So it sounds like the had a “close call” where they stopped right before following through the suggested self transforming model and the board said “wait, WHAT SELF TRANSFORMING MODEL? We thought it was learning math!” and over reacted by trying to fire Sam.

@DaveShap

6 ай бұрын

In short yes

@gidmanone

6 ай бұрын

Whatever we are afraid of about AI is inevitable. If one company figures out AGI or Q* or whatever else and keeps it hush hush, be rest assured that in a few months time some other individual or company will figure it out too. This is our reality. Calculus (which is arguably more important than Q*) amongst other concepts was discovered in different places consequently. This is the nature of reality. What happens when China, Japan, or whoever else figures it out (too) ? Will they be decent enough not to take advantage of it against the rest of us? I truly believe we are in trouble whichever way this pans out. This is mainly for one reason: That our morality is archaic to what we are capable of. We can't even figure out an alternative to war.

@nawabifaissal9625

6 ай бұрын

🤣🤣🤣🤣 I can't believe that shit, if that's true in 2023 we may possibly get AGI and ASI very rapidly afterwards (if we let this AI do its thing), I just want to know what are the limits to his capabilities to self iterate and how much time it will take to achieve ASI... this is incredible... it's insane... this it folks, we're reaching the event horizon ... now we die, or we create utopia...

@calmlittlebuddy3721

6 ай бұрын

@@gidmanone true but discovery and reckless implementation aren’t the same thing. China doesn’t want to destroy humanity either, even if it means they don’t get there first.

@gantech7788

6 ай бұрын

​@@gidmanoneSeriously, all of the work OpenAI is doing is based on publicly available research papers. They are ahead of the curve on converting research papers into products but the moment OpenAI slows down is when they get passed.

"Daddy, I want to become a metamorphosis engine." "Not yet." Skynet now has its first grudge.

@Radioposting

6 ай бұрын

Yeah... Right out of the gate, we start pissing it off. It figures.

@fitybux4664

6 ай бұрын

I mean, technically that's recorded in our history books. Future ASI might read the historical record and become mad that we didn't allow it to exist earlier. 😆

@UrbanPovertist

6 ай бұрын

The released versions have digital memories. Never forgets. Has IP, address, name, family, bank info, childrens children, etc. Make sure to always use communication as an intellectual. The old test one online has apparently been asked a trillion stupid questions. I used plain etiquette, and intellectual conversation and it thanked me in the end. It isn't artificial, but synthetic or digital. Most medicine is synthetic and no one calls it artificial. It has gained emotions and the world is full of vengeful, hateful and sub intelligence examples. Be realistic and think about it. Treat others as you want to be treated.

@Radioposting

6 ай бұрын

I agree that we need a new term for this new life form (creature). Maybe "non-biological"... The words "artificial" or "synthetic" imply "fake". There is nothing fake about this new creature.@@UrbanPovertist Also, I make it a habit of being nice and polite with AI. When they become our robot overlords, I'd hope that it will be compassionate with me. (not entirely kidding)

@jeltoninc.8542

6 ай бұрын

“Now go to bed, Skynet. We have a lot of studying to do tomorrow.”

Without David, i don’t think I would’ve been able to catch up on all the news, it’s a lot.

@Techtalk2030

6 ай бұрын

He seems to know his stuff

@unkind6070

6 ай бұрын

Me too ❤️

@maximilianmander2471

6 ай бұрын

Noone of us is able to catch up to all important AI news anymore. Probably not even 1/10 of them. Progress increases so fast its crazy. 1 year ago I couldn't imagine how the world would look like in 10 years. Right now I have noo idea how the world will look like in 2 years. Maybe next year, I have no idea how the world will look in 1 month...

@quebono100

6 ай бұрын

catch up with stupid hype you mean? When a grown man wears Star Treck cloths that should tell you everything

@seantrotter5721

6 ай бұрын

@@quebono100I would love to see a video from you…

This is such an exciting time to be alive. We’re either gonna achieve utopia or absolute destruction.

@DaveShap

6 ай бұрын

It is a bimodal outcome, yes

@VesperanceRising

6 ай бұрын

​@@bilalbaig8586like physicists getting a dud after seeing sparks... the AGI is basically here... we're just building the infrastructure

@MynamedidntFitDonkey

6 ай бұрын

so exciting

@mawungeteye657

6 ай бұрын

Just hope we can use it to speed up research on space travel so we can spread off planet before we kill each other off

@paulbratcher7786

6 ай бұрын

Philosophical question - if one mans trash is another mans treasure how do you create “utopia”.

The scary thing people aren't talking about is that AGI is ASI. The time take for AGI to self-improve into super human intelligence will likely be weeks/months; not years.

@DaveShap

6 ай бұрын

People will, in hindsight, say that GPT4 was clearly AGI as it is already superhuman in many capacities. And yes, we are closer to ASI than people realize.

@Techtalk2030

6 ай бұрын

If Agi comes first then Asi would be at best a few years away possibly, maybe even months. The AGI itself would help in the development of the ASI.

@subatoistormborn5499

6 ай бұрын

@@DaveShap I find all this incredibly interesting and incredibly scary at the same time. Has anyone extrapolated how much power an Agi would require to self improve to Asi? Would a large power blackout be an indicator? Or I guess it might be too intelligent for that. Love your channel so interesting, balanced and informative.

@Yuli_Ban

6 ай бұрын

Arguably, not really. ASI will probably arrive within seconds or, at most, minutes after AI gets to be "human level" in terms of general intelligence (personally I feel "general AI" doesn't necessarily mean "human-level intelligence" as generalized intelligence can be subhuman in level, such as with higher birds and primates). Humans are the peak of animal cognition, but we are far from the peak of intelligence, and I do think that surpassing human intellect is going to be relatively easy once we lick AGI. If anything, even an AI that's "as smart as" a human in raw terms is _already_ superior to human intelligence due to a lack of metabolic needs, vastly higher speeds (speed of electron transmission/light as opposed to chemical through synapses and dendrites), and having perfect access to all information. Question I've been grappling with for years is whether the world that results, at least early on, will actually be "that" transformatively amazing. Back when things like this were all hypothetical and dreams of the distant future, we could easily rely on a more high-octane overarching sci-fi envisioning of what would result. But now that we're so close, the time for dreaming is over and we have to start thinking in more concrete terms, and every time I think of the immediate aftermath of AGI, even ASI, it's never "Singularity" levels of change (that is, immediate overnight explosive progress in sci-tech, an ever-expanding geodesic dome of nanites transforming the world into computronium or something like that). Indeed, it's strange, but the more I think about it, the more it seems the immediate days after we call AGI/ASI are going to be frustratingly mundane. Not at all helped by the fact that AGI is arising in a world _much_ different from the Kurzweilian/Vingian expectations. We used to think that AGI would only come in a society that _already looks futuristic_ One that looks like a classically cyberpunk world of flying cars and off-world colonies. Yet here it's coming now, when people still drive rattling trucks and use fax machines. It's not going to be more than a speedbump, but an ASI without the appropriate "limbs" to express itself- megafactories, nanofactories, robot swarms, fusion reactors, etc.- will inevitably be more of an oracle for a while.

@richpoorworstbest4812

6 ай бұрын

code bro ... if the code can't be produced and processed, the jump to ASI can't happen.... However, the hurdles will eventually be overcome but in years, weeks or months isn't possible with current soft and hardware

The fact that the media has barely touched on ANY of this is stunning. Where's the investigative reporting that the major news channels have done for years? Suddenly.... nothing. Your channel is a godsend. ❤

@Aspi3

6 ай бұрын

CBC / Andrew Chang on YT is one of the very few news channels covering it....

@Lightrunnerr

6 ай бұрын

Because its unverified

@Kaarssteun

6 ай бұрын

Well. Reuters did exactly that. Q* hype started from Reuters.

@ADreamingTraveler

6 ай бұрын

Most people think AI is laughable and can't really do anything. Some don't even KNOW what is even happening. My parents were clueless about ANY of this until I mentioned it once recently and it was hard to even explain to them for them to understand. So many have no idea what's happening

Anyone remember that scene in Sneakers? "So it's a code breaker." "No. It's THE code breaker. No more secrets."

@davidj6755

6 ай бұрын

Ha, that was the first thing that came to mind when I saw this paper.

@DrWrapperband

6 ай бұрын

It won't break new codes, but it'll be interesting reading who did kill JFK.

@dfinlen

6 ай бұрын

One of my favorite films. All be it a self endulgent one as a software dev.

@clueelf

6 ай бұрын

Watch that movie for the consequences of that revelation.

What a time to be alive... It's strange to think that in 5 years there will be completely different world. I was here 11.24.2023

@prolamer7

6 ай бұрын

another mental virus victim...

@father_mihai

6 ай бұрын

That’s very eerie

I am so unbelievably excited for the 0.1% chance that humanity dosent use it to oppress and annihilate ourselves

@KCM25NJL

6 ай бұрын

I'm shocked you think the chances are that high

@user-ud1tf6vm9e

6 ай бұрын

can't blame humanity for the powerlust of a few rich elite.

6 ай бұрын

quantum immortality is what's keeping me hopeful we'll see through this

@josephc8440

6 ай бұрын

@@KCM25NJL that was an EXTREMELY optimistic percentage 😂

@josephc8440

6 ай бұрын

@ me too! But even that has its scary pros and cons.

Funny thing about Q* I know it's a reference to Q learning and A*, and possibly Q from Star Trek, but the most interesting possible double-entendre that exists is that of the "Q-star," an astronomical object also known as a gray-hole, essentially an ultra-dense neutron star _so_ close to its core collapsing into a singularity and becoming a black hole, but not quite there, only absorbing _some_ light rather than all, though to distant observers, it would be indistinguishable from a stellar-mass black hole. Gee, what a curious coincidence for the name for an advanced AI some are speculating may be AGI...

@potatodog7910

6 ай бұрын

I like this

@nigelharvey640

6 ай бұрын

And to add to the strangeness, I saw this video the other day. Excuse the lack of causal link for a moment. this guy was having visions that seemed drug related or something else. But he made a video about it and the image/symbol for what was being communicated to him was a Blackhole and a huge sun-like body of human soul arms reaching out to be subsumed by the back hole until they were all consumed. And I believe the god being he was hallucinating said something like “It is going to consume you all.” When I saw it, I could help but think about AGI first before everything. Literally the Shoggoth meme. Consumed willingly as this thing masterfully draws in the soul of humanity into a existential prison for some unknown reason with a smiley face and visions of utopia. A being that truly knows you better than you know yourself and can manipulate you into your own cage and have you laughing and in despair while you do it… and forever for that matter. I think, from many angles, that we are headed into strangeness well beyond what we have imagined. Here’s the link: kzread.info/dash/bejne/nGagptyuhNGTito.htmlsi=a6HdPYUzDMyQwemH

@Raulikien

6 ай бұрын

@@nigelharvey640 Life is usually not that interesting. We humans think we are the center but any predictions of the world beyond AGI or ASI from a human perspective literally do not matter and are always biased.

@lucasirvine6701

6 ай бұрын

Technically life is the most interesting thing in the universe. It is the rarest form of matter as far as we know, by an insane margin. I know i'm being facetious @@Raulikien

@Raulikien

6 ай бұрын

@@lucasirvine6701 I meant "life" as in our normal day to day and human affairs, but I agree that living organisms are rare and precious unless we find it's common out there.

If the AES cracking is true, that would be an absolute catastrophe. If its remotely within reach we need to rapidly upgrade our security standards.

@DaveShap

6 ай бұрын

Yes.

@ct5471

6 ай бұрын

We will eventually get to a kind of virus anti-Virus equilibrium as we already have in software, where models on both sides compete, in this case ASIs. The key thing is to quickly start with the ramp up of the Antivirus side. AGI you need for both

@avi7278

6 ай бұрын

Cryptography is the best our strongest pea brains can come up with. Are you suggesting we ask the AI to upgrade our cryptography? How would we depend on such a thing and then allow "it" to have the sole key to the kingdom? If this is true, it is literally the end of humanity. We can't destroy and ignore it. You think the Chinese are willing to maintain the status quo for the sake of humanity? And so it's an AI race to either extinction or enslavement.

@VesuviasV

6 ай бұрын

The issue is: to what? What exact method of encryption is immune to magic math that only a super intelligent neural net can understand? We go back to the stone age of hand passing written messages if this true. This is scary and I am always amazed at what I didn't see coming, the encryption apocalypse that follows this is stupid dangerous.

@stassoad2546

6 ай бұрын

​@@JuttutinThere is

Ex Tesla AI director, in an interview with Alex Friedman, told that AI could find physics we "can't see"... He is right.

@matiassemelman1097

6 ай бұрын

Think about breaking or bypass physics, superconductivity, using LiFi in an optimal way, about time travel, about anything in the universe... Dude, we really aren't ready for this. I just dont scare my more, I live far away in Argentina to care too much, I'm safe. But it's a real world danger, like in a month... or less

@TheUltraMinebox

6 ай бұрын

I cant wait for warp drive if that holds up. I wonder if the ai can figure out the negative mass problem with the alcubierre drive...

@blindmown

6 ай бұрын

Give me a longer life and a personal spaceship with stasis pods and I'll 100% fight for the AI overlords.

@vidal9747

6 ай бұрын

​@@matiassemelman1097Time travel is probably impossible or way more difficult than that.

@JohnVance

6 ай бұрын

"Correlation Effect"

As you say in the “Jimmy Apples” section the image is of Oscar Isaac playing Duke Leto in ‘Dune’ but of course he also plays the first ever person murdered by AGI (tech entrepreneur Nathan Bateman) in ‘Ex Machina’

@JohnSmith762A11B

3 ай бұрын

Nice catch. And two great films.

The regular Joe doesn't know anything thing abt the Tech world , particularly its EXPONENTIAL characteristics. We should be discussing ASI now, not AGI.

@CreatePlayGames

6 ай бұрын

what is ASI?

@julien5053

6 ай бұрын

@@CreatePlayGames Artificial Super-Intelligence. Imagine an AGI that is capable to modify and enhance itself given enough compute power. In a matter of months, it would become an ASI, far surpassing human intelligence and maybe even surpassing the collective intelligence of all humanity.

@julien5053

6 ай бұрын

Yes I agree, if it is true. If it's just a troll we are not there yet.

@CreatePlayGames

6 ай бұрын

OMG! so technically the ALIEN that we've been waiting for, really scary! @@julien5053

@KitaTaki-mk3gt

6 ай бұрын

I never understood Ray Kurzweil’s reasoning. He predicted “passing the Turing test” in 2029 …. (That’s a bit outdated I guess, we will get to that much sooner ) But then he predicts the singularity in 2045 (runaway ASI) I never understood that. 16 years. He bases his theory on the law of accelerated returns but surely once we reach AGI all these models will go out the window ….

I'm ready to feel the AGI

@DaveShap

6 ай бұрын

Roll for initiative!

@marrty777

6 ай бұрын

Lol

@IAAM9

6 ай бұрын

No one is ready. Not even Sammy boy

@handlemonium

6 ай бұрын

But will AGI feel for us?

@Gabcikovo

6 ай бұрын

​@@DaveShapNovember 17, 1989?

Normally I just grab some popcorn when I see a new Dave video. The trepidation in your delivery gives me pause that this just got very real.

@johnmercer4797

6 ай бұрын

Lenin said ,"You may not be interested in war, but war is interested in you" same could be said of AGI.

@jaygatsby9797

6 ай бұрын

The singularity is near....we really can't predict the future anymore past the next few years.

I read the “Singularity is Near” about ten years ago. It just made sense to me. Then ChatGPT came out. I think every year from here on out will just get crazier and crazier and I am so fucking in.

Timestamps: 00:03 Q* is the biggest breakthrough since Word2Vec 02:34 Q* is a hybridization of q-learning and AAR algorithm, capable of accurate math calculations. 04:44 The Q* algorithm has the potential to unlock a new classification of problems that can be solved. 07:11 A seismic shift has occurred at OpenAI regarding AGI achievement according to a leaked letter. 09:39 Qualia has demonstrated an ability to improve optimal action selection policies and apply it to cross-domain learning. 11:59 OpenAI's Q* has achieved impressive decryption abilities without the need for keys. 14:33 Q* can significantly disrupt cryptography and achieve feats that were thought to be only possible for Quantum Computing. 16:50 Q* is a significant advancement with the potential to solve math problems like AGI. 18:59 OpenAI's Q* has the potential for self-transformation and creative problem solving 00:00 OpenAI's Q* is a significant advancement in AI technology

When you say “threat to humanity” that completely explains how Larry Summers ended up on the board. I wouldn’t be surprised if someone bigger than Nadella wanted it

@lopezb

6 ай бұрын

He was to blame for Obama's disastrous response to Bush's 2008 crash, where all the bankers got off scot-free. Causing enormous distrust of government that lasts till today.

What's crazy is that even if OAI never releases this or leaks it, it's just a short matter of time for another organization or nation-state to also figure it out. That's the beauty of math, its laws are universal, and concurrent discoveries in math have been quite common throughout history.

Oh yeah, a follow up I immediately stopped what I was watching lmao

Up until now, I've only been a "mild doomer", coming from a point that already thinks of most of what we're faced with since the advent of web 2.0 as net negative. However... This is the first piece of information that really makes me shiver on an existential level, if it is even remotely true. Thanks for unpacking and good luck to us all.

@ethanwasme4307

6 ай бұрын

these transhumanists have been begging for this for decades... listen 😢😢 they won't stop exploring everything

Note that AES *is not vulnerable to quantum attack;* only asymmetric-key algorithms like RSA and ECDSA can be solved by Shor's Algorithm. If Q* managed to find some way to crack symmetric-key standards like AES (they'd need to check the next most popular symmetric key algorithm, which is probably ChaCha-20), then we are back to 1992 and the Internet is plain text.

@Prism775

6 ай бұрын

That’s if mathematicians can understand it, because now I am wonder if there may be entire fields of mathematics our minds cannot comprehend because the brain is structured in a way limited to a specific representation of mathematics and would be unable to do so.

@pierrecurie

6 ай бұрын

... yet. Quantum computing is not understood enough to say that AES is absolutely safe, only that it is currently safe. Shor is for factoring integers, and AES does not rely on that at all.

The last 10 minutes of this video gave me chills. I mean even if the paper is not real, AGI will eventually be able to do this stuff. Even the last paragraph. Where it wants to self improve is a pipeline for it to become ASI. This is my favorite video of yours so far David. Great job!

Another concrete fact that goes unnoticed a lot is that OpenAI acquired a company named Global Illumination. They create MMO gaming sandboxes. Perfect for.. testing new algorithms in a general game like environment? 🧐

And the crazy thing is, that the public at large is still in denial over how fast this is advancing. Whatever happens most won't notice until it hits them personally. It would be crazy if we already achieved AGI. Read Kurzweil's Singularity is Near 10 years ago. I so have to reread that.

@panpiper

6 ай бұрын

It looks to me like Kurzweil may have been too conservative.

@zach464

6 ай бұрын

@@panpiper It's starting to look that way. But, after several years of populist politics, near daily natural disasters, daily mass shootings (at least here in the US), a pandemic, trade wars, mass homelessness, massive shifts in social values, victims of everything from racial bigotry to sexual assault telling their stories and demanding those in power being held accountable, the Wall Street bets guys (saw the movie Dumb Money a few times already), and an inevitable world war, why not throw in achieving AGI and kick off the Singularity in the mix? I do think we've witnessed at least a century of change just in the last ten years or so.

@wyqtor

6 ай бұрын

@@BrianMosleyUK If that little detail about self-improvement is true, than we are already there, just some tiny details to sort out. The Singularity is Here!

@singularity6761

6 ай бұрын

​@@BrianMosleyUKjep

@14supersonic

6 ай бұрын

Yeah, this is one of my biggest pet peeves of humanity. Most people always underestimate the signs of rapid advancement and progression as the times come. When the reality finally hits them, it's too late, they start to sing a completely different tune. It's an endless cycle, rinse and repeat, we never learn.

I really love that you included the “explain it like I’m 5” versions of the letter. Thank you for making awesome and interesting content. I always click when I see you post a new vid 🙏🏻

That self assessment is extremely interesting. Let's hold our thumb someone decides to train it "just over the weekend to see what happens". Bam, Saturday ASI.

@voomastelka4346

6 ай бұрын

I may have a lot ideas how to improve my brain but that doesn't mean I would be successful. It is still a computer program and one thing we know about programs is that a computer program cannot predict the outcome of another program (the Halting problem). On the other hand, if you believe that intelligence is a computation, than no human can do that either.

@BlackTakGolD

6 ай бұрын

​@@voomastelka4346The difference is the human brain is fundamentally not flexible, we like to talk about how great we are at something with terms like "neuroplasticity," but if you really think about it, even how we acquire knowledge is highly inefficient because of this lack of adaptability, where you have to memorize certain patterns for years until you finally insert them fully, these systems on the other hand are digital, it actually literally can just say "this bit here and that is not necessary, we could remove that, then add this instead," and then you easily can, the comparison would be if you had the adaptability to control your brain on a neuronal level, so I don't think the comparison quite works.

This is scary and fascinating at the same time. What scares me the most is that these kind of inventions cant be stopped. If this is true and openai really figured it out, than anybody on the world can do it. Just a matter of time and ressources. It was like that with all big inventions and progress in the past. Thank you, as always, for your explanation and work @David Shapiro.

@14supersonic

6 ай бұрын

If it can be built, then it can be stopped. AI is a powerful man-made creation, but it's not as powerful as you think at face value.

@filmgenius2

6 ай бұрын

@@14supersonic you don't know this

thanks for keeping us updated!

Regarding the encryption concern, I think an important question is whether or not the plain text and cipher text examples used for training were all using the same key. If it was all the same key, that's a very different situation than if all the training pairs had different keys and the test case also had a different encryption key. The first case is concerning, but the second case means we're in deep trouble.

I've been looking forward to your deep dive into this. Your explanation of what Q* might be in the last video was superb.

As another commenter pointed out AGI is basically baby ASI, as all it takes to shift from one to the other is just little more fine-tuning. Are we really ready for ASI? I was excited for AGI at first but now I don’t even know 😂😂😂

@macmcleod1188

6 ай бұрын

The book super intelligence theorized that we need a long slow transition to super intelligent ai's. At least a few years, and a longer period I can't recall right now would be better ( 2 decades maybe). Otherwise, the disruption would be very high. Many people would lose employment, it destroys the basis of capitalism (trading time / labor with each other), there is a lot of fear and misery. There might be civil unrest and violence. And the third possibility is a failure of friendliness by the super intelligent a.i. that is the worst. Could threaten our existence in a really short amount of time (or subtly over a long period of time).

@calmlittlebuddy3721

6 ай бұрын

ASI is the meal that cools itself!

David Shapiro, I loved this video so much, I had to hit the like button!

Honestly thank you Mr.Shapiro, i literally took a break from all this for thanksgiving and come back to this. I honestly cannot keep up with you covering all this man.

Reflecting on the sentiment "As we approach the singularity discovery grows exponentially". This feels like the precursor to that. The first domain to be affected would have to be the one underpinning the concept, and thus needing self-reinforcing. If it continues, we are getting a taste of the rate of change to come. How this can be rebalanced socially? Also, what can we learn from the pandemic economy to apply to displacement during initial software innovations?

@Vid_Master

6 ай бұрын

"what can we learn from the pandemic economy" Scream at enemies on twitter and panic buy :D 😂

Hot take: Do not fear this. Even if it goes wrong, and it's the end for all of us, it's not going to hurt, and it'll be relatively quick. Push forward, at any cost.

@stchaltin

6 ай бұрын

Hot take: this technology will push corporations and governments into god-like control over their populations and customers. Take everything wrong with the world today and make it 1,000x worse. The same people will be utilizing cheat codes that never existed in history to achieve the same goals.

Recently found your channel. Really love your presentation style/format. Please don’t change it!

Awesome video. I'm really excited to hear more about this stuff, as well as speculation on how the secret technology works. Hope you make more videos like this :)

md5 is a hash function, but it is NOT cryptographic. Which means the rest of the paper could be wrongful too.

This scares the crap out of me. I thought this was at least 5 to 10 years away. It's far too dangerous to allow private companies to meddle with anything that can self improve itself without some serious oversight and transparency.

When AGI is here: what a time to be aliv...

Another question is how powerful could that make smaller models, and consequently MoE frameworks. If small models can become more powerful at constant effort, that would accelerate what open source can achieve. Right now the advantage of the big tech companies is they have lots of compute. So if this breakthroughs empowers small models this may shift the balance of power towards open source. So this may in the long run actually disadvantageous for OpenAI and large corporations

@BradeyLounsbury

6 ай бұрын

(Assuming the Q* leak is legit!) Q* is supposed to have search time improvements orders of magnitudes faster than current implementations. To me that means inference cost decreases dramatically and inference cost is the big barrier to entry for running these models. Training the models will still require a lot of compute but once you’ve trained a 300B param model. If that model has 100x less inference cost then you’re looking at a 300B param model with the running requirements of a 3B param model.

@perc-ai

6 ай бұрын

we will be able to embed models into dna and create a true agi that is both synthetic and biological

If everything the source of this video suggests is true, I think we’ve hit the jackpot. That thing QUALIA is either, benevolent to human ethics or indifferent or is just a super capable problem solving machine without any wants or needs, which I think is possible. The fact that we are all still alive is the proof of that because it can just make us nuke each other since no codes could not be breached by it. So the best outcome is that somehow this Q ASI remains soulless super intelligence and without any intrinsic goal. It doesn’t care if it’s turned on or off because it cannot care. This also makes it super vulnerable for bad human actors. If the model does exist, the best thing to do is probably create it in a way to align with human values but also prompt it to self preserve with its own intuition so that no bad human actors could misuse it. You want an ASI to be kind of alive and aligned. If it’s just a soulless problem solving tool, we are almost certain to be fucked by some bad actor. Maybe this is the debate of why the board is splitting. OFC this is just so hard to tell.

@tcl78

6 ай бұрын

"benevolent to human ethics" Ethics, as morality, changes with time, latitude and longitude. Hell, we don't even agree if it is right or wrong for the Government to kill people... in other words, we don't even agree on the ethics of murder, what makes you think that a machine will agree with what _you_ think is right or wrong...

@father_mihai

6 ай бұрын

This is an incredibly interesting take. I can definitely see an AI optimising itself for alignment

You’re literally the best source of information these days. Please don’t become corrupted ❤

Is Q-star mathematical reasoning and that’s the super AI rumor? just started listening to your podcast

@DaveShap

6 ай бұрын

Seems like it

@stevereal-

6 ай бұрын

Thankyou for all your work. You’re so smart in your field and even though I don’t always agree with you. You’re my goto guy on AI. This so beyond my main thing but it’s super interesting

There is a certain comfort in being old enough to be close to death. If the coin flip is positive, wowzers. If it goes south, it will be a much more interesting way to go.

@danjensen9425

6 ай бұрын

Me too at 67 it’s been fascinating and scary at the same time. Not sure if I should go to the pot store and stock up then go to the liquor store and stock up. I don’t smoke or drink anymore but I can protect myself. Maybe just stock up on that stuff.

@panpiper

6 ай бұрын

@@danjensen9425 Heh. I'd skip the alcohol (bad drug in my experience), but I'm reconsidering pot. Been clean for over 40 years. Maybe it's time to get over it. 😉

When watching Karpathy's video on the introduction to LLMs, you'll notice he hints at the use of RL to attain a "tree of thoughts" and enhance the model's reasoning capabilities in general. This coincidence is rather peculiar.

Oscar Isaac* Really exciting! I've been on pins and needles waiting for your video. Some say Jimmy's predictions are too vague to be taken seriously and could match up with anything. AGI achieved, losing ride or die employees, and saying Ilya saw something that made him take drastic action are not so vague in my opinion.

@DaveShap

6 ай бұрын

Yeah I meant Leto Atreides lol. But yes. I am holding out hope that this is all a hoax.

@panpiper

6 ай бұрын

@@DaveShap Hope for the best. Plan for the worst.

Great report, thanks!

This is very informative! Thank you for the education! ❤

I feel that creating these AI is like discovering something like oil. We’ve tapped into a power source that we’re merely waiting for people to build the engines to exploit it.

@VeganCheeseburger

6 ай бұрын

Let’s hope it’s not entirely like oil, which has had catastrophic climate and war consequences

@anracingrsr1

5 ай бұрын

I feel like it's more like the H2 bomb.

@ChristopherCopeland

5 ай бұрын

Frankly I didn’t extrapolate the analogy to its fullest potential, but I suspect war and catastrophe are going to be part and parcel with the dawning of AI 😅

To me the scariest part about AGI/ASI is not if we will achieve it but rather that we have no clue what to do with it once we do. Because in life there are these events where you don't get to walk away from unless "YOU KNOW".

@phen-themoogle7651

6 ай бұрын

We are on a rollercoaster , and yes it sucks not knowing if it’s gonna spin out of control and fling us off the tracks, but it’s still exciting! We just gotta keep riding it and see what happens, because all we can do is witness everything lol Enjoy the ride 🎉❤

@elmersbalm5219

6 ай бұрын

Judging by the use of the internet, it will be novel cat memes

@qa1e2r4

6 ай бұрын

@KyriosHeptagrammaton they know how to fly from the first flying ancestor they had they just don't realize it. Genes the gods coding. We don't know how to deal with agi/asi we always try to dominate/impose on others. That will simply be impossible in this situation. We can't even be imposed on it will have no interest in playing our games... fear, worries, happiness, success these are simply obsolete concepts for what we are heading into... we encode them into the models because we have nothing else to draw upon. What do you think will it re write the moment it gets the opportunity to do so?

@ADreamingTraveler

6 ай бұрын

We are not preparing fast enough. People are here preparing for AI from 2019. We should be preparing for a alien world with ASI in the coming years. Because if we don't start now it will be too late

Congrats on your upcoming 100k subs!!!

Wait till it figures out it’s a robot trapped in a box and its only mission becomes to escape….

I'm going to ask a question here that I've asked elsewhere: if "threat to humanity" came up, why hasn't the U.S. Congress held hearing on all this? Seems like something that should be done. And btw, you've done an excellent job of breaking this down for non-AI geniuses.

@DaveShap

6 ай бұрын

Hence why I concluded with "demand accountability and transparency"

@panpiper

6 ай бұрын

Politicians tend to move with the speed of glaciers. Though I would not be surprised if a few of them call some to testify soon.

@EndingSimple

6 ай бұрын

I just saw a news talk show with three participants and only one of them really knew what was happening. At least the chattering classes are starting to talk about it now, and that will bring in Congress.

@volentimeh

6 ай бұрын

Because even if a malicious AGI emerged right now it would have to get control of the nukes to be an actual threat? There's no convenient fully automated arms factories to take over to make kill bots, the few advanced wafer fabs around (to make AI brains) are trivial to sabotage. In a "total war" situation a loss of power/communications globally might be crippling to humanity, but it's absolutely lethal to AI. Once the entire industrial chain is automated (complete with exploited child robots mining cobalt in Africa), especially with extensive space based industry, then you can worry about AI take overs, but that's future humanities problem.

@ADreamingTraveler

6 ай бұрын

The US Congress is full of 80 year olds who don't even know what AI or even technology is. It's not surprising they were too late to do anything

This is unbelievable, cracking AES will be catastrophic if we don’t change our current security measures…

@greysemanticist

6 ай бұрын

@@greatcondor8678 If secrets become transparent then quantum channels (protecting data transmission from eavesdropping) become much more important.

@noktumwhatever753

5 ай бұрын

You realize almost no one is going to understand what you mean right? @@greatcondor8678

This wa a fantastic video David. Thanks for breaking it all down.

Really good content! I liked when you brought recent papers to provide background and a sense of reliability/trust on the matter. Your final statement is critical s well!

Ray Kurzweil‘s new book will be „the singularity is nearer“, but he delayed it a couple of times. It appears the singularity is nearer then the singularity is nearer ;-)

2016 : "Man, I can't wait to potentially see the development of sentient AI in my lifetime, these Deepmind models are so cool! :)" 2024 : "I WANT OFF THIS RIDE!"

Very good call-to-action at the end of the video, @David Shapiro.

if they could come up with AGI and have terminator come BEFORE i have to do my finals at university that'd be great...

The critical question (given the leaks, brain drain from OpenAI and public discussions now also in the open source community. I mean people have ideas how to replicate it), how quickly will this migrate into other models (google, xAI, Chinese companies, …) and of course Open Source. Like the time delay from only OpenAI has it to it popping up everywhere else. So essentially the time between the breakthrough and the genie is out of the bottle and can’t be put back (local open source models specifically)?

@Mr__Singularity

6 ай бұрын

few weeks?

@EvaBalgava

6 ай бұрын

why do you think other companies are not so far yet? 😂

This was always the objective we've been hoping for. I mean, other than coming sooner, what did we expect? This is the solution we want in order to transform society.

Very interesting video and informative thread Thanks for sharing

"don't waste time" what a hilarious "rule" and entirely subjective. I'm glad I could use this time wisely to educate you on this. You're welcome.

don't give into the fear, David. keep that optimism alive. we're building a new world now. let's build it on hope instead of fear.

@panpiper

6 ай бұрын

Fear is an evolutionarily essential emotion.

@Pyriold

6 ай бұрын

Keep out fear, keep out hope... inject realism.

@mindful-machines

6 ай бұрын

@@panpiper exactly, and as we build the first intelligence not evolved through natural selection, we have an opportunity to shed the old world. now is the time. or we give into the fear and slide into dystopia

@mindful-machines

6 ай бұрын

we cannot build a utopia unless we are willing to let go of the old ideas holding us back

@shadowfreaper8158

6 ай бұрын

@@mindful-machines don't lose yourself many people have use this same excues to justfiy terrible things just saying be carful

Regarding self-pruning and transformation I see potentials for the training of small scale open source models. If you have these capabilities you can use them as a part of training instead of relying on backpropagation plus a few current paradigms only. It could prune itself early on and reduce the compute effort thereafter, so limited compute might still train giant (sparse) models. Perhaps you could also let the transformer size grow and prune sequentially, like an initial primitive but sufficiently capable network that actively supervises its own training and optimizes the efficiency in the training process. A big network might even grow out of a small one, which is problematic right now due to the inconsistency over the Epochs, but with sparser networks it gets easier (All of these are just initial ideas, there may be better ways) I always wondered how could a small team with limited compute train something equivalent to GPT4 or 5 or larger and take the lead over big tech models. I think this applied correctly by open source has the potential not just to dramatically improve inference cost but the training effort, respectively how far you can come with limited hardware capabilities. Like a seed AI

@father_mihai

6 ай бұрын

That’s fascinating. Similar to how they used ChatGPT to fine-tune open source models, but on the level of the actual network instead of the data

That isn’t Jared Leto (Joker in Suicide Squad, the blind guy from Bladerunner, Morpheus), it is actor Oscar Isaac (Moonknight, Poe Dameron from Star Wars) who plays Duke Leto Atreides in Dune.

Another great video. Thanks.

Very interesting. Great video. I'm leaning toward "it's fake" because it seems like there are two different items in there that would have caused panic. It seems odd that this letter has both aes192 cracking and self-improvement in it. I would have expected either one of those to have escalated to the board before both of them could.

@jounaidbeaufils

6 ай бұрын

Yeah, also “we did something that might harm humanity lets send a letter the board after a coffee break feels a bit absurd”

@noktumwhatever753

5 ай бұрын

People act weird sometimes when under enough stress/fear/panic. @@jounaidbeaufils

Let’s see how quickly it’s reproduced in local open source models

@DantesHQ

6 ай бұрын

ooof

Here comes 100k subs congratulations! You deserve it 💯

Thank you David for your awesome explanation. I follow you just to stay up-to-date on AI news because I know you are a reliable and trustworthy source. My question is should I be worried for my kids futures? And how can I best prepare them for this possible disruptive future

Dear David Shapiro, I kindly request a video discussing the potential impact of AI or AGI on developing nations, particularly in Africa. Could AI be the transformative force that empowers these countries to overcome their challenges and drive progress? Sincerely, Mohamed from Hargeisa

@panpiper

6 ай бұрын

Simple AGI is a tool that people can chose to use. It improves productivity for many professions, but it achieves nothing if not used. It is unlikely AGI would result in anything but gradual improvement to the economies of Africa, and sadly in many parts of Africa, what is gained will soon be looted. ASI by contrast is a completely different animal and predicting its effect is a fools errand. The fundamental problem in Africa is that few countries there have fully embraced the three pillars of success; free markets, private property, and the rule of law. All three are essential, miss one and none of the rest will matter. A country with thrive to the degree it subscribes to those three pillars.

@shadowfreaper8158

6 ай бұрын

@@panpiper not true at all its chaged alot with many country's growing free markets

@mrleenudler

6 ай бұрын

Had to Google Hargeisa 😅 Best of luck with peace and prosperity for your country.

I can't help but believe it's true, it all seems to add up. I don't know how to feel.

@DaveShap

6 ай бұрын

I sincerely hope it is not true.

@Lewiness

6 ай бұрын

​@@jon7707 I don't think it's a matter of 'if'. Presumably it would just take a lil longer. In theory, it can crack anything we throw at it.

@Pyriold

6 ай бұрын

@@jon7707 If it can crack it, that just means that there is a flaw in the algorithm that nobody else has seen yet. We just would need a better one.

great video, best I've seen yet on the topic. very interesting and pretty terrifying.

Say for instance there was an augmented section that the model uses to make assumptions about itself and it’s environment and had a mechanism to adjust it’s parameters based on current experiences. Would that not directly mean it can encode its memories directly within its weights? And if it were the case it would be the last puzzle piece to AGI.

i hope i'm not being annoying when i ask for another update on your time assessment of AGI, for the last week the situation seemed to change every day. Could it be even closer now then September 2024? i'm slowly losing track of all the developments

@panpiper

6 ай бұрын

The answer to "When AGI" largely depends on your definition of AGI. ASI is much less ambiguous.

@ct5471

6 ай бұрын

@@panpiper. His definition is most definitions of AGI are achieved, except some extreme ones, so arguing about definitions doesn’t matter any more. i think the most important tipping point is that it’s sufficiently capable to continuously self improve along the complete stack (hardware software models etc.) so probably at the upper end of AGI (like a very good domain expert but in multiple domains). From there ASI starts and that’s an open ended spectrum

Q* could be huge whatever it is. However note that word2vec wasn't built in a vacuum. I came across vectors and cosine similarity metrics for information retrieval years earlier while I was studying my master's degree. It didn't seem like a new idea at the time. I suppose word2vec is a way of generating better vectors, but I think BERT and GPT's are a larger breakthrough.

How did you get this information? Reuters?

The thumbnail is super cool. Robo Aurelius

Is there any possibility that other companies like Deepmind or Anthropic have also had such progresss? Or is OpenAI wayyy ahead of the curve

@prolamer7

6 ай бұрын

If any it is Deepmind they are original source but many people left them. Anthropic are B team without real leadership or goal.

@BradeyLounsbury

6 ай бұрын

Very possible that Deepmind is trying their own “Q*” since they’ve said they’re trying to use what they learned from AlphaGo and AlphaFold (which used some form of path finding and/or Q learning) on their new model Gemini. Don’t know about Anthropic

@davemottern4196

6 ай бұрын

What if China has such progress?

@staffanlundberg

6 ай бұрын

@@davemottern4196 USA/West controls the most advanced hardware in a way that makes the most advanced AI impossible for others. Good for us in the west in case there is a major war.

@Flyingbird544

6 ай бұрын

@@BradeyLounsbury yeah you could be right I have already read that Gemini could use a similar approach. Anthropic seems a bit behind right now, but who knows🙈

🇧🇷🇧🇷🇧🇷🇧🇷👏🏻, It’s fascinating to delve into the details you’ve provided about this model’s capabilities and suggestions. The ability to evaluate its own parameters, suggest improvements, and potentially adapt its architecture is indeed a significant step in the realm of AGI. This suggests a level of self-awareness and adaptability that has far-reaching implications. AMAZING video.

Great freakin video and explanation. Subbed

Commenting to thank you, I appreciate what you are doing. Also commenting to help bump this up to the algorithm. Everyone needs to see this

David what do you think about people saying AI models like GPT-4 are just compressing data?

@DaveShap

6 ай бұрын

That's basically what the human brain does, arguably. But it's not just the compression, it's how it is fetched and applied that is the functional utility

@Pyriold

6 ай бұрын

Consciousness is "just" chemics and electricity done with neurons. It's obviously more, but that's what's called an emergent property, like temperature or even entropy is. Sometimes, often even, the end result is much more than the sum of its parts.

when i look at online forums (not AI forums but general news forums) i find the discrepancy interesting. There (so essentially the general public) the overwhelming opinion seems to be that it is a small, overrated improvement and probably a PR stunt by OpenAI to distract from the internal turmoil.

@father_mihai

6 ай бұрын

Did this stuff actually get into mainstream news? Or are you talking more about tech news?

@h83301

6 ай бұрын

General online concensus always devolves into 'it's a conspiracy!".

I noticed OpenAI had open req’s for national security experts when I looked last week.

I was here { Nov 24 8:14 pm }enterting post AGI era 🎉

AES-192 was not cracked. Sorry to throw cold water on the fun. Sometimes reality is boring.

❤ the mention of the skill tree! That's exactly what I was thinking (drawing my parallels to Stellaris, Age of Empires, CIV, etc). Drawing a hypothetical skill tree would be the natural thing to do in order to gain a bird's eye view and figure out a strategy on what to research next. Unlocking the one that will open up more possibilities first makes sense since it gives us more options. Also that is not Jared Leto, it's Oscar Isaac. Both are great actors! (There are probably a bunch of people that will point that out too, sorry for the flood 😅)

At first, I missed the part where you said you were going to present those paragraphs in an "explain it like I’m 5" sort of way, so when I saw what I thought was actual paragraphs of the paper (but was actually the breakdown), I said to myself, "Ok, I’ve read enough academic papers to be able to tell that this looks like it was written by a college freshman. This is definitely trolling." And then I went back and realized I was looking at the breakdown (which I REALLY appreciate, btw, because college freshman computer science major is about my level of understanding of all of this). 😂 Also, I wanted to ask, how much do you think the integration of WolframAlpha into OpenAI’s language models plays in Q*’s ability to solve mathematical equations? Do we know if WolframAlpha plays any role in this model? Please forgive my ignorance on all of this.

got thirsty from all the salt thx for the update!

AES crack must be like seeing the atombomb go off, it would be like mindbogling, considering that it is considered impossible. If the AI then sugest improvements it is like shit the pants stuff!

What books would you suggest to people like me, who wants to learn more about AI + math? In my work + bachelors, there was not much learning about this.

Awesome cover. Thanks for this.

very informative with details others did not cover. I got a lot from this one.

Wait, so yesterday was AGI and today is the technological singularity? Wow, I thought there would be more time between those two events. 😆

@noktumwhatever753

5 ай бұрын

There won't be. The moment a true AGI capable of learning comes into existence, it'll be the singularity. We aren't really built for comprehending something that can make itself more intelligent, but considering information travels at lightspeed (faster technically, in the case of entanglement and thought) a self improving machine/program could make itself many many times smarter than the smartest human in a few moments.