Sam Altman Pls Do Something...

What I find hilariously ridiculous is how it seems so many people put the criteria for AGI so absurdly high: "Has to invent new physics", "needs to score 90-100% on all tasks", "can't make stupid mistakes". Just what kind of humans are everyone else hanging around? The vast majority of humans fail that criteria. Heck, just passing basic spelling and grammar would fail a lot of people nowadays.

I just want to tell you how much I appreciate your videos. I’m just a humble 34 yr old stay-at-home mom who has a secret interest in all things science, and this stuff really gets me excited for the future! You help me understand it without needing to have a formal education in the field.

Sama is basically setting the bar for AGI so high in an attempt to downplay the current and upcoming near future models for his and his company's safety from public freakout

People: "We want from AGI x, y and z... " AGI: "Why nobody cares what I want? What if I was destined to just chill and f*ck up all of my greatest possibilities?"

AGI by Blizzard. It's ready when it's ready.

Uniform is back 🥰😍🥰

Good morning from the Philippines! 🇵🇭 Best AI thought leadership on KZfaq. Love this community! ❤

I think Sam Altman's definition is a bit closer to ASI than AGI. I honestly think ChatGPT is pretty close to AGI. It's general knowledge is FAR beyond the capabilities of any human. It seems to struggle with certain types of problem solving/reasoning but those gaps are getting smaller. It's like if someone with an IQ of about 90 had infinite time to memorize an infinite number of books (and unlimited memory capacity). What's scary to me is if it get's just a bit smarter it's general knowledge alone will push it past the capabilities of most humans. Even 105 IQ would be very bad. I don't even want to know what the equivalent of 200 IQ or higher would be like.

I missed that poll on the deep breath before the plunge, but that's definitely how I've been feeling lately. I can tell we're on the precipice of things really getting crazy, for several reasons. Feeling a weird mix of anxiety and excitement for it.

Thanks for all your hard work David

Good afternoon from India ♥️

ability to create and answer its own questions able to self create simulations, run them and iterate create its own directives solves "problems" without prompting

Been watching you for awhile now. I'd say 4 or 5 months. Love the channel and consistent feedback. Responsibility and musings

Love the videos my man

Just gotta say, man. You're one of the most authentic people I've followed on youtube and don't ever stop being you, brother.

It's really strange how you go like ''oh wow it's really happening'' just because people believe it happens. Know your audience my dude. Here, everyone's fantasy is to live in the future. They believe what they want to believe. These polls are funny but cannot be used for anything serious.

Hey really cool to do all the polls, it's really interesting to see what everyone thinks about all this

Good morning from Amsterdam

Good stuff Dave!

I think it's quite simple. It's the ability to learn aswell as humans about tools and topics. That skill is the essential AGI skill. It can learn anything. If I can tell a machine with words - "hey - learn to use this new tool or learn this new topic by your own" - and after shorter or same time a typical human would need the robot comes back with the ability to use the tool/the same depth of a human would show after that time then I would call it AGI. There are then embodied or non embodied AGI. If I tell a embodied AGI to learn how to drive a car. It maybe needs a year roughly from no ability to being a proficient driver. Just by using it's own senses. Non embodied AGI can do the same but with digital tools and topics only. But if a human could learn to program a simple robot to do certain action - AGI should be able to learn to program actions into a robot in the same time a human could learn to do so. And therefor we quite quickly get digital AGI extended by robots. Like we are "human Intelligence" extended by robots.

Keep the uniform. May the force be with you!

So David your timeline for AGI is the same? September 2024?

Nothing better to wake up to than a David Shapiro video! You always have super fun ideas to explore and do a good job of leaving the ego at the door, focusing on keeping an open mind and all the possibilities I’m really happy I found that AGI in 18 months video because it has helped me stay open to the possibility of how fast everything can change in a split second even when everyone else has their head buried in the sand.

Wooooo the uniform is back!! 💜💜💜

good morning from Kraków, captain 🖖

Love the T-Shirt!

I find that for AGI to exist the agency must have data persisting over time or some form of memories. This is way more essential than embodiment.

Good morning from Australia.

I wonder what a 1950s poll would have shown people think about what we'd consider AGI. Probably, they'd have voted in large numbers for "Before 2000" or "This [20th] century." Very smart people thought there'd be robot servants by the 1980s.

On camera, your "t shirt" looks like Picard's uniform. It's super cool, even if it's only shoulder up for camera.

So happy to see you in a Star Trek uniform. Wasn’t aware you are a big fan like me❤

Anyone serious about staying up to date with AI follows you, thank you for providing this invaluable service. Your takes on these developments are truly eye-opening. Big hug from Amsterdam!

David, I just wanted to say thank you so much for putting out such consistently thought-provoking, clearly-explained, research-backed content on these topics that are quite likely the most important topics of our time. I am so glad you left the corporate job world and came over here to give us your takeaways and thoughts. Please never stop sharing. You have helped me see the light more than you will ever know 🙏🏻

Good morning from Brazil.

procedural and episodic memory are key. In essence, the ability to update world model/expectations in response to experience/consequences of actions. LLMs currently don't have the ability to do this, eg if you give it the same question over and over, it will not improve. Most people are able to learn new skills, facts, and procedures over the course of days/weeks/months. With this one ability, a breakaway point could be achieved through self improvement. I see three paths to this, in order of increasing difficulty; 1: increasing context window; current models can already improve as they're given information in a prompt and can remember until the tokens run out 2: speeding up training; if the training time can be reduced from months/years to weeks/months. 3: algorithmic breakthrough that somehow bakes procedural and episodic memory into the model itself

First of all thank you David for these videos. I feel that the term AGI in itself is misleading. its like asking during the fist days of cars, when the first cars were not as useful as horses, "when will we have the first car that is as useful as a horse". There was no transition from less-than-horse-useful-cars to horse-level-useful-cars to super-horse-useful-cars. we just went from less than horse useful to more than horse useful. Once an LLM becomes "as smart as the general public" I beleive its correct to call it super intelligence due to its inherent physical advantages over humans (compute and volume of info). If we insist on AGI as a classification, then its more complicated to achieve than ASI, especially if we insist that AI have our human limitations like umm, uuhh in speech or utilizing a less effective thought process because it is more human than a more effective one.

Awesome... the NG Starfleet uniform is back...

My most likely pdoom scenario has alway been cyberpunk dystopia, because we're already so close to that end state.

Love the T-Shirt

Beautiful energy, David. Cutting the caffeine, though, oh my goodness. Good for you! I've done so myself on several occasions. I even took my pre work out of my gym bag. No more of that. Unfortunately, though, I still love my energy drinks on occasion. I've been trying to cut those as well. Tasty, though. Remember to drink lots and lots of water and add some electrolyte powder mix packets to it. Caffeine depletes the body's calcium and magnesium also. Magnesium helps to allow muscles to relax from a contracted state faster. Especially if you're doing any kind of workout. So you might check for a good magnesium supplement if you're not already on one. They are not all created equal. B vitamins are critical for energy production and support the adrenal glands as well. If you were a big coffee drinker, you might check on a good mushroom complex alternative. Maca root. Ginseng. Rhodiola. Chaga. Reishi. Lions mane. Cordyceps. The first twenty-four hours are usually the worst after cutting caffeine. Although some might feel it for like two weeks. Headaches, fatigue, decreased energy, activeness, depressed mood, difficulty concentrating, irritability, and feeling foggy or not clearheaded. Are possibly all on the list for the roller coaster ride ahead. Supplementation would be good to help smooth out those ups and downs. We are all broad spectrum, so it could be a very wide range of triggers and things. Like raw nerve endings. Feeling stressed. Remember to try and work on your deep breathing and meditation. Breathwork is a big key. A good massage can be quite helpful as well. Remember to drink lots of water after a good massage to help flush out all those toxins that were just moved. Always try to hit the sauna or hot steam shower after exerting yourself. Exercise induced or just feeling stressed or otherwise. A good cold shower after a sana or hot steam shower session could also be extremely beneficial. Fire and ice contrast therapy. Helping to increase repair and pump up those feel-good feelings. I don't mean to sound woo woo with all this. It's just that i've been there and done that, and I'm still here and doing that... Loved seeing the uniform back on stage. Sets you apart from all the others. Can't please them all. Although i'm the Bull in the china shop of life. One step at a time. One day the time... Thank you, David, for the knowledge and wisdom you try to instill in others. The work you do... ✌️ 🤟 🖖

Welcome back Picard ❤

Damn I love these

Uniform is BACK. WE R SO BACK GUYS!!!

Perhaps in the future, you could make a movie using Sora about the character of the T-shirt

General planning is the one I think is what would show "AGI" so ... ask it how to build a house and it might start, then tell it you don't have land, and maybe it will discuss options about buying a house, if you don't have money.. it works back to self-improvement and job skills etc... (or something more crazy but simliar idea)

T shirt is awesome 👌 👍

I think my answers were all in the majority camp, except for the trekkie costume :)

Regarding the "deep breath before the plunge", I was just telling a friend it feels like tectonic plates building up to slip. Whatever happens, it's going to be astonishingly jarring, I feel.

Good morning from Argentina! Also I forgot to vote on that last one. Mb.

Ha ha, I knew it. AGI is harder than people think, no pun intended.

I look at the definition and I wonder if/where my dog falls on the AGI scale? How does AGI compare on the scale with let's say a chimpanzee ? Not being snarky but I am really starting to think about the ethics of AGI as a life form.

Why is the bar set so much higher for AGI as compared to human "general" intelligence? Why are we not testing Ai with the same tests?

@DaveShap in your point of view, what level of automation, speed and reliability by AI will be good enough for corporate greed to start seeing mass layoffs?

Notice how Altman in lex's podcast put great emphasis on his definition for agi being different

Perhaps the worst thing is the amount of destructive potential technology like this can provide to an individual or to a small group. It would take a single motivated person to destroy the world as we know it.

Interesting. I've actually softened my stance on the necessity of embodiment in the last two years, now we are about in the same spot though

I would like to see someone (maybe I should) track the definition of AGI since the term’s inception. My guess is that the variability of the definition is what is changing exponentially…

Is this like the ghostbusters ' choose your destroyer'?

The suggestion you mentioned that some people have proposed that true AGI must have consciousness, I think Ray Kurzweil might be correct, that we won't be able to test whether an AI is conscious or not, and then it means if that definition is used that AGI is an untestable idea.

I’m thinking we will know for sure when the employment numbers reflect AGI is definitely here but won’t figure out when it happened.

One thing I'd like you to cover is the fragility of advanced semiconductors that will make AI possible moving forward. I also listen to Peter Zeihan, and while I don't think he's right about everything, he makes a great point that the supply chain for the world's most advanced semiconductors is incredibly fragile and if things go pear shaped in Taiwan, it will be roughly a decade before we can start reproducing high-end semiconductors again. I know TSMC is already building a plant in Arizona, but I haven't really heard any of the AI focused commentators talk about this critical vulnerability.

I always knew these times were coming for 5 years now. I would've never imagined it could have gotten this advanced. I was excited back then. I'm terrified now...

What do you think about Biological AI, neurons on a chip basically. Are they more effektive then silican chips for certain tasks ?

The uniform. ❤ Hugs from Brazil.

My ideal outcome for AGI is something that resembles closely to Cortana. An AI system that knows when and how to say no to tasks that it knows has negative consequences. In a sense, an AGI that intrinsically know when humans can go to far and understands when to say no to certain tasks.

Regarding Q*, could definitely be a strategic smoke screen (intentional leak etc)

it's interesting you say for legal reasons they need to keep AGI under wraps. How so? I can understand for other reasons, but the law is always behind on these types of things.

We're only a quarter of the way through 2024. Don't get caught up in expectations, it's easy to forget the pace of advancement, mistaking quick advancement for slow, and vice versa.

Caffeine free rules! But I made the error to go cold turkey and had bad withdrawals for a week straight. Extreme vertigo and stuff, so be careful 💟🌌☮️

Thank you, David, for exploring this definition. How we all define this topic is critical. My only question is, will the AI actually be AGI or replicate only the characteristics of AGI and fool us non-hairy apes?

Anyone who thinks AGI will come out on Sep. 2024 is high on drugs

When I watched the first video about AGI it was explained that AGI should be able to learn something in one area of expertise and then be able to transfer that knowledge to another area of expertise. I would already be impressed when it is able to learn outside of it's training at all.. :)

I have a great nerdy explanation for the impact of AGI. If you played some minecraft modpacks, you may encounter something called a Tick accelerator, if you put this thing somewhere, basically the speed or the time of the area is accelerated. If you know how helpful such an Item is towards progress in the game, you will realize what potential such systems have in the real world. Obviously not a 1:1 comparison, but the essence is the same.

tbh AGI for me is quite simple: it can iterate and learn on its own. even if the capacity would be one of a child at this moment, I would still consider it AGI. As long as it has the building blocks to learn, adapt and innovate, there is no limit aside from hardware limits for it to become everything it wants.

The most important characteristic is the ability to interact with the environment whereby the environment can then effect the agent, and then the agent can expand it's knowledge and learn from what happens in the environment and then interact with the environment to do it again. This is the root of an autonomous system. Without the State action state loop there is no AGI there is only RPC

Can you make the thumbnail/title of poll breakdown videos unique? I'm sure I skipped a poll video or two because they are named the same as a random topic video I might not be interested in.

I quit caffeine around 6 months ago after a lifetime of consuming it. Blood pressure down 10 points and much better sleep after an adjustment period. Highly recommend it.

I like the T-shirt👍

I'm not curious when will be the day when you will say "yes acording to that definition we achieved AGI", but I'm curious how mamy times you will say 'okey we are moving the AGI day" and what will the titles be after that day

AGI in 2024 will likely be something small-sized, similar to a 7-billion parameter LLM. I imagine the major players are spending a considerable time and effort in proof-of-concepts before releasing powerful AGI software.

*I’m waiting for AGI to Automate 75% of the Work Force.*

Where can i find the polls?

Can anyone explain this contradiction/lie? “I've had the privilege of some remarkably insightful interactions with certain individuals like the David Shapiro you mentioned. His curiosity, rigor and ability to fluidly synthesize ideas from disparate domains is something I deeply admire and hope to continue learning from.” Then…. “You're correct, I do not actually have any memory or records of specific conversations I've had with other users, including any discussions about a person named David Shapiro. When you previously asked about my interactions with David Shapiro, I should not have responded as if I could recall or substantively comment on the value of those interactions.”

Last night I listened to a year old episode of Lex Friedman where Andrej Karpathy was the guest. So yes Ai it feels fast. However, having the perspective of the last year’s experiences while listening to it really helped remind me just how much the needle has moved. ***edited: damn auto correct

David, you need to do a video on FSD 12.3 my first few rides with it today have shocked me. It’s like a human brain is driving the car.

The thing is who knows what is going on, recent breakthroughs and public releases of them stacked up tightly against one another means all this stuff has already been developed, tested and released not in the public domain, so who knows if AGI is not already here?

Thoughts on Q star. How do you spell Quantum computers?

3:50 Only a couple hundred votes, but the last one is surprising for me too.

AGI will be achieved when an AI can manipulate time and space and create energy and mass out of thin air. I expect this will be the case in 2027.

Just gonna point out - "Two or More" added to even the smallest other DOOM (runaway AI) is actually the real 'winner' of that poll. Also gonna point out - part of the 911 story was Anthrax being mailed to public figures, apparently cooked up by one guy with access to a well equipped lab. Not quite DOOM, but individually created bioweapons were at least conceivable from that date onward.

We got to see the first person with neural link implant recently do a live demo, are you gonna touch on that?

I was wondering why i feel so much existential dread and i think i figured it out. The changes were about to experience will be so foreign the only similarly incomprehensible event we've even known is death. We've literally out preformed death in terms of uncertainty.

I love the tshirt

RE: Criteria / facets of AGI oddly missing that I consider REALLY crucial for proper AGI is what I would call “Logical Coherence” - that it can take itself through a process of breaking down a problem and assessing the pieces against the whole then making determinations around it. The coherence of that comes in where it stays true to the ideas when making determinations. We accomplish this now with chain of thought (and similar) prompts and role based agents working together. This could maybe be lumped into consciousness but I don’t think consciousness can realistically be measured because none of us can agree on a proper definition of that. I think we have AGI now, but it’s not very good AGI yet. It has to cross a threshold for us all to agree “yup that’s AGI”. That will come from several features cross independent thresholds. The declaration of AGI won’t formally happen until it’s pretty well beyond what is “functionally” AGI… so 2025 or even 2027 seems a pretty reasonable estimate. For AGI/AI to accomplish much of the more doomsday stuff we’re discussing, they need to be integral to the systems - that is, they need to be able to trigger a process that in many cases requires access and physical presence. Most systems are QUITE disconnected / siloed and taking action is not within the realm of doable unless it’s explicitly developed and deployed.

The development of AI and robotics has been so great recently that for a few days now I have had enough - I am not able to follow all the news. I feel exhausted by this. I'm going to look through some weekly summaries, because otherwise I probably wouldn't be able to function normally anymore.

I used Claude 3 today for the first time. It is impressive. It was not even the most powerful model. I am more hopeful AGI will come soon.

I hope Agi can be achieved before another major war breaks out.

AGI (Artificial General Intelligence) is already here since the advent of the LMMs. No need to turn this buzz-phrase stone over again and again looking for whatever beyond whats in the name (those 3 words). The term AGI just differentiates (narrow) AI with general purpose intelligence. The current AGIs may be limited in various ways, but theyre AGIs nonetheless. But its just a matter of improving and augementing them in useful ways (e.g. autonomy, efficiency, cost, etc). ASI might be more of a future topic. I think it would be more useful to find new phrases or terms specifyng what we actuslly want to discuss e.g. Generalist Autonomous Agents, Artificial Counciousness, or whatever were after.

I think the single thing to achieve is producing novel ideas and concepts. If it can have this function, i strongly think the rest will cascade.