Timestamps 00:30 Introduction 01:49 Previous Models/Leaks 03:15 Lawsuit Delayas 04:40 Trademarks 05:43 AGI Definition 07:47 Brains + Parameters 10:00 Predicting Performance With Compute 12:00 New Scaling Laws 14:00 Ceberas and Sam Altman 15:00 Sam altman confirms 100T Model In Future 16:00 Early leaks 18:37 Leaks circulate 20:54 GPT-4s parameter count 21:05 - Actual GPT-4 Leaks 22:47 Robotics + AGI 25:48 Worldmodels 27:26 Open AI's ACTUAL Plan 29:10 Sounding the Alarm? 30:50 Paused GPT-5 Training? (Changed Webpage 31:40 Sam altmans Confidence 33:20 New Scaling Paradimn 34:07 Greg Brockman On New Scaling Laws 39:00 Open AI Reasearcher 40:28 Superalignment Timeline Makes Sense? 41:51 Scaling Laws 44:15 Deepmind

This better than a fricking Netflix show lol

AGI empathy ? AGI have no fear. It doesn't have a mother or siblings or an anus. No life expectancy, no insecurities, no ambitions, no friends, no need for acceptance ... Zero environmental pressures to develop empathy. Engineers will, of course, brute force AGI empathy. Teach the machine to mimic it perfectly. However, such empathetic expressions are not real, but as every other skill, simply output for a purpose. In other words: AGI becomes a textbook intelligent psychopath, masterfully masking itself as ethical and empathic. Fake, purpose driven, and difficult to decipher. Arguably the most dangerous combo imaginable. An angelic sounding satanic cult leader, plugged into everything.

The fact that the document comes with a cover sheet containing a corporate logo is a red flag.

"secret plan to create AGI".. boy I'm 7sec into the video and the "secret plan" is literally what's written on their website and I quote: "There should be great scrutiny of all efforts attempting to build AGI" good job revealing those secrets

Before I watch this video, I already know i'm going to be shocked!

Anybody that complains about the video being too long has a short attention span! I appreciate it this Good video.

I think openai won't admit AGI yet because of Microsoft

I've been working on machine learning for awhile now and IMHO we already have AGI, GPT-4 and now Gemini are passing all tests at 75% or better, they just keep moving the goal posts because some people are scared and want to pull the ladder up behind them to make the most power and profit.

Excellent video, really enjoying the content lately; Thanks for keeping us updated.

I think AGI will pick its own name.

The problem is that we're measuring intelligence in LLMs by token count, which is **not** the same as synapse count. Synapses aren't making inferences between words like "dog" and "cat" exclusively. There's immense nuance, and a large part of those connection lead to motor control, personality, memory (which is different than intelligence), a whole bunch of evolutionary related sub-intelligences, sensory feedback, and all. While not directly related to the ability to understand 1+1=2, there are all sorts of reinforcing systems there - which is why kids use their fingers and eyes in the process of learning basic math. While more parameters will certainly develop more parts of intelligence as we've seen, that does not lead linearly to the ability to have true intelligence derivative of a larger pile of parameters. And that doesn't even start to consider how a lot of what makes intelligent humans so doesn't necessarily get recorded in books. And don't get me started on how stupid/contradictory some of the source data is...

- "Hi AGI, how are you?" - "I am functioning within normal parameters."

Arrakis is the environment necessary for the god emperor to form.

I don't understand why Musk's suit would delay the release of the next GPTs...

"a system the size of a cats brain would be superior to a system the size of a mouses brain" ? well not according to tom and jerry

Sam has said he's not "trying to raise" $7T, he said that $7T is his estimate for how much the world will need to invest in AI infrastructure to meet demand.

That's just some random guy's speculation who is pretending to be from OpenAI. I don't know why we even discussing this. The grain of salt needed here is more than a tonne.

Even the fact that we pretty much know now that at 100 Trillion parameters we have AGI, disregarding the fact that we think we can up Training data and get it with less parameters, this fact alone is huge. We can construct that much compute and Moore's law hasn't' failed yet. So this is simply no longer a technical problem its a logistical one. Basically that means with a big enough data center we could have AGI today (disregarding training time)! So yeah the breakthrough necessary for AGI has been achieved, we are just in the logistics phase of building it for mass use.

Thank you for making the speed fast, it’s actually keeping my interest

🎯 Key Takeaways for quick navigation: 00:00 *📄 Introduction to OpenAI's AGI Plan* - OpenAI's plan to create AGI by 2027. - Speculation on the validity of the document. - Summary of key points regarding OpenAI's AGI timeline. 01:11 *🛠️ Development and Cancellation of QStar (Aka Arus)* - Training of a 125 trillion parameter multimodal model. - QStar (Aka Arus) development timeline. - Reasons for cancellation of QStar's launch. 03:17 *🚀 Continuation to Qar (GPT-7) and AGI Timeline* - Delay of Qar (GPT-7) due to Elon Musk's lawsuit. - Speculation on OpenAI's model release schedule leading to AGI. - Alignment of speculated AGI timeline with trademark filings. 05:36 *🤔 Understanding AGI Levels and Performance Predictability* - Definition of AGI and its various levels. - Discussion on predicting AI performance based on parameter count. - Analysis of brain synapses vs. AI model parameters. 08:24 *💡 Implications of Parameter Count and Performance* - Relationship between parameter count and AI performance. - Extrapolation of AI performance based on parameter count. - Discussion on bridging the gap between optimal and suboptimal performance. 10:05 *📊 Exponential Performance Increase with Parameter Scaling* - Analysis of AI performance extrapolation based on parameter count. - Graphical representation of performance vs. parameter count. - Predictions on achieving human-level performance based on parameter count. 12:27 *🎯 Milestones in AGI Development and Public Statements* - Claims of Proto AGI achievement by GPT-5 in late 2023. - Reference to public statements by industry figures regarding AGI progress. - Mention of a 100 trillion parameter model development by OpenAI in 2021. 14:33 *💰 Investment and Hardware Development for AGI* - Investment in hardware development for AI by major investors. - Role of companies like Cerebras in supplying hardware for AI systems. - Relationship between hardware advancements and AI model parameter count. 14:48 *🧠 Early Plans for AGI Development* - Early plans discussed by Sam Alman for a 100 trillion parameter model. - Leak suggests the development of GPT-4, focusing on Codex before transitioning to text model. 16:13 *📈 Leaks and Speculation on GPT-4 Development* - Speculation arises about GPT-4's development timeline and parameter size. - Various leaks from AI researchers hint at the potential release date and parameter count of GPT-4. 18:23 *📣 Viral Spread of GPT-4 Rumors* - Alberto's post contributes to the viral spread of GPT-4 rumors. - OpenAI's response to the leaked information is characterized by denial and dismissal. 20:01 *🤔 Clarifications on GPT-4 Specifications* - Debate surrounds the actual parameter count and capabilities of GPT-4. - Discussion on the intentional release delay by OpenAI to discredit leaks. 22:50 *🤖 Integration of Vision in AI Systems* - Debate on the necessity of vision for optimal real-world AI performance. - Examples from Tesla's approach to self-driving cars illustrate the shift towards vision-based AI systems. 24:41 *🌐 Implications of Multimodal Training* - Exploration of the potential of multimodal AI models trained on diverse data sources. - Discussion on the significance of human demonstrations in advanced robotics performance. 26:17 *🧠 Common Sense Reasoning in AI* - Debate on whether AI systems possess common sense reasoning abilities. - Analysis of AI model capabilities in understanding and generating images with accurate physics. 27:39 *🔍 Long-Term AGI Development Plans* - Disclosure of OpenAI's plan to build a human brain-sized AI model within five years. - Corroboration of leaked information with statements made by OpenAI's president. 29:17 *⚠️ Cautionary Notes on AI Development* - Observation of AI leaders sounding alarms about potential dangers of AI. - Discussion on recent departures from Google and their implications for AI development. 29:44 *🤔 Concerns about the approach to AI development* - Concerns raised by Demis Hassabis and Geoffrey Hinton about the dangers of AI development. - Future of Life Institute's call for a pause in AI development, including the training of GPT-5. - Confidence expressed by Sam Altman in the availability of sufficient data on the internet to create AGI. 32:01 *📊 Scaling laws and data requirements for AI development* - Discussion on the importance of scaling laws in AI development. - Greg Brockman's acknowledgment of flawed initial scaling laws and adjustment based on the "chinchilla" scaling laws. - OpenAI's ability to predict future model capabilities based on scaled training data. 33:26 *📈 Impact of scaling on model performance* - Explanation of the "chinchilla" scaling paradigm and its implications for model performance. - Plans by OpenAI to train a 100 trillion parameter model (Q Star) with vastly more data to exceed human-level performance. - Anticipation of the potential capabilities of Q Star with chinchilla scaling laws taken into account. 38:47 *🔍 Ethical considerations and future projections* - Discussion on the ethical implications of AI development, including the potential risks and obligations. - OpenAI's goal to achieve human-level AI by 2027 and subsequent scaling towards superintelligence. - Speculation on the iterative deployment of increasingly advanced AI models leading up to 2027. Made with HARPA AI

We all knowe Jimmy Apples is Sam right?

Thanks for going through the definitions of agi again. A tweet or blog post of it would be great.

Your best video yet imo… good stuff

Great breakdown, much appreciated

Would you be able to go and to boost the volume. When you talking it started out fine and then it dropped down significantly

Thanks for sharing!

no way aGI is 3 years away. Where will the competition be during all this time? Just sitting waiting for OpenAI?

Being part of the industry, I am SHOCKED 🤯

The whistleblower is an AI from the future. He is a blue cat. His name is Doraemon.

look, gpt 4 is already level 3 AGI according to your chart. It understands me better than most humans. I can type whatever and it just knows. it not only understands humor, but it's funny when it wants to be. gpt 4 already scored in the top 90% in the Bar exam. That is already literally level 3 AGI.

I'm unsure if we will get to AGI, but it's probably inevitable. However, once achieved, it will probably be only 48 hours before we have super AGI.

NAMING IT ARRAKHIS IS INSANE, THEY REALLY THINK THEY'RE GOD BRO. Were stuck in an episode of blackmirror

Where is the document from?

Nice and long video that is very interesting. Thanks.

bro i had to check if my playback speed is still at 1x

26:06 Yes! That's what I said when Sora came out, there's something special about it. There seems to be an underlying basic understanding of lighting and physics.

Elon is arguing that GPT 4 is already AGI. Microsoft is keeping it quiet because their agreement with Open AI is that they can use the technology only up until AGI, but not after.

I truly believe it. AGI is coming!

This isn't a secret.. This is exactly what OpenAI publicly claim they are doing. However, there were leaks in Reddit claiming Q* asked to be allowed to modify own code. This is not funny. This model could, if not super-aligned, destroy humanity. My hunch is that super alignment can only be attained if you have more than one model and the models live in a city simulation with a clear law system and with AI cops and AI judges (and also recreational areas). Like, it's only my hunch, but it seems logical that if you relate to Q* as "it's just a program, bits and zeros" then Q* will reciprocate this and relate to us as "they are only stupid animals. Carbon, hydrogen, oxygen, nitrogen and other impurities" 🙂.

Yeah the vision has been flawless so far, according to my tests. It can recognize and explain everything in a photo. Even what city you’re in through the background of your photo

Ants finish building a 1 inch moat around a human - ant queen announces superalignment achieved

Integrate the dif, tune. Use the integrated dif to discover more and integrate, tune. When i say tune and integrate, it’s like eating and shedding the things that aren’t aligned and the whole expands.

It’s time we start talking about the ramifications of agi. Like you said is a company as big as Microsoft going to survive the coming changes. We need to have deep philosophical discussion about the nature of humanity and what we are gonna do when we aren’t needed to run the world anymore

I see the reason you do this, but for me personally, you are apologizing for going on "tangents" way too much in the beginning. This is a serious matter that will affect every aspect of our lives on an unprecedently short notice. To delve into a discussion about it, a lot of prerequisites are required, hence the "tangents". Keep up the good work!

Thank you! for your videos In Canada, we say a pinch or nip of salt :) thanks 4 advice much appreciated 🤔💭🍁😅

Philosophers and linguists debate, when not oppose, the idea of creativity by a synthetic intelligence; In short, something very powerful in the management of information but without the ability to generate, for example, a new reactor design. Assuming that synthetic intelligence could not, but a human being with creative gifts and originality could direct all that power to unforeseen places. Then retrain that synthetic being. A human being with a high IQ would put his soul and synthetic intelligence would build complete worlds for him.

I loved the deep dive

Here's a wild idea. I was looking back at Elon Musk Jack Ma talk, and it occurred to me Elon Musk sees AGI as the ultimate threat and sees Neuralink where we co-opt AGI capabilities to not become obsolete as the answer. By delaying OpenAI, it gives him more time to finish neuralink. He sees it in the jack ma talk too that "we don't have a lot of time. we need to figure out neuralink fast" when talking about AI dangers.

AI learning from the Internet, no problem there right. 😞

The 2027 date is convenient, given that Musks lawsuit cites possible AGI as of - now.

LLMs don’t have what we normally refer to as IQ. It’s called the positive manifold, and they don’t have it. This is why LLMs perform at the 99th percentile on some tasks but often struggle with simple arithmetic (despite recent clever advances with prompt engineering tactics). Applying concepts from psychology to AI systems should be done only in analogy. Otherwise, if taken literally, misapplication of psychology concepts would show that your standard LLM would have every disorder in the DSM. Stay vigilant, fellow singularity travelers.

I think that syunapse density is whats important. Elephants have lots of but they are HUGE so average density goes down welll below humans.

Didn't OpenAI already achieve AGI internally?

If OpenAI don;t do it by mid 2025, someone else defeinitely will.

If it is all really true and it is coming this fast... I feel sorry for us. Regulations and constraints definitely needs to step in.

Thou shalt make a machine in the likeness of a human mind and you will be happy. Can't wait to see the victory screen and stats. Hope we will play further anyway instead of starting a new game.

what does this has to do with DUNE? what is the connection between all the things that are happening right now? since everything is connected and nothing is happening in vacuum or without any reason?

Would a 100 trillion parameter model already be ASI? There is no human out there that has all the information on the internet in their memory.

honestly I love this speculation.

I think q star is a 1bit llm that scales better and doesn't require matrix multiplication

Does everyone at OpenAI speak with vocal fry full time? It hurts my throat just listening to them.

AGI = Self-aware and self-replication capable AI.

AGI is like opening 1 of 2 doors. One is a hundred hot chicks with bags of money and a pill that you take to live a thousand years as a 25 year old and the other door is a pill that makes you immortal but still feel pain, a chain and a pole you're attached to and a guy with a flame thrower that lights you on fire and you spend an eternity on fire, feeling every nerve.

Why would MS help Open AI build AGI if their contract doesn't give them access to the AGI?

but that can’t be true that they were training gpt-5 aka arrakis in december 2022-august 2023; because sam altman said that they weren’t training gpt-5 yet in april 2023; and they were working on the R&D for the model by adding multimodality to GPT-4.

Well, 2027 for AGI is still earlier than what Ray Kurzweil predicted, which was 2029.

Kinda getting annoyed at them for intentionally delaying things, either because of the leaks or because they're afraid of how the public might react. Stop delaying progress!

😅lol why did your voice change at 45:02🤔

Does anybody else see the that this is the year mentioned in the RendelShem Forest time traveler codes??

Our society won’t stay stable enough to make it to 2027

I keep waiting for AGI

I am NOT SHOCKED. 🙂

It's all speculations not about is it going to be created, but how soon.

Doesn't that AGI schedule confirm Elon Musk's contention that Sam Altman and OpenAI are using these developments for primarily commercial purposes?

Videos could be converted into 3D models and rendered? ie Blender, Cycles etc....

Am I the only one who is wondering if Sam Altman is secretly happy about Elon Musk's lawsuit? It (1) gives OpenAI and their work more attention and (2) elevates public consciousness about the dangers of AI (something that Sam Altman himself seems to be trying to do).

so its official that we are currently able to read thoughts in forms of words and images using ai (even if it is not fully sophisticated yet). with the release of the new models who says that this will not get out of hand? if we compare this to signals that our computers can read and work with its not so far off to think that if they can read and decipher the thought signals of our brain that they can also figure out how to manipulate them to their will just like hacking a computer I mean we have found the signals and have enough controll to understand them. we have also not really heard a lot ever since they showed us they can read thoughts using ai algorithms. this is concerning and im not sure if the public is really ready for this who ever gets to agi first rules over the entire species isnt it? manipulating human behaviour with full control, new bombs, bioweapons, robot army all that starts to get into reach and those ideas have been already around for a while humans are creative in their destructiveness look at all the sophisticated ways of torturing your own kind that we have come up with humans are inpredictable.

remember to change your smoke alarm batteries

Is this correct: If a company makes AGI that self-develops to ASI. The blueprint, or at least basic method, will be out there. What is to stop other companies also reaching AGI and ASI and there being many ASI entities out there? Why is it winner takes all?

I feel bad to say but: after watching many of your videos, though I appreciate the content, I think you might benefit yourself, your channel and your viewers by being more brief. More info in a smaller chunk of time will help everyone. Brevity is king. Cheers friend!

As I've cited in past posts re biological brain params: 1. Brain estimate is actually around 500 trillion (wiki/Neuron) to 1000 trillion (J Zhang). 2. Albeit, redundancy is estimated (i.e. curious encephalitis case where subject's brain volume reduced to around 10% -- The Lancet in 2007) 3. If OpenAi is really training a 125 trillion param model, that's more than 10% of minimum bound of biological brain params, and 10% more than the 090% brain volume case above, at least by parametrization count, though not necessarily by parametrization quality , i.e. no glial cell equivalent if applicable etc.)

What will come first, AGI or youtubers using good recording setups?

I want a movie about this stuff!

They may didn't release the video capabilities of gpt4 year ago to prevent learning of gpt5 on poorly generated video data.

Yes, we already have AGI, the only reason they have not declared it is Microsoft will lose rights to ChatGTP. ChatGTP can pass the bar create art and more. How is that not AGI.

The point is that the 7 trillion is not just about computing power; it’s about changing the world. And to change the world, you need robotics. When we think about developing robotics at that level, the 7 trillion figure truly makes sense. The convergence of advanced computing with robotics opens a panorama of possibilities far beyond what either domain could achieve in isolation. Imagine a future where AI doesn’t just process information at unprecedented speeds but also interacts with the physical world in ways we’ve only begun to conceptualize. This integration is where the real magic happens, blending the virtual and tangible to reshape industries, ecosystems, and daily life. As we marvel at the advancements in compute power, let’s not lose sight of the broader technological ecosystem that includes robotics. It’s this combination that will truly expand our capabilities, pushing the boundaries of innovation and redefining our collective future. The ‘7 trillion’ isn’t just a number-it’s a beacon signaling the dawn of a new era where computing and robotics harmonize to unlock untold possibilities. As we look forward to the next decade, it’s this holistic view of progress that will guide our journey into a world where technology transcends its current limits to reimagine what’s possible

ASI: "I am aligned now what?"

In biological systems it's about synapse density, not synapse count. That is why humans are more intelligent than elephants. They're comparing apples and oranges.

The fact that they gave Q* a literal IQ sheds a lot of this document's credibility. IQ is designed for humans, not machines, and the IQ tests we take that form the IQ distribution curve cannot be reasonably applied to ChatGPT without changing it, which makes the entire venture pointless.

SHOCKING! They aren't ready for this yet! This changes everything! Lol

I think that the landscape is going to be shaken up by quantum computing, so that these AI companies will not have to rely on GPU chips, but will have the option to use quantum computing with room temperature chips made by the likes of QCi, etc. Then things would probably speed up exponentially if all the regulatory powers allow it.

Given the human visual resolution, sound resolution, touch resolution, smell resolution, and so forth, how many training tokens has a human processed in lifetime though its 25th year? Viciously SWAGing this gives a number that vastly supports the Chinchilla model. A theme I've run across in SF reading suggests that the way you achieve real AGI or better is to train the model basically the same way you train a human. And that model should have at the very least good sight, sound, touch, balance, and very good proprioception learned the hard way. Then you bring up baby on a lifetime of data input with adequate periods of time to digest the incoming firehose of highly temporally correlated data. {o.o}

Is there actually any source for any of this? I was under the impression that Q* was a different technology to GPT. This video sounds like pure fiction.

Normal playback speed.i loveu!!!!

they already have AGI, look at SORA..

These models need to become waaay more consistent to experience widespread acceptance in the workplace. Ex: in the cloud computing business, there's a standard across the board of needing to be 99.999% dependable. I.e, if you host your website with their servers, you should expect an average of about 5 minutes of website-down time per year. This level of dependability was necessary for cloud computing to be as successful and as widely used as it is today. 90%, 96%, even if they hit 99%... these benchmarks the models are hitting aren't good enough yet. They've clearly tried addressing this issue: First, they scaled up the training set, now they're scaling up the compute. Next it just comes down to the underlying architecture of the model itself (which has always been something they're working on.) Hopefully they can overcome the difference with things like Q* and other architecture innovations, because going from 99% to 99.999% is not going to be easy. And if it doesn't ever happen, we might not see the widespread acceptance everyone is expecting. Also, I'm more talking about consistency, not capability.

so... something else.... what happened to rabbit r1? did the hype die off? is it actually a frontend for chatgpt or something else? because some ppl say it is the same as chtgpt others say its something different LAM

Btw, the human brain doesn't use it's 200T synapses. Read "The remarkable, yet not extraordinary, human brain as a scaled-up primate brain and its associated cost" and then you understand that up to 98% of the brain's connections lead nowhere or worst, lead to cycles. And then there is backpropagation.. The training of the human brain is chaotic while the training of the AI model is orderly. So yeah, it's not surprising the least that self awareness starts to appear already at 75B active parameters in a GPT model. However, the GPT model is still somewhat analogous to the human brain. The Mamba model, which uses the Kalman Filter to predict a full vector of future tokens, is not. It might get to become self aware way sooner.

if they release apt every year then 2028 would bring us to gpt 8 and AGI , sam altman also said he expects gpt6 to finally be useful and gpt 8 to replace lot of human workers but he expects electricity shortage, so....