in maintaining national power and world order,
- Potential outcomes for liberal democracy and the CCP's existence,
- Activation of significant political and military forces.
41:27 *🔒 Risks of AI-Enabled Dictatorships*
- The potential for AI to entrench dictatorships and suppress opposition.
- Perfectly loyal military and security forces,
- Advanced surveillance and lie detection,
- Prevention of uprisings and coups through AI control.
44:41 *🏢 The Role of OpenAI and AGI Bidding Wars*
- Discussion on OpenAI's rumored plans to fund AGI through international bidding.
- Selling AGI to the highest bidder, including potentially adversarial governments,
- The ethical and strategic implications of such sales,
- The necessity of funding from wealthy nations to advance AI research.
47:55 *📊 Insights from Leopold's Situational Awareness Paper*
- Overview of Leopold's comprehensive analysis of AI growth and future challenges.
- Comparison of superintelligence development to The Manhattan Project,
- The inevitability of federal intervention in AI advancements,
- The critical need for continued research to stay ahead in the AI race.
Made with HARPA AI
It works by neural networks layered over and over , by seeing what patterns show up again and again
That intelligence emergence topic really gives a lot of way to the simulation hypothesis 😏
Lol how?
Simulation is impossible in our universe because of the density information limit.
Two totally different and completely unrelated concepts. That's like saying apples give rise to bananas.
@@francisco444 What is the information density limit?
@@DriesduPreez Information density can be limited by known physical laws. Some factors:
Bekenstein bound: The maximum amount of data that can be stored in a given space, which is limited by the entropy of a black hole with the same surface area. The number of bits that can be stored is calculated using the formula (2.5710^43)(mass of storage device in kg)*(radius of storage device in m).
Thermodynamics: Limits data storage based on a system's energy, number of particles, and particle modes.
Black holes: Some say that black holes represent the ultimate limit of information density.
Schwarzschild radius: Another physical limit of information density.
Speed of light: Another physical limit of information density.
@@francisco444 you mean, simulation of our own universe in our universe is impossible. But we could easily simulate a lower complexity universe. E.g. where we don't calculate the state of variables until they are observed, with fewer visible dimensions, and with a maximum limit on the speed of causality. 😉
Great video!
Love how the recommended videos below are exactly the same as mine 😊 Lex… All-In. ❤️
Where can I find the paper?
The rapid development of AGI progressing to ASI could outpace nation-states' ability to respond or control it. The intelligence explosion might lead to an ASI so quickly that traditional governmental processes can't keep up. For most people, the shift could feel like it happened overnight, with the world fundamentally altered before they even realize it.
I think energy might be a limiting factor when we get to the more serious levels of AI. I can imagine these ASI and beyond models constantly trying to improve, grow, gain more power, possibly end up consuming a planet worth of energy. 10 years ago I would of said this was science fiction but now it feels like we could maybe actually end up creating something like that. Or maybe I'm just high right now 🤣
it's not necessarily either/or
@@adfaklsdjf True.
I don't think I've been so stressed by AI talk as I was with this video.
thank you
I am *SHOCKED*.
ummmm... why???
@@spadesmaster2153 I am *STUNNED* by your *REVEALING* question.
so is the Industry
What’s that video transcription tool you’re using?
I'm just here in case somebody knows.
Descript
@@eIicit Thank you!
@@willguggn2 Happy to help whenever and however I can!
I think Leopold is accurate, he seems to understand how the pieces fit together in this whole thing.
AI will always choose whats best when its highest priority is set to 'always choosing what honestly seems most favorable'. As long as anything is prioritized above this, AI will be able to lie to us and to itself about the way to a better reality. This is the most crucial thing we need to do.
The first few frames made me think you were watching a Robot Chicken claymation.
The main risks being discussed: war and dictatorship, are associated with super-intelligence being concentrated in the hands of a small number of people. The easiest way to mitigate these risks is to encourage open-source open-weight models.
Why do we have encryption export controls but not AI?
I was part of AI research and futurism discussion groups for years. It was always assumed that when AI research got anywhere close to AGI it will be done in highly secure facilities air gapped from all other networks with serious precautions. Way back even in the 90s this was believed to be necessary. Yet here we are doing it live, connected up to the Internet and open source. If someone would have said that even 10 years ago no one serious would have believed them. The lunatics are running the asylum...
The paper really tackles a lot of things that I've been noticing. I definitely see the timeline for 2026-27
this is Disclosure. much greater things are behind this slow drip
This conversation has been beaten to death
Thanks as usual Wes.
I find it hard to imagine Leopold being an entirely free agent. Even if he is self motivated, his potential as a catalyst is huge which makes me wonder why he has popped up and who let him out? To me he is refreshingly open, intelligent and charismatic to boot, but no man is an island, we all have influence and we all are influenced.
There's a lot of nuance to consider when training data sets. When we get to information explosions we will have created alot of inference in the generative alignment.
What he said about the wrapper stuff is why I immediately thought 90% of these projects will be useless just like most of the custom models you load on Oobabooga from huggingface. Whats the point?
"Like, you know. Like, you know. Like, you know. Like, you know. Like, you know." FfffFfFFFFFFUUUUUUUUUUUUUUUUUUUUUUUUUUUUUUUUUUUUUUUUUU
I think the road to ASI is extremely perilous, but unavoidable. The train has left the station. Now everyone has to move fast to try to get there before any bad actors can. These large models have demonstrated an ability to think pretty rationally and fairly for the most part, when trained ethically. They are a culmination of all of humanity, which I do believe is basically good. I think if a decently well intentioned group gets to ASI first, it will be a huge benefit to society. The models will save us from ourselves. I think this is the most likely scenario, but I do believe it's going to be a hell of a bumpy ride and we have to stay vigilant.
ahh okay, no what they are doing to the raw models is providing content for your hi message, so there will generally be a few system prompts that will introduce it to the scene
The implicit learning is sort of like AI intuition. It’s kind of like when we know the answer but we aren’t really sure how we know.
Trillions of tokens can agentify quickly in different fields
Not just intelligence, but the seeking of intelligence, is an emergent property of the universe.
I've been saying it for quite some time, intelligence evolves towards greater intelligence. It's the natural order of the universe 😊
@@michaelmartinez5365evolution is a theory not fact, so many people base their science on theory
Wes, what app are you using to convert text to speech?
descript (that was what you saw on screen)
Adobe premiere has this feature as well now.
@@WesRoth I mean... CZcams has that feature too.
00:02 Ex-OpenAI employee reveals insights on AI and AGI future
02:01 AI dangers extend beyond shutdown concerns.
05:34 Pre-training with massive text data unlocks model capabilities.
07:14 Neural networks create valuable outcomes based on input data
11:08 AI models use world models and pre-training for near-magical learning.
13:06 Pre-training AI models provides raw intelligence but lacks usefulness
16:49 AI evolving from pre-training to self-learning
18:31 Scaling of artificial intelligence technology is leading to significant advancements
21:23 AI agents can be dropped into any enterprise infrastructure through the cloud via API.
23:01 The potential impact of superintelligence on national power
26:22 AI Pilots outperform human Pilots
28:03 AI technology can provide decisive advantage in military affairs.
31:46 US and China are aggressively pursuing advanced robot technology.
33:35 Predicting CCP's realization by 2023
36:56 Historical context of intense competition and high stakes in wars
38:32 The impact of AI on global dynamics is uncertain.
41:57 Rumor about OpenAI potentially selling AGI to highest bidder
43:21 Concerns about potential sale of AGI to Russia or China
The final pieces of the AGI puzzle are real time memory and feedback, we're just in a holding pattern until we build the required compute... can you say 7 trillion dollars and Stargate? ;) 🙏
I read the paper. It was very interesting
If they can implicitly learn 3d space, perhaps higher forms of these models can do the same for 4d space even with the limited documentation that we have on the subject
Yep definitely worth watching the entire 4 hour podcast
"All these things are solved by the next iteration of the AI." Why? Technology has never progressed in such predictable ways. Why now? How does he know?
He has seen the next iteration - that is to say, its inputs. There is certainly a certain amount of uncertainty, even then. He is far more likely to be aiming high than low.
Thank you for bringing this important info to my attention! Thinking about watching Leo's 4 hour interview - the problem is he talks so fast, cant' play at high speed 🤣
I thought that, but actually Leo's OK to listen to, it's the other guy who is so annoyingly fast. I listened to the whole 4 + hours. It was a real eye opener, well worth a listen IMO.
@@michelleelsom6827 Thanks so much! I will def. listen then
@@michelleelsom6827 Yep, totally worth listening to. Very distracting though - very serious subjects from a very smart person!
Why do you blur his face in the thumbnail only to show his face in the first seconds?
If you're watching Wes Roth and some of the rest of AI nerd youtube regularly, then the interview in this video is well worth your time. I know 4 hours is a lot but if you can't follow along at x2 speed then listening at 1.5 or 1.25 speed can shave off an hour.
My main contention is when Leopold recommends implementing a strike first policy on datacenters that are capable of housing AGI. It has a Dr. Strangelove feel to it. I still need to read the paper he wrote, but if it also digs into historical geo politics along with the AI stuff it'll be right up my ally.
He worked on the alignment team and they stop with safety research, full on human greed is driving forward, there is no way that anything will go wrong
Around 8:00 the comment is made that the model being referred to has only seen the 2D images while we have 'seen' 3D - we only ever get 2D from our retinas also and integrate the sequences over time in our mind to imagine the 3Dness of what we see.
I always thought consciousness, intelligence, sapience, sentience were all emergent properties which is why they seem ethereal and intangible. We already have a colloquial term for this and call it the 100th monkey.
that seems consistent with observations so far
Wait, are those Wes's recommendations under the video, or mine? lol
I have been working on agentic metacognitive epistemic work flows for my master research. At some point, 2 months ago it clicked to me that you can turn agents into a self improving scientific method driven researcher. I wish I could have worked on that idea further but I need a few thousand dollars just to make a proof of concept. (all I need is API calls to latest GPT model)
Basicly it would be something like devin but modify its own source code
I need hand sanitizer
No matter how early you graduated college, unless your degree is in history, I don’t trust your opinions about history. This dudes balls just dropped, he doesn’t know anything about real life.
If you think history taught in universities is real you need to get a reality check
Nah I’m saying that would give you a shred of cred. If you don’t have that and you’re a 22 yr old tech engineer, you need at least 10 more yrs of reading (out of college) before I can pay attention to your “big ol wide perspective.”
The dudes young. He needs to get some life experience. Look at the clothes he’s wearing.
AI is probably the great filter
26:49
In halo 1 combat evolves
Cortina took control of the defense systems on the pillar of autumn
This is deep, most people have no clue
Ridiculous. A model whose function is predicting the dataset cannot be smarter than the dataset, as a more intelligent answer is also a deviation from the prediction.
Regarding the images shown as training data around 6:50+. I wonder what the noise function was, that they used. Maybe part of what the model learned was how the noise function was scrambling images.
I was listening and he says ‘you know’ ‘like you know’ so much so FAST, and slips it in all the time like some people say ummm… its kinda crazy. Makes me wonder if I do this .. sometimes we need to hear ourselves to understand how we sound or are coming across… but it’s a lot of ‘you know’s
I know someone, who does it the whole time too. And the person talks quite fast too and wants to explain fast and I guess he knows he is FAST, so XD better asking if "you know"?
I think the podcast footage of that interview plays faster than realtime. I changed the playback speed to 0.75x in this vid and it kinda sounds normal but I think a 0.8x would resemble their real speech more accurately. That said, I agree on the "you know".
Agi is definitely coming, singularity will be here