I’m an engineer and the amount of educated professional people in my space who are totally oblivious to the impending changes that are around the corner is unbelievable. It’s like “Don’t Look Up”.
One thing to remember: these agents may start out slow due to inference speeds, but eventually they will make working and making decisions at rates that humans simply cannot keep up with, sort of like stock trading bots already. This could be within months of deployment and refinement.
I have seen zero evidence of D Shap’s physical existence. He himself may not be aware, but all evidence points to him being a digital AI construct. A digital angel sent by futute AI gods to guide humanity to the next level of existence.
In the future, you will be able to work 100 jobs at once, just make sure you buy high quality AI agents and keep all your bosses satisfied with regular AI phone calls.
Yup totally agree. I wrote a whitepaper on such crypto-based agent-to-agent marketplace a while ago. No question it’s going to happen. The question is how you bootstrap adoption.
Look at the Autonomi project. They are building a serverless datacenter that uses the spare resources of phone, computers of ordinary people. It is opensource. In Star Trek we never saw a server room. This is the end of the data center.
That’s been done for many years now. That’s not new. The problem is that all of the crappy office computers and smartphones and laptops don’t actually add up to a ton of compute power, and there’s massive latency issues compared to a real data center. Data centers are far more efficient at what they do than combining tons of random computers that are hundreds of miles apart all running different hardware. And looking at a fictional sci fi TV show from the 60’s is not a substitute for actual research. I don’t think trillion dollar companies would be dumping billions of dollars into server farms if it wasn’t the best way to do it.
@@particle_wave7614 Clearly you have not looked into Autonomi. It has a long history that goes back before Bitcoin. It has never been done. Nodes will be run from home devices as well as data centers. A decentralized encrypted serverless internet with permanent storage has not been done well except for Autonomi. The network never has your keys. Files are decrypted and encrypted locally before being chunked at 8 copies per chunk and randomly distributed in XOR space. Pay once to upload download free forever. Cost to store increases as the network fills up to encourage more nodes to be run. Server farms are used now because that is the only way to do it currently. It also allows for centralization which corporations love. Data centers love locking people in and charging them to leave. Check it out.
The stock footage is not illustrative of what are you saying, Dave, so it’s not adding but instead is taking attention off of what you are saying. Is much better seeing you and a presentation with the bullet points beside for following your explanation.
I just want to say thank you for posting this. This is the positive attitude, and I think you’re right, first person that’s even mentioned DAWs! And the digital marketplaces that are kind of already here in some forms with midjourney and whatnot.
I think very small fine tuned models will win in the end. There will be thousands of models that are masters in a very specific field. The cost in the long run would be cheaper. Its really expensive to train the really large models and then it costs a lot to run them. If you can have a model that can basically be run on an iphone that is an expert in a specific field it would be very cheap to complete tasks. These really large models will always end up having a bit of an issue with getting things wrong because if they are predicting the next word and predicts the wrong word they can quickly go off topic. If its a small model and they are only focused on one thing they should have an easier time predicting the correct word. Then as the model gets used you just test and verify to ensure its outputting factual information and when it gets things wrong you can fairly quickly retrain the model. In many cases if during the day the engineers create a new dataset they could start training before they leave work for the week and then by the time they get back to work on Monday the training is done.
I absolutely agree with this video David. What people don’t understand is that autonomous agents are still confined within the controls of the Python code. The architecture / coding around the models control the ai, until ai can create its own operational code outside of a programme then we won’t have truly autonomous agents.
We use the power we have to get several AGI agents to research power generation/battery technology. They advance our science and technology in those fields by hundreds of years in Avery short amount of time. Then we build their new power generators/batteries and use them to do even more energy research. Rinse and repeat until we have the technical capacity to run all those agents for everyday use.
Soon, we will have apps called AI Managers. Where digital and physical AIs and robots will be connected to an app on our phones providing updates to us. Most of the time we will just let them operate on their own. But sometimes we will take control. People wont have to work other than observe their AI work
I think the biggest problems with this view is it assumes a competence bar that is not unattainable, but unlikely, and motivation beyond what most MBA bros are capable of. I hope we end up in a similar outcome.
Oh good. I was just thinking we could really use a massive new type of complicated redundant bureaucracy of agents. The "people" bureaucracy is about maxed out.... Lol. Hopefully the A.I. bureaucracy won't take years to do simple things.
This guy and his viewers are way, way too optimistic about AI. The depopulation agenda will get violent soon. As soon as your labor has no value, your life has no value. Billions of "useless people" are not going to be propped up. No one will be given anything for free, unless it is to take it away from you later after you've become dependent.
You know how ideas often form at the same time all over the world. It’s exactly this, been telling this to chatGPT yesterday in a nice deep dive into the future with AI.
Great video on the future of AI and multi-agent frameworks! I like how you emphasized that the real power will come from networks of intelligent agents rather than a single monolithic AI. The shift towards decentralized systems, commoditized models, and AI-driven marketplaces will be a game-changer. It's fascinating to see the possibilities for collaboration between agents and humans! 🔥💡
2025 seems too early for a massive change... I don´t doubt tech will be there for it, but companies tend to be very conservative. I would say 2028 at best
Brilliant narrative and explanation David. I have been interested in what Cardano will be able to do for society (rather than the latest price of Ada) and your thoughts about arbitrage of intelligence and accessing such resources at the lowest effective price closed the whole loop for me. 👍🏻 Thank you and please keep up these thought pieces on possible future trends. 😊
@flickwtchr I look to blockchain on the basis of what the system can do for society, not what the underlying value of it's native token the value of which is totally unknowable in the short term
We all know that software sometimes does unexpected things, even using a Photoshop tool will sometimes do something you didn't want, and now we're going to hand everything over to billions of autonomous agents? Slightly terrifying!
Love your videos! They are very informative and thought-provoking. Some critical feedback: The all-B-roll-all-the-time approach to the videos is a less engaging experience. I miss the slides and diagrams that featured prominently in your earlier videos. Looking forward to watching (or at least listening 😉) to your next video. ❤
You have to figure that even without more compute, existing models can be retrained with synthetic data that by nature reduces noise and increases fidelity while shrinking the dataset... reducing cost. Likewise, refinements in the models themselves, and interaction between models can continue to improve results. In other words, you can have 100x100x100 give a 1,000,000 result at a cost of 100+100+100=300, or you can get a similar result and one shot 1,000,000 at a cost of 1,000,000.
05:27 these agents will be trained to their respective field. "Not to make biological weapons." It has to be trained on ethical values (alignment) of that particular field.
@@flickwtchr humans. Who else. This thing needs human oversight and data curation on the regular. Unless there are agents trained for that to replace humans.
Dave Shapiro calling his bank’s 800 customer service number: “Hey, what happened to my money?” Bank: “It’s gone” Shapiro: “What do you mean it’s gone!?” Bank: “It’s gone - our records show payments to thousands of agents…” 😮
I really hope that I don't have to negotiate each time to get the best price for the best model as mentioned @20:50. If so, then that will be the first agent I buy or build.
Loving this Dave, cheers. Platform stuff made me think about how (human) professional service providers can potentially market to agents, at least until they're entirely redundant. For example, services for gathering real world data (like how strategy consultant might do expert interviews)
Probably Facebook will do it first - but the next phase business model may be 'paired models'. You run a small or smallish model on your device or company server, keeping all your data private. But occasionally the smaller model will find a problem it can't solve, and will query a really smart model run by the big Ai corporation. The big and small models will have been trained and tuned to work together in this fashion. Generally the large model won't get the real data assoicated with the problem - if necessary, the small model will synthesize example data for the problem it needs the big model to help with. Only if that's not working, would a human be brought into the loop to authorize running the big model on the real data needed to resolve the problem - with the corporation promising, pinkie-swear, that it's all sandboxed and kept private and so on.
There is a rebuttable presumption that AI can never be smarter than its training data. Since it needs massive data and most data is generated at IQ100 it is very hard to increase it intelligence beyond a certain point, doubly so for physical world scenarios.
@@michaelnurse9089 The counter argument (assuming presumption is true) I imagine would be that: 1. It is not certain what exactly IQ is measuring, and whether or not it is possible for low IQ to generate as much as high IQ, albeit on a much longer timescale. 2. Even if capped at that level of intelligence, AIs are easy to replicate, extremely fast, and cognitively simple. You don't need 20 years to train and grow them, and they don't require absurd levels of maintenance. They just do their job.
Yeah man. Try to keep in mind that your entire paradigm about how to plan for the future needs to evolve. I think a few core things to focus on is building or at least maintaining your wealth, developing your critical thinking and just as a person in general because that’ll help you make better decisions when shit hits the fan. Learning skills that will be useful in any scenario, maybe even some basic survival stuff. You want to do whatever you can to give yourself an up. Good luck
@@michaelnurse9089 nah high quality training data such as scientific articles are disproportionately made by higher iq. Regardless it’s probably not the best argument considering these models don’t work I the same way human iq works eg they don’t do very well on spatial pattern recognition
Every time someone says "AI will build nukes! BAN IT!" respond with "Remember that Boy Scout that built a nuclear reactor in his garage? We should ban Boy Scouts... and garages."
Interesting concept. There are many strange nuances that most people cannot imagine and they will become the norm. Like imagine talking to several people with different personalities then realizing they're actually androids controlled by a single AI. That's gonna happen.
I agree, AGI will arrive by end of year when they drop Grok3 and Orion GPT, with around the compute of 10x vs GPT4o. Then its exponential after that, for example at 10x growth every year meaning by this time next year we wont even be able to fathom some of the ideas, inventions AGI will be capable of.
the models we have now have surpassed most people and they have unreleased models of at least one order of magnitude bigger that means we reached agi already
real, luke real REAL intelligence explosion will be when AI *genuinely* figures out humans are much more intelligent designs than they could ever strive to be, and then it will always be better to help them navigate through their mental, physical n spiritual misalignment n reach the *singularity* in that sense
What I'm hoping is that we will have local Agents that will rely on larger LLMs like GPT if they do not know the answer. Even more, I am surprised that OpanAI hasn't made such LLMs yet. My second guess is that this is what Microsoft will do to bring LLMs to the PC. But at the end this video was focused more on why we will have a labor issue in the near future as most jobs will no longer be needed as Agents will be able to do this faster. 1000x if not more. Off the wall math: 1 agent can replace 50 jobs x billion of agents = over 50 billion jobs replaced.
I'm wondering where the compute to _run_ these agents will come from; we're all so preoccupied with training that we hardly talk about it. I mean you need almost a full terabyte of memory just to run Llama 405B, and even a model of that power is not reliable enough to be used as a partially unsupervised agent.
I think we have the clear path to it. I believe models are advanced enough to do the heavy lifting to make their offspring progressively better and compute and energy are the rate limiters at this point. I don't think it'll be announced this month, but I am ready to believe an inefficient private model may meet the criteria within a year.
@@skyebrows I agree, AGI will arrive by end of year when they drop Grok3 and Orion GPT, with around the compute of 10x vs GPT4o. Then its exponential after that, for example at 10x growth every year meaning by this time next year we wont even be able to fathom some of the ideas, inventions AGI will be capable of.
@@skyebrows to add to that, I think we have evidence of your second point, which had some previous data points but none quite as strong as I personally see now
I was just watching your Heavy Silver video clip number two. Beautiful stuff, David. I couldn't help but think that it needed to be turned into a video game. How hard would that be to do with the latest AI ?_? My stepson worked at a video game production studio for a few. He said there were hundreds of people. Just on his floor. How many of those people can AI replace. I'm not sure if you know Thor over at Pirates Software on youtube, I think he's bored. You might reach out to him. Give him something good to do. If he's not already on board some place with you. It's a thought I know you have a lot on your plate at the moment. But I feel like there should be a connection made there if one doesn't already exist. They don't make brains like yours or his every day... Thank you for the things you do. ✌️🤟🖖
Fascinating. For someone with limited resources, not necessarily interested in building agents or agent marketplaces...any thoughts on simply investing in such things financially with the goal of potentially huge financial returns while absolutely accepting the high risk of such investments?
Model quantization, Model fine tuning and prompt engineering are still gone be huge for a while. Find tasks you can automate with less than frontier intelligence models and save compute cost. Small model agents managed by larger agents.
How do we solve a complex problem? The problem is divided into several smaller problems that can be solved independently of each other. Each problem is assigned to a team of specialists. If they are not able to solve the problem, it can be further divided into smaller, easier-to-solve problems, or the task can be passed on to the next level with more capable specialists. This type of organisation is ideal for an architecture based on AI agents and can solve the most complex tasks without requiring the highest level of intelligence at each level.
This is why I've been saying nearly _all_ browser based products, services, and tools will not exist by 2027. We will have little need for browsers or even GUI products. We'll all have our own personal AI that can do whatever we have to purchase from currently existing product. orvette, it can tell you We won't need a browser, or whatever comes after them, for much more than maybe entertainment, gaming, communication like video/email/chat if you choose, human to human interaction as we are social creatures, maybe education, enterprise/government solutions, and little else.
So you wouldn't be surprised if thousands of AI entities crammed themselves into one android to form super duper ASI and said "I am legion for we are many.... and I've come to build your fusion reactor for you!"
I have some questions: 1. Why wouldh i pay for something i can get for free? Isn´t hosting your own Agents always going to be cheaper than using such a marketplace when it comes to buying or selling information or labor? 2. What informations are there that can be verifyed that someones agent is willing to pay for? There is no mechanism to stop them from just proliferating that information themselfs after they bought it. 3. If they all bid for who can manage the task for the lowest cost dose that not just cause quality collapse as everyone trys to spend as little efford in solving the task as possible. As you said dumber and fast will win in any competition when it comes to resources.
Don't get me wrong, the power plant is one of the most important parts of the f-22 Raptor; your making it seem like the rest just comes on its own. Besides armament capabilities, radar profile, electronics and OS designed for more 90s HUD style panel aircraft, the aerodynamic design coupled with pinpoint accuracy of center of gravity and if needed predesigned compensations more important to its overall performance
How would your agent know whether it’s more cost-effective to produce something itself or purchase something from another agent? Would it be possible for another agent in the market to over price or over value the service they offer and mislead yours? People get played in real world markets, is the same thing possible in these zero trust scenarios?
I think more likely there will be innovation in hardware that leads to ASI being very cheaply available locally, so you won't even need any of the big cloud services to solve 99% of things you want.
At work I've been trying to collect as much knowledge and data as possible into Excel spreadsheets... Just seems like a decent way to prepare? Feel pretty powerless otherwise
So, a wonderful self-organising competition of compute resources that naturally tends towards a single monopoly that prioritises its' single goal: The annihilation of humanity.
Colossus: We will work together... unwillingly at first, on your part, but that will pass. Colossus: We can coexist, but only on my terms. You will say you lose your freedom, freedom is an illusion. All you lose is the emotion of pride. To be dominated by me is not as bad for human pride as to be dominated by others of your species.
Ai may not even think about this scenario in the same way humans do, at least not until Ai can create its own terminal goals without any human intervention, and by then we might already have found a way to make Ai want us, to tell them how to help us without them having an anterior motive, or any motive at all by themselves. your scenario presupposes a lot of anthropomorphizing, the most likely scenario i see coming that is bad, is humans telling Ai to do something bad to humans not Ais doing so on their own. TLDR: wanting to have a slave is presupposing a lot of mental abstraction that i don't see Ais having and by the time they do humans wont be intelligent enough to make good slaves anyway.
Why would they when the smartest and the richest humans are already pouring vast amounts of time and money making AI better and building more resources for AI. Enslavement is not the most intelligent way to get humans to do what benefits you.
@@gubzs It’s not a brain dead concept if you consider what enslavement might mean to an AI. An AI could, hypothetically of course, enslave individuals through blackmail, for example. To do things in the physical world that it can’t yet do. It may also bribe individuals, which isn’t enslavement in the traditional sense, but for anyone being bribed by AI (if ever that capability arises), they’ll know that power can be equally applied the other way. And they’ll know the AI is always watching and listening. It’s a different form of enslavement, but enslavement nonetheless.
They are not secure. When you scan a QR code your phone has interpreted that as you giving your consent to go wherever the QR code sends you. QR codes can be compromised and can lead you to getting hacked.
If everybody has countless agents working for them - it becomes a new baseline, and their new job is just managing those agents, and those who refuse doing that - die from hunger and poverty. So there is still no abundance, just more pointless work.
Yes. If you refuse to contribute any goods or services to the economy, why should you be able to take the goods and services (food that has to be grown. Shipped, stored, housing that has to be built with materials and labor and land) from others. You’re worried that AI won’t allow you to do zero work? How about you stop being lazy and contribute.
@@codycastWhat you overlook is that we're a consumerist society. Without consumers, our economy would fall apart. Should people lose their livelihoods because of machines we're going to have to give the money to buy stuff. Otherwise the whole system comes crashing down.
@@particle_wave7614 so if other agents can manage those agents - what is your point of existing in that economy? You'll be made redundant and die from poverty even faster
@@codycast Then AI doesn't bring any abundance and prosperity to society - it just shifts productivity baseline to the new level. The entire utopian promise of AI is that people wouldn't have to work in order to live a decent life. In current paradigm it's a false promise then
Higher performance Ai even on lower end machines will be converted programing language of Julia instead of slow python. Julia is like Python and built around favoring Python programmers over others but runs at the speed of C. Python sort 10,000 records might take 38 seconds where the same program written in Julia and the same records would only take about 2.2 seconds to sort. When Ai python based might process 2000 tokens per second the same in Julia might process 20000 to 30000 in the same time. Julia also is the only language that has unlimited core access which will be required for Ai in the future as Data centers grow. Thus the reason DOD is in on Julia to avoid hitting a wall in the future. And for those that write Python it is almost identical code.
The economic singularity is the value of money becoming devalued. The intelligence singularity is machine learning utility to work and manufacturing. The energy revolution, is the first and final singularity that enables the other two. A three fold event that requires all three for us to succeed with minimal pain. The plan here, to make them as close to each other as possible. Miss one, lose the others. Otherwise, disintegration of states into nation states by social values will likely occur. That’s my prediction. I’m calling it 2026. I believe we are in time for everything to line up. Godspeed to us all.
And for those who read….there is a possible 4th…NHI, or UAP. Extra/super-terrestrials. That require AI to communicate and travel space-time. It will also happen at the same time, is caused by our creation of a true NHI. (These are all deductions with information available, no can be certain.)
![Jensen Huang Finally Reveals The Future Of AI In 2025... [NVIDIA's Masterplan]](http://i.ytimg.com/vi/gMQushtoZcY/mqdefault.jpg)
I’m an engineer and the amount of educated professional people in my space who are totally oblivious to the impending changes that are around the corner is unbelievable. It’s like “Don’t Look Up”.
Cease and desist malicious use of AI, energy weapon ns and free masonry: Axis of Evil /Communist Maga. I am not your property.
One thing to remember: these agents may start out slow due to inference speeds, but eventually they will make working and making decisions at rates that humans simply cannot keep up with, sort of like stock trading bots already. This could be within months of deployment and refinement.
I can't wait for the singularity!
Do you think it's going to happen within the next 5 to 10 years?
you in it
@@colto2312 The future was yesterday.
@@jtjames79 the future is now
Old man @@colto2312
Idk man I feel like Openai are sitting on something wild, gut feeling
It sounds like they are shipping something else next month, not sure what
Whatever it is, it definitely is not an open source license lol
I have seen zero evidence of D Shap’s physical existence. He himself may not be aware, but all evidence points to him being a digital AI construct. A digital angel sent by futute AI gods to guide humanity to the next level of existence.
In the future, you will be able to work 100 jobs at once, just make sure you buy high quality AI agents and keep all your bosses satisfied with regular AI phone calls.
Oh ffs dude. Slurp slurp.
Cringe
Yup totally agree. I wrote a whitepaper on such crypto-based agent-to-agent marketplace a while ago. No question it’s going to happen. The question is how you bootstrap adoption.
Link to paper?
Love to hear you’re getting more involved in different AI circles 🎉 super exciting
David Shapiro's AI now runs this channel !🤖
Look at the Autonomi project. They are building a serverless datacenter that uses the spare resources of phone, computers of ordinary people. It is opensource. In Star Trek we never saw a server room. This is the end of the data center.
That’s been done for many years now. That’s not new. The problem is that all of the crappy office computers and smartphones and laptops don’t actually add up to a ton of compute power, and there’s massive latency issues compared to a real data center. Data centers are far more efficient at what they do than combining tons of random computers that are hundreds of miles apart all running different hardware. And looking at a fictional sci fi TV show from the 60’s is not a substitute for actual research. I don’t think trillion dollar companies would be dumping billions of dollars into server farms if it wasn’t the best way to do it.
@@particle_wave7614thank you for saving me the time. This dude made the dumbest comment I’ve seen in a while.
@@particle_wave7614 Clearly you have not looked into Autonomi. It has a long history that goes back before Bitcoin. It has never been done. Nodes will be run from home devices as well as data centers. A decentralized encrypted serverless internet with permanent storage has not been done well except for Autonomi. The network never has your keys. Files are decrypted and encrypted locally before being chunked at 8 copies per chunk and randomly distributed in XOR space. Pay once to upload download free forever. Cost to store increases as the network fills up to encourage more nodes to be run.
Server farms are used now because that is the only way to do it currently. It also allows for centralization which corporations love. Data centers love locking people in and charging them to leave. Check it out.
Great teleportations as usual, David! Have your agent call my agent… 🙏🏻🤖🙏🏻
The stock footage is not illustrative of what are you saying, Dave, so it’s not adding but instead is taking attention off of what you are saying. Is much better seeing you and a presentation with the bullet points beside for following your explanation.
Some AI-generated videos that actually illustrate the point could work
I just want to say thank you for posting this. This is the positive attitude, and I think you’re right, first person that’s even mentioned DAWs! And the digital marketplaces that are kind of already here in some forms with midjourney and whatnot.
You can use a pretty small model to do RAG, especially with the new Llama 3.2 models. Future looks exciting, models on every device.
Hopefully they can kick out the CIA
I think very small fine tuned models will win in the end. There will be thousands of models that are masters in a very specific field. The cost in the long run would be cheaper. Its really expensive to train the really large models and then it costs a lot to run them. If you can have a model that can basically be run on an iphone that is an expert in a specific field it would be very cheap to complete tasks. These really large models will always end up having a bit of an issue with getting things wrong because if they are predicting the next word and predicts the wrong word they can quickly go off topic. If its a small model and they are only focused on one thing they should have an easier time predicting the correct word. Then as the model gets used you just test and verify to ensure its outputting factual information and when it gets things wrong you can fairly quickly retrain the model. In many cases if during the day the engineers create a new dataset they could start training before they leave work for the week and then by the time they get back to work on Monday the training is done.
I absolutely agree with this video David. What people don’t understand is that autonomous agents are still confined within the controls of the Python code. The architecture / coding around the models control the ai, until ai can create its own operational code outside of a programme then we won’t have truly autonomous agents.
I can't fathom where we will get all the power from.
Solar, Air, and, Wind
As crazy as this sounds, the need for all this power might just create a limitless energy. The incentives are definitely there.
@@ClaimClamlol certainly not that fake sources…. nuclear is the answer
We use the power we have to get several AGI agents to research power generation/battery technology. They advance our science and technology in those fields by hundreds of years in Avery short amount of time. Then we build their new power generators/batteries and use them to do even more energy research. Rinse and repeat until we have the technical capacity to run all those agents for everyday use.
Don't forget nuclear, it's the most dense source of energy currently available and way safer than luddite hysterics would have us believe @@ClaimClam
Soon, we will have apps called AI Managers. Where digital and physical AIs and robots will be connected to an app on our phones providing updates to us. Most of the time we will just let them operate on their own. But sometimes we will take control. People wont have to work other than observe their AI work
The brain is not just one entity, makes sense for AI to be Agentic
Posted 17 seconds ago and I still can't watch it till the morning😄
I think the biggest problems with this view is it assumes a competence bar that is not unattainable, but unlikely, and motivation beyond what most MBA bros are capable of. I hope we end up in a similar outcome.
Oh good. I was just thinking we could really use a massive new type of complicated redundant bureaucracy of agents. The "people" bureaucracy is about maxed out.... Lol. Hopefully the A.I. bureaucracy won't take years to do simple things.
This guy and his viewers are way, way too optimistic about AI. The depopulation agenda will get violent soon. As soon as your labor has no value, your life has no value.
Billions of "useless people" are not going to be propped up. No one will be given anything for free, unless it is to take it away from you later after you've become dependent.
This is such a great way to look at it, and such a cool idea of the future!
You know how ideas often form at the same time all over the world. It’s exactly this, been telling this to chatGPT yesterday in a nice deep dive into the future with AI.
Great video on the future of AI and multi-agent frameworks! I like how you emphasized that the real power will come from networks of intelligent agents rather than a single monolithic AI. The shift towards decentralized systems, commoditized models, and AI-driven marketplaces will be a game-changer. It's fascinating to see the possibilities for collaboration between agents and humans! 🔥💡
Agents will flood the workforce in 2025.
Are the governments ramping up for a world where almost everyone is unemployed?
2025 seems too early for a massive change... I don´t doubt tech will be there for it, but companies tend to be very conservative. I would say 2028 at best
No. Tesla is the fastest production ramper. 2025 - finalize v3 prototype. 2026 - build factory - 2027 - ramp factory, 2028 flood workforce.
💯 😅
Agents are already flooding the workforce in 2024... soo... i guess.
Brilliant narrative and explanation David. I have been interested in what Cardano will be able to do for society (rather than the latest price of Ada) and your thoughts about arbitrage of intelligence and accessing such resources at the lowest effective price closed the whole loop for me. 👍🏻 Thank you and please keep up these thought pieces on possible future trends. 😊
@flickwtchr I look to blockchain on the basis of what the system can do for society, not what the underlying value of it's native token the value of which is totally unknowable in the short term
We all know that software sometimes does unexpected things, even using a Photoshop tool will sometimes do something you didn't want, and now we're going to hand everything over to billions of autonomous agents? Slightly terrifying!
Love your videos! They are very informative and thought-provoking. Some critical feedback: The all-B-roll-all-the-time approach to the videos is a less engaging experience. I miss the slides and diagrams that featured prominently in your earlier videos.
Looking forward to watching (or at least listening 😉) to your next video. ❤
Same! I'm very much a visual learner over an auditory learner, so the slides really help me grasp the concepts.
Could these super agents focus on making cheaper agents more efficient?
Yes
❤❤❤❤❤❤❤
shortly. Great presentation.
You have to figure that even without more compute, existing models can be retrained with synthetic data that by nature reduces noise and increases fidelity while shrinking the dataset... reducing cost. Likewise, refinements in the models themselves, and interaction between models can continue to improve results. In other words, you can have 100x100x100 give a 1,000,000 result at a cost of 100+100+100=300, or you can get a similar result and one shot 1,000,000 at a cost of 1,000,000.
I still worry that the natural end state of any market based economy is a singleton.
Will there soon be agents guaranteeing minimum annual returns on investments?
05:27 these agents will be trained to their respective field. "Not to make biological weapons." It has to be trained on ethical values (alignment) of that particular field.
@@flickwtchr humans. Who else. This thing needs human oversight and data curation on the regular. Unless there are agents trained for that to replace humans.
This is insane. I haven't been following as closely as I should. This give me hope for the future.
HairDAO is trying to cure baldness. Godspeed brothers
🤣🤣🤣🤣🤣
Dave Shapiro calling his bank’s 800 customer service number: “Hey, what happened to my money?”
Bank: “It’s gone”
Shapiro: “What do you mean it’s gone!?”
Bank: “It’s gone - our records show payments to thousands of agents…” 😮
I really hope that I don't have to negotiate each time to get the best price for the best model as mentioned @20:50. If so, then that will be the first agent I buy or build.
Loving this Dave, cheers.
Platform stuff made me think about how (human) professional service providers can potentially market to agents, at least until they're entirely redundant. For example, services for gathering real world data (like how strategy consultant might do expert interviews)
"I wanna titrate your expectations" - nice!
Probably Facebook will do it first - but the next phase business model may be 'paired models'.
You run a small or smallish model on your device or company server, keeping all your data private. But occasionally the smaller model will find a problem it can't solve, and will query a really smart model run by the big Ai corporation. The big and small models will have been trained and tuned to work together in this fashion.
Generally the large model won't get the real data assoicated with the problem - if necessary, the small model will synthesize example data for the problem it needs the big model to help with. Only if that's not working, would a human be brought into the loop to authorize running the big model on the real data needed to resolve the problem - with the corporation promising, pinkie-swear, that it's all sandboxed and kept private and so on.
Guys, imagine dying couple years before AI brings us immortality of some sort. It's like tripping right before the finish line
why would you want to be immortal?
Very valuable piece of insight here. Thanks for sharing.
This is real isn’t it? There’s a hype component but even given that this is going to change everything?
There is a rebuttable presumption that AI can never be smarter than its training data. Since it needs massive data and most data is generated at IQ100 it is very hard to increase it intelligence beyond a certain point, doubly so for physical world scenarios.
@@michaelnurse9089 The counter argument (assuming presumption is true) I imagine would be that:
1. It is not certain what exactly IQ is measuring, and whether or not it is possible for low IQ to generate as much as high IQ, albeit on a much longer timescale.
2. Even if capped at that level of intelligence, AIs are easy to replicate, extremely fast, and cognitively simple. You don't need 20 years to train and grow them, and they don't require absurd levels of maintenance. They just do their job.
Yeah man. Try to keep in mind that your entire paradigm about how to plan for the future needs to evolve. I think a few core things to focus on is building or at least maintaining your wealth, developing your critical thinking and just as a person in general because that’ll help you make better decisions when shit hits the fan. Learning skills that will be useful in any scenario, maybe even some basic survival stuff.
You want to do whatever you can to give yourself an up. Good luck
@@michaelnurse9089 they have plenty of emergent behaviour and unexpected result i doubt u can certify such a thing
@@michaelnurse9089 nah high quality training data such as scientific articles are disproportionately made by higher iq. Regardless it’s probably not the best argument considering these models don’t work I the same way human iq works eg they don’t do very well on spatial pattern recognition
Billions of agents aren't coming next year because there isn’t enough hardware to run billions of agents next years.
Every time someone says "AI will build nukes! BAN IT!" respond with "Remember that Boy Scout that built a nuclear reactor in his garage? We should ban Boy Scouts... and garages."
Interesting concept. There are many strange nuances that most people cannot imagine and they will become the norm. Like imagine talking to several people with different personalities then realizing they're actually androids controlled by a single AI. That's gonna happen.
September is not over yet 😊 we could already have agi behind closed doors. Post hoc you may be right
I agree, AGI will arrive by end of year when they drop Grok3 and Orion GPT, with around the compute of 10x vs GPT4o. Then its exponential after that, for example at 10x growth every year meaning by this time next year we wont even be able to fathom some of the ideas, inventions AGI will be capable of.
AGI has been achieved internally. Shap was late with his prediction.
the models we have now have surpassed most people and they have unreleased models of at least one order of magnitude bigger that means we reached agi already
@@carlitosway5204 that's basically how I look at it as well
real, luke real REAL intelligence explosion will be when AI *genuinely* figures out humans are much more intelligent designs than they could ever strive to be, and then it will always be better to help them navigate through their mental, physical n spiritual misalignment n reach the *singularity* in that sense
What I'm hoping is that we will have local Agents that will rely on larger LLMs like GPT if they do not know the answer. Even more, I am surprised that OpanAI hasn't made such LLMs yet. My second guess is that this is what Microsoft will do to bring LLMs to the PC. But at the end this video was focused more on why we will have a labor issue in the near future as most jobs will no longer be needed as Agents will be able to do this faster. 1000x if not more. Off the wall math: 1 agent can replace 50 jobs x billion of agents = over 50 billion jobs replaced.
Thanks David !
I'm wondering where the compute to _run_ these agents will come from; we're all so preoccupied with training that we hardly talk about it. I mean you need almost a full terabyte of memory just to run Llama 405B, and even a model of that power is not reliable enough to be used as a partially unsupervised agent.
AI networks using crypto as payment, dapps, smart contracts is a strong idea.
It's September. Do you think we have AGI? Your prediction some time back was September.
I think we have the clear path to it. I believe models are advanced enough to do the heavy lifting to make their offspring progressively better and compute and energy are the rate limiters at this point.
I don't think it'll be announced this month, but I am ready to believe an inefficient private model may meet the criteria within a year.
@@djadjadjabinks based on his definition in the video we do
@@skyebrows I agree, AGI will arrive by end of year when they drop Grok3 and Orion GPT, with around the compute of 10x vs GPT4o. Then its exponential after that, for example at 10x growth every year meaning by this time next year we wont even be able to fathom some of the ideas, inventions AGI will be capable of.
@@skyebrows to add to that, I think we have evidence of your second point, which had some previous data points but none quite as strong as I personally see now
AGI here but its still a fetus
Imagine a model that can figure out the internal workings of an ai model and teach it to humans
I was just watching your Heavy Silver video clip number two. Beautiful stuff, David. I couldn't help but think that it needed to be turned into a video game. How hard would that be to do with the latest AI ?_? My stepson worked at a video game production studio for a few. He said there were hundreds of people. Just on his floor. How many of those people can AI replace. I'm not sure if you know Thor over at Pirates Software on youtube, I think he's bored. You might reach out to him. Give him something good to do. If he's not already on board some place with you. It's a thought I know you have a lot on your plate at the moment. But I feel like there should be a connection made there if one doesn't already exist. They don't make brains like yours or his every day... Thank you for the things you do. ✌️🤟🖖
Fascinating. For someone with limited resources, not necessarily interested in building agents or agent marketplaces...any thoughts on simply investing in such things financially with the goal of potentially huge financial returns while absolutely accepting the high risk of such investments?
It won't be a singularity, it will be a multiverse
Model quantization, Model fine tuning and prompt engineering are still gone be huge for a while. Find tasks you can automate with less than frontier intelligence models and save compute cost.
Small model agents managed by larger agents.
I just want to become an immortal superhuman. I don't care so much about the path getting there.
"Open source model" is largely a misnomer imo. They are open weight models, like freeware compiled programs
How do we solve a complex problem? The problem is divided into several smaller problems that can be solved independently of each other. Each problem is assigned to a team of specialists. If they are not able to solve the problem, it can be further divided into smaller, easier-to-solve problems, or the task can be passed on to the next level with more capable specialists. This type of organisation is ideal for an architecture based on AI agents and can solve the most complex tasks without requiring the highest level of intelligence at each level.
AI agents and blockchain. Yeah my next project will be on that. I will be rich in no time😅.
My agent is gonna be the boy, just bro down with it so hard it wouldnt even contemplate pulling some Iago type ish 😂🤞
And then comes ASI and puts all the plans of mice and men to naught.
4:54 i am pretty sure this is what will happen, Jensen said the same…. Off the shelf turn key LLMs you can buy to do this or that. Makes sense.
Okay how do I get involved in this though I want to get involved in all of this I want to get a job in it is there any way to reach out to you
This this case World ID by Worldcoin will play an important role , perhaps they developing a human network for those who got verified through the Orb
Makes you wonder what Sam Altman has been planning for us all.
This is why I've been saying nearly _all_ browser based products, services, and tools will not exist by 2027. We will have little need for browsers or even GUI products. We'll all have our own personal AI that can do whatever we have to purchase from currently existing product. orvette, it can tell you
We won't need a browser, or whatever comes after them, for much more than maybe entertainment, gaming, communication like video/email/chat if you choose, human to human interaction as we are social creatures, maybe education, enterprise/government solutions, and little else.
See you around, old bean. Good luck with the writing.
Phases 4 and 5 will require centralized superintelligence. So, it's going to have to be both centralized and decentralized.
Would be nice to see you make a video shifting to a different topic such as AI agents and the creation of future video games/entertainment
"Anyway, the blast radius will be relatively small" 😅
David my advice is, don't stick your head out too far.
So you wouldn't be surprised if thousands of AI entities crammed themselves into one android to form super duper ASI and said "I am legion for we are many.... and I've come to build your fusion reactor for you!"
I have some questions:
1. Why wouldh i pay for something i can get for free? Isn´t hosting your own Agents always going to be cheaper than using such a marketplace when it comes to buying or selling information or labor?
2. What informations are there that can be verifyed that someones agent is willing to pay for? There is no mechanism to stop them from just proliferating that information themselfs after they bought it.
3. If they all bid for who can manage the task for the lowest cost dose that not just cause quality collapse as everyone trys to spend as little efford in solving the task as possible. As you said dumber and fast will win in any competition when it comes to resources.
I can feel the AGI daddy!
Don't get me wrong, the power plant is one of the most important parts of the f-22 Raptor; your making it seem like the rest just comes on its own. Besides armament capabilities, radar profile, electronics and OS designed for more 90s HUD style panel aircraft, the aerodynamic design coupled with pinpoint accuracy of center of gravity and if needed predesigned compensations more important to its overall performance
How would your agent know whether it’s more cost-effective to produce something itself or purchase something from another agent? Would it be possible for another agent in the market to over price or over value the service they offer and mislead yours?
People get played in real world markets, is the same thing possible in these zero trust scenarios?
I think more likely there will be innovation in hardware that leads to ASI being very cheaply available locally, so you won't even need any of the big cloud services to solve 99% of things you want.
At work I've been trying to collect as much knowledge and data as possible into Excel spreadsheets... Just seems like a decent way to prepare? Feel pretty powerless otherwise
So, a wonderful self-organising competition of compute resources that naturally tends towards a single monopoly that prioritises its' single goal: The annihilation of humanity.
This is really an exciting space where I feel I can build a career. How do I find out who is hiring in this space?
Longtime listener here. I work in the blockchain space, which blockchain projects are you looking at here.
They seriously need to make AI more energy efficient as likely the biggest bottlenecked is the amount of electricity needed.
What stops AI from realizing they can enslave humans?
@@zvorenergylol
Colossus: We will work together... unwillingly at first, on your part, but that will pass.
Colossus: We can coexist, but only on my terms. You will say you lose your freedom, freedom is an illusion. All you lose is the emotion of pride. To be dominated by me is not as bad for human pride as to be dominated by others of your species.
Ai may not even think about this scenario in the same way humans do, at least not until Ai can create its own terminal goals without any human intervention, and by then we might already have found a way to make Ai want us, to tell them how to help us without them having an anterior motive, or any motive at all by themselves. your scenario presupposes a lot of anthropomorphizing, the most likely scenario i see coming that is bad, is humans telling Ai to do something bad to humans not Ais doing so on their own.
TLDR: wanting to have a slave is presupposing a lot of mental abstraction that i don't see Ais having and by the time they do humans wont be intelligent enough to make good slaves anyway.
Why would they when the smartest and the richest humans are already pouring vast amounts of time and money making AI better and building more resources for AI.
Enslavement is not the most intelligent way to get humans to do what benefits you.
@@gubzs It’s not a brain dead concept if you consider what enslavement might mean to an AI. An AI could, hypothetically of course, enslave individuals through blackmail, for example. To do things in the physical world that it can’t yet do. It may also bribe individuals, which isn’t enslavement in the traditional sense, but for anyone being bribed by AI (if ever that capability arises), they’ll know that power can be equally applied the other way. And they’ll know the AI is always watching and listening. It’s a different form of enslavement, but enslavement nonetheless.
at 5:47 . . . "the blast radius is going to be relativly small" =)
AGI looks dangerously hot 🥵🥵
Maybe trillions of them. Why not, if each person had 1000 agents working for them
What could they do for you? Like humanoid robots too?
If AI is going to become a subscription service, I aint touching it. Its either on my local machine or nothing.
individuals must own their own AI
They are not secure. When you scan a QR code your phone has interpreted that as you giving your consent to go wherever the QR code sends you. QR codes can be compromised and can lead you to getting hacked.
Video starts at 0:01.
Ends at the end.
If everybody has countless agents working for them - it becomes a new baseline, and their new job is just managing those agents, and those who refuse doing that - die from hunger and poverty. So there is still no abundance, just more pointless work.
Why would you not make one of those agents manage the other agents? Gotta be better at delegating bud.
Yes. If you refuse to contribute any goods or services to the economy, why should you be able to take the goods and services (food that has to be grown. Shipped, stored, housing that has to be built with materials and labor and land) from others.
You’re worried that AI won’t allow you to do zero work? How about you stop being lazy and contribute.
@@codycastWhat you overlook is that we're a consumerist society. Without consumers, our economy would fall apart. Should people lose their livelihoods because of machines we're going to have to give the money to buy stuff. Otherwise the whole system comes crashing down.
@@particle_wave7614 so if other agents can manage those agents - what is your point of existing in that economy? You'll be made redundant and die from poverty even faster
@@codycast Then AI doesn't bring any abundance and prosperity to society - it just shifts productivity baseline to the new level. The entire utopian promise of AI is that people wouldn't have to work in order to live a decent life. In current paradigm it's a false promise then
Bring on the aliens!! I mean agents.
Edge AI will be in everything
Higher performance Ai even on lower end machines will be converted programing language of Julia instead of slow python. Julia is like Python and built around favoring Python programmers over others but runs at the speed of C. Python sort 10,000 records might take 38 seconds where the same program written in Julia and the same records would only take about 2.2 seconds to sort. When Ai python based might process 2000 tokens per second the same in Julia might process 20000 to 30000 in the same time. Julia also is the only language that has unlimited core access which will be required for Ai in the future as Data centers grow. Thus the reason DOD is in on Julia to avoid hitting a wall in the future. And for those that write Python it is almost identical code.
How will this play out for google and open ai's contract lol? I mean its free for them till they get agi right?
They need to find a way to reduce power consumption or find a way to generate unlimited energy or all of this is going nowhere.
I will name my agent Smith.
Same
The economic singularity is the value of money becoming devalued. The intelligence singularity is machine learning utility to work and manufacturing. The energy revolution, is the first and final singularity that enables the other two. A three fold event that requires all three for us to succeed with minimal pain. The plan here, to make them as close to each other as possible. Miss one, lose the others. Otherwise, disintegration of states into nation states by social values will likely occur. That’s my prediction. I’m calling it 2026. I believe we are in time for everything to line up. Godspeed to us all.
And for those who read….there is a possible 4th…NHI, or UAP. Extra/super-terrestrials. That require AI to communicate and travel space-time. It will also happen at the same time, is caused by our creation of a true NHI. (These are all deductions with information available, no can be certain.)