Thank you! I plan to release a video every other week so it gives me some time to research the topic area in depth. The next video should hopefully be out by the 21st. I appreciate the support.
Nice film. Don't forget that, while AI is advancing a lot, robots are not advancing as quickly. So skilled manual labour is likely to be the preserve of humans for some time to come. I can't really imagine hiring a robot plumber any time soon.
Thank you! That is a fair point. The advancement in robotics is slower than generative AI due to the complexity of modelling human senses and how they react to the inputs they receive. However, at the pace that robotics is advancing, like Tesla’s Optimus bots, it makes me wonder how long we will have until we start to see AI taking on more and more real world actions.
Coming sooner than you think. Genesis was likely the biggest advancement of 2024, on par at least with OpenAI o3 and other major developments. The ability to give a robot 430,000 hours of physics accurate simulated world training and experience in ONE HOUR of run time on a graphics card you can run at HOME was an Earth shaking advance in robotics, all on FREE open source software anyone can download and run: th-cam.com/video/IAmrSaDW88I/w-d-xo.html At $16K, and combined with Genesis above, I think humanoid robots are advancing just fine. th-cam.com/video/y2KF2DnvN9Y/w-d-xo.html But, yes, due to the variability of working on plumbing in ever changing scenarios in a home, robotic plumbers making house calls will be in the latter tiers of deployment, but they WILL COME.
1. You're WRONG about a "shift away from AGI" and it being "hype" and a "lot of money". First off, Deepseek V3 showed how to build a model that COMPETES with a frontier model, but use 1/20th the COST OF TRAINING. That means for the LOW LOW bargain cost of $5.6M they were able to MATCH a frontier model that cost $111M to train. Simply going the OPPOSITE WAY with the math that V3 showed us means that you COULD train the equivalent of a $1T model for ONLY $50M - 1/20th of the cost (and HALF the cost of GPT-4). However, it is unlikely it would scale in the inverse that far despite the potential for enough data to be available through RL. But the reality is that it SHOWED how we can CUT COSTS BY 95% and get a SIMILARLY CAPABLE MODEL. Next, R-Star-MATH just showed how a mere 7 billion parameter *SMALL* LANGUAGE MODEL can BEAT a 1.7 TRILLION PARAMETER FRONTIER MODEL (GPT-4) at math. It also beat OpenAI o1 preview at math. If you are unaware, a 7B model can be RUN ON A MODEST PC AT HOME, whereas GPT-4 needs a host of data center GPUs and lots of memory. We'll have AGI in late 2025 or early 2026 and MANY of the top experts are in agreement on this. 2. We absolutely HAVE seen the "next big leap" or "GPT-5" (despite it NOT being named as such). GPT-4 was followed by o1, a NEW NAMING SCHEME due to THINKING MODELS being introduced in September (would have BEEN GPT-5 had there NOT been good reasons to use a new naming convention). And then JUST 3 MONTHS LATER, o3 was announced in some cases DOUBLING the performance of o1 (which would have been GPT-6 if not for the name change). The "new things" Sam mentions in your video snippet is WHAT IS PLAYING OUT with the "o" series. So, NO, there will NOT be a "shift away" from AI development. Not in a world where every HOUR of the day, 7 new scientific papers get published, and when you just get a night's sleep, you have FALLEN 48 published papers of AI advancement BEHIND, since there are 5,000 scientific papers published EVERY MONTH. AI progress will proceed JUST FINE as Agents get released in 2025. 3. NO, prices will NOT continue to RISE for access to models, as above Deepseek V3 showed how to CUT THE COST of training a frontier model by 95%. Deepseek is ALREADY AVAILABLE to the public and TONS OF USERS ARE FLOCKING TO ITS LOWER PRICE POINT PER TOKEN. As this becomes a normal best practice (*NO ONE will spend $111M to train a model that COULD have cost them $5.6M using what V3 leverage), access for end users will CONTINUE TO FALL IN PRICE due to COMPETITION, except for the VERY high end "thinking" models that use large amounts of inference compute time (which is what the ChatGPT Pro $199/mo. plan offers from OpenAI, and is DIRT CHEAP to a researcher who is, for example, being paid $150-200,000/yr. or more while trying to CURE CANCER). The GENERAL PUBLIC would almost certainly get ZERO benefit out of access to a model like ChatGPT Pro as they are NOT doing cutting edge research in a field based upon their PHD. No one needs a model that advanced to ask AI the questions that the general public is asking. ALL of the $20/mo. models are well beyond "high school level". And NO, there ARE Open Source models that DO compete well with frontier models, like Llama: "Meta's Llama 3.1 405B is currently the most capable open-source model competing against frontier models. It is described as the "first frontier-level open source AI model" and is competitive with leading foundation models across a range of tasks, including GPT-4, GPT-4o, and Claude 3.5 Sonnet" 4. Stating that 750,000 robots are in use at Amazon is HUGELY MISLEADING. First of all, the use of Digit, a humanoid robot is ONLY IN THE TESTING PHASE. That 750K number of robots exists ONLY in the form of Sequoia, which is NOT even a "robot" in the the sense of the word that most people will think of. It is an AUTOMATED INVENTORY MANAGEMENT SYSTEM, with conveyor belts, bar code readers, etc.: th-cam.com/video/c55fSE-yd7I/w-d-xo.html 5. You are simply WRONG about the ability of AI to create, to imagine things unimagined before. Genex can synthesize a navigable world from a single IMAGE. Genie 2 can create a "playable game world" from a single image, where it understands the physics that would be in effect. Sora and other AI video generators as well as a half dozen image creation AIs create incredibly details fantastical worlds. AI is on a CLEAR TRACK to SURPASS HUMAN KNOWLEDGE and be a MULTIPLE of the knowledge of the entire human race. There will ultimately be NO humans that could even invent something CLOSE to what super intelligent AI will create for us. This commentary was BADLY researched and often simply dead wrong or misleading.
@@brianmi40 Firstly, I just wanted to thank you for taking the time to give a detailed insight into the current state of AI. It is progressing at an exponential pace, and you can’t wait too long until something groundbreaking comes along, especially in the open-source space, thanks to the ongoing work of developers and researchers making their work publicly accessible. You’ve highlighted some interesting points of consideration that I’ll take on board going forward, but I feel that you may have misinterpreted what I was saying towards the end. My intention wasn’t to say that AI can’t produce the unimagined but that humans can’t. We take inspiration from varying, seemingly unrelated sources of information to create our work, which makes it exciting. I asked an AI to write the script for this video, and it did a great job explaining how AI will progress, but at no point did it think to mention George Lucas, Star Wars, or the role of humankind. This uncanny ability to be inspired by seemingly unrelated sources allows us to invent and come up with the unordinary. Sure, we may have machines that are more intelligent than we are, but I believe that it’ll be on us to bring this intelligence together to solve the problems of tomorrow.
Legacy knowledge based companies will be wiped out by new entrants who go digital agentic first from the ground up by those who leave the legacy companies and team up with technology experts.
@ I am much more pessimistic about the upcoming times. The current capitalistic system and Extreme changes due to AGI are going to be catastrophic in the short term for general public. I would love to discuss these issues with you. we plebs gotta stick together, because the power is going to consolidate even further.
While a vast number of models and research are publicly available, a majority of the models that the average individual interacts with are centralised to just a handful of companies. It makes me question whether or not this data being collected through our interactions to improve the models, will create a higher barrier to entry for other AI companies and indviduals, ensuring that the existing giants in the industry remain at the top. Feel free to reach out to me on LinkedIn, I'd be happy to discuss further and hear your thoughts on the current state of AI.
It's frustrating listening to a video wherein someone with little understanding of technology discusses technology. I stopped at 4:23 because you are dead wrong about advances in LLMs. For example, the reason that generative pre-trained transformers (GPT) are expensive to train is that their training time is quadratic and based on the input context window. Newer architectures, such as mamba reduce this to a linear time drastically reducing training time and cost. I hate to be harsh but everyone and their brother is now putting out AI videos, with little value to add, all in an attempt to get clicks.
Hi, thank you for sharing your thoughts and providing detailed feedback, I appreciate constructive criticism. The goal of my videos is for me to better explore and understand the rapidly evolving AI space by sharing the lessons I’m learning with others. I aim to justify my conclusions in research, and at 04:23 in the video, I discuss the relationship between increasing parameter sizes and reducing loss. This was based on findings like those presented in the OpenAI paper, Scaling Laws for Neural Language Models arxiv.org/pdf/2001.08361. The paper highlights that as resources (compute, data, and parameters) increase, loss decreases, but the improvements follow a pattern of diminishing returns. This was the reason for my conclusion about the escalating costs of training larger models while achieving marginally smaller gains.
There is a flaw in the thinking. PhD level is not relevant or needed. The global economy is 2% PhDs. AI is currently better than essentially all of the workforce and agentic systems will only enable white collar work to be performed entirely over the next 2-5 years, probably a lot less. Compute and power are not problems, that is very short term thinking. There is another trillion fold improvement coming between now and late 2030s and a lot of it will be software and mathematical gains, 100s of times over the next 5 years. The next computing photonics chips will blow Nvidia out of the water over following the next 5 years of development and another ten years of production scaling and technology efficiency gains at a fraction of the power consumption. Nvidia will reign supreme for the next 15-20 years max.
That’s an interesting point. It seems though that regardless of how good our generative AI gets, there’s this tradeoff where the more it generalises, the less performant it becomes on individual tasks. Therefore, in order to have a model solve a specific challenge, it needs domain expertise on that specific problem, which is why I suggested PhDs and research being of more significance, since people can now develop their specific research area and then guide a model in understanding that area itself. I don’t necessarily mean everyone will be entering into academia, but I feel companies will start to have ongoing R&D as well as think tanks within their businesses, as the need to do the traditional process-driven tasks falls. Past a certain point in training, it also seems like throwing more compute at the model yields marginally less results, which is why I feel that there needs to be this focus on higher-quality data that the models are trained on. At the pace that technology is improving, however, I’m sure these advancements will be realised a lot quicker than we expect.
@@krishparikh1 Keep in mind that we live in a world with a $100 Trillion economy. Remember, money is not relevant anymore. If they spent $1 Trillion on compute and it got more output it would be worth it. There is so much growth in software and hardware improvements in AI, we will see a trillion fold improvement over the next 15 years or so. Go read the research papers there are thousands coming out each month. Remember, there is probably $40 Trillion to $60 Trillion of that is knowledge and efficiency driven, meaning AI agentic systems could automate and expand knowledge work tasks and enhance physical processes and product design and development. Imagine a world where a design for a product is automated autonomously and a production process and equipment designed and then built and assembled by humans and eventually robots (2040s or sooner). AGI is all about job automation, systems, and organizational workflows. SGI is all about PhD and new scientific research and expanding discoveries, establishing new materials and uses, research frontiers (not the 98% of work that is done by essentially the regular people of the world). AGI will automate current types of work while SGI will make new work (for machines) that grows/expands the economy output and differentiates the output to more complex levels never achieved before. Humans will be in the loop, then humans will not be able to be in the loop anymore as the machines will be far to intelligent and to fast. We will barely be able to figure out what it did in one moment in time. AI will be better, faster, cheaper, and more valuable and useful for economic and management of work. We will likely have most redundant repetitive jobs automated over the next 10 years. We will have organizations ran by AI systems. We will have breakthroughs in science and medicine ongoing. Humans then can relax and enjoy life and think and learn and have peace, though this will the rest of the century process to fully happen, probably.
That’s an interesting point. I hadn’t thought about the irrelevance of money given the inflationary world we are in today. So, investment in training the next AI to improve productivity gains would be justified, as it becomes “cheaper” over time. Once AI takes on redundant, repetitive jobs and humans can “relax and enjoy life”, I do wonder what this looks like. Part of the enjoyment comes from fulfilment in what we do. So, I feel that we would still continue to work, just not in the traditional sense of being paid to do repetitive tasks, but instead create economic value in other ways through creating stuff.
@@krishparikh1 Just think of the one billion jobs that will be automated over the next 10 years or less. More than likely the economy will double over the next 25 years due to AI, AEVs, drones, humanoid robotics. The ROI on that alone, for white collar type jobs is worth ten trillion plus efficiency gains of long term economic value and outside of agentic system advancements, not really based on gains in improvement, but rather adding functionality like multi-modal and action based learning to automate digital processes.
10:55 “Employees will have their roles shift from producing work to evaluating work.” That is powerful.
Thanks!
a very nicely structured video, so waiting eager for the next video!! what frequency will you publish from now on over this AI/human life,🤔 thread?
Thank you! I plan to release a video every other week so it gives me some time to research the topic area in depth. The next video should hopefully be out by the 21st. I appreciate the support.
Nice and truthfull Insight
Thanks for watching!
Nice film. Don't forget that, while AI is advancing a lot, robots are not advancing as quickly. So skilled manual labour is likely to be the preserve of humans for some time to come. I can't really imagine hiring a robot plumber any time soon.
not long and at the end of the day it's mind on silicon
Thank you! That is a fair point. The advancement in robotics is slower than generative AI due to the complexity of modelling human senses and how they react to the inputs they receive. However, at the pace that robotics is advancing, like Tesla’s Optimus bots, it makes me wonder how long we will have until we start to see AI taking on more and more real world actions.
Coming sooner than you think. Genesis was likely the biggest advancement of 2024, on par at least with OpenAI o3 and other major developments. The ability to give a robot 430,000 hours of physics accurate simulated world training and experience in ONE HOUR of run time on a graphics card you can run at HOME was an Earth shaking advance in robotics, all on FREE open source software anyone can download and run:
th-cam.com/video/IAmrSaDW88I/w-d-xo.html
At $16K, and combined with Genesis above, I think humanoid robots are advancing just fine.
th-cam.com/video/y2KF2DnvN9Y/w-d-xo.html
But, yes, due to the variability of working on plumbing in ever changing scenarios in a home, robotic plumbers making house calls will be in the latter tiers of deployment, but they WILL COME.
@@brianmi40 That's what I wanted to say
1. You're WRONG about a "shift away from AGI" and it being "hype" and a "lot of money".
First off, Deepseek V3 showed how to build a model that COMPETES with a frontier model, but use 1/20th the COST OF TRAINING. That means for the LOW LOW bargain cost of $5.6M they were able to MATCH a frontier model that cost $111M to train.
Simply going the OPPOSITE WAY with the math that V3 showed us means that you COULD train the equivalent of a $1T model for ONLY $50M - 1/20th of the cost (and HALF the cost of GPT-4). However, it is unlikely it would scale in the inverse that far despite the potential for enough data to be available through RL. But the reality is that it SHOWED how we can CUT COSTS BY 95% and get a SIMILARLY CAPABLE MODEL.
Next, R-Star-MATH just showed how a mere 7 billion parameter *SMALL* LANGUAGE MODEL can BEAT a 1.7 TRILLION PARAMETER FRONTIER MODEL (GPT-4) at math. It also beat OpenAI o1 preview at math. If you are unaware, a 7B model can be RUN ON A MODEST PC AT HOME, whereas GPT-4 needs a host of data center GPUs and lots of memory.
We'll have AGI in late 2025 or early 2026 and MANY of the top experts are in agreement on this.
2. We absolutely HAVE seen the "next big leap" or "GPT-5" (despite it NOT being named as such).
GPT-4 was followed by o1, a NEW NAMING SCHEME due to THINKING MODELS being introduced in September (would have BEEN GPT-5 had there NOT been good reasons to use a new naming convention). And then JUST 3 MONTHS LATER, o3 was announced in some cases DOUBLING the performance of o1 (which would have been GPT-6 if not for the name change).
The "new things" Sam mentions in your video snippet is WHAT IS PLAYING OUT with the "o" series.
So, NO, there will NOT be a "shift away" from AI development. Not in a world where every HOUR of the day, 7 new scientific papers get published, and when you just get a night's sleep, you have FALLEN 48 published papers of AI advancement BEHIND, since there are 5,000 scientific papers published EVERY MONTH.
AI progress will proceed JUST FINE as Agents get released in 2025.
3. NO, prices will NOT continue to RISE for access to models, as above Deepseek V3 showed how to CUT THE COST of training a frontier model by 95%. Deepseek is ALREADY AVAILABLE to the public and TONS OF USERS ARE FLOCKING TO ITS LOWER PRICE POINT PER TOKEN.
As this becomes a normal best practice (*NO ONE will spend $111M to train a model that COULD have cost them $5.6M using what V3 leverage), access for end users will CONTINUE TO FALL IN PRICE due to COMPETITION, except for the VERY high end "thinking" models that use large amounts of inference compute time (which is what the ChatGPT Pro $199/mo. plan offers from OpenAI, and is DIRT CHEAP to a researcher who is, for example, being paid $150-200,000/yr. or more while trying to CURE CANCER). The GENERAL PUBLIC would almost certainly get ZERO benefit out of access to a model like ChatGPT Pro as they are NOT doing cutting edge research in a field based upon their PHD. No one needs a model that advanced to ask AI the questions that the general public is asking. ALL of the $20/mo. models are well beyond "high school level".
And NO, there ARE Open Source models that DO compete well with frontier models, like Llama:
"Meta's Llama 3.1 405B is currently the most capable open-source model competing against frontier models. It is described as the "first frontier-level open source AI model" and is competitive with leading foundation models across a range of tasks, including GPT-4, GPT-4o, and Claude 3.5 Sonnet"
4. Stating that 750,000 robots are in use at Amazon is HUGELY MISLEADING. First of all, the use of Digit, a humanoid robot is ONLY IN THE TESTING PHASE. That 750K number of robots exists ONLY in the form of Sequoia, which is NOT even a "robot" in the the sense of the word that most people will think of. It is an AUTOMATED INVENTORY MANAGEMENT SYSTEM, with conveyor belts, bar code readers, etc.:
th-cam.com/video/c55fSE-yd7I/w-d-xo.html
5. You are simply WRONG about the ability of AI to create, to imagine things unimagined before. Genex can synthesize a navigable world from a single IMAGE. Genie 2 can create a "playable game world" from a single image, where it understands the physics that would be in effect. Sora and other AI video generators as well as a half dozen image creation AIs create incredibly details fantastical worlds.
AI is on a CLEAR TRACK to SURPASS HUMAN KNOWLEDGE and be a MULTIPLE of the knowledge of the entire human race. There will ultimately be NO humans that could even invent something CLOSE to what super intelligent AI will create for us.
This commentary was BADLY researched and often simply dead wrong or misleading.
@@brianmi40 Firstly, I just wanted to thank you for taking the time to give a detailed insight into the current state of AI. It is progressing at an exponential pace, and you can’t wait too long until something groundbreaking comes along, especially in the open-source space, thanks to the ongoing work of developers and researchers making their work publicly accessible.
You’ve highlighted some interesting points of consideration that I’ll take on board going forward, but I feel that you may have misinterpreted what I was saying towards the end. My intention wasn’t to say that AI can’t produce the unimagined but that humans can’t. We take inspiration from varying, seemingly unrelated sources of information to create our work, which makes it exciting. I asked an AI to write the script for this video, and it did a great job explaining how AI will progress, but at no point did it think to mention George Lucas, Star Wars, or the role of humankind. This uncanny ability to be inspired by seemingly unrelated sources allows us to invent and come up with the unordinary. Sure, we may have machines that are more intelligent than we are, but I believe that it’ll be on us to bring this intelligence together to solve the problems of tomorrow.
Legacy knowledge based companies will be wiped out by new entrants who go digital agentic first from the ground up by those who leave the legacy companies and team up with technology experts.
You have gained a lifetime follower. I would love to meet you. I think the same.
@@Page_max Thank you so much! I really appreciate this message, it has made my day.
@ I am much more pessimistic about the upcoming times.
The current capitalistic system and Extreme changes due to AGI are going to be catastrophic in the short term for general public.
I would love to discuss these issues with you.
we plebs gotta stick together, because the power is going to consolidate even further.
While a vast number of models and research are publicly available, a majority of the models that the average individual interacts with are centralised to just a handful of companies. It makes me question whether or not this data being collected through our interactions to improve the models, will create a higher barrier to entry for other AI companies and indviduals, ensuring that the existing giants in the industry remain at the top.
Feel free to reach out to me on LinkedIn, I'd be happy to discuss further and hear your thoughts on the current state of AI.
@@krishparikh1 Ok
It's frustrating listening to a video wherein someone with little understanding of technology discusses technology. I stopped at 4:23 because you are dead wrong about advances in LLMs. For example, the reason that generative pre-trained transformers (GPT) are expensive to train is that their training time is quadratic and based on the input context window. Newer architectures, such as mamba reduce this to a linear time drastically reducing training time and cost. I hate to be harsh but everyone and their brother is now putting out AI videos, with little value to add, all in an attempt to get clicks.
Hi, thank you for sharing your thoughts and providing detailed feedback, I appreciate constructive criticism. The goal of my videos is for me to better explore and understand the rapidly evolving AI space by sharing the lessons I’m learning with others.
I aim to justify my conclusions in research, and at 04:23 in the video, I discuss the relationship between increasing parameter sizes and reducing loss. This was based on findings like those presented in the OpenAI paper, Scaling Laws for Neural Language Models arxiv.org/pdf/2001.08361. The paper highlights that as resources (compute, data, and parameters) increase, loss decreases, but the improvements follow a pattern of diminishing returns. This was the reason for my conclusion about the escalating costs of training larger models while achieving marginally smaller gains.
@krishparikh1 the paper you cite is from 5 years ago. The technology is moving so fast that papers dated 5 months ago can be outdated.
There is a flaw in the thinking. PhD level is not relevant or needed. The global economy is 2% PhDs. AI is currently better than essentially all of the workforce and agentic systems will only enable white collar work to be performed entirely over the next 2-5 years, probably a lot less. Compute and power are not problems, that is very short term thinking. There is another trillion fold improvement coming between now and late 2030s and a lot of it will be software and mathematical gains, 100s of times over the next 5 years. The next computing photonics chips will blow Nvidia out of the water over following the next 5 years of development and another ten years of production scaling and technology efficiency gains at a fraction of the power consumption. Nvidia will reign supreme for the next 15-20 years max.
That’s an interesting point. It seems though that regardless of how good our generative AI gets, there’s this tradeoff where the more it generalises, the less performant it becomes on individual tasks. Therefore, in order to have a model solve a specific challenge, it needs domain expertise on that specific problem, which is why I suggested PhDs and research being of more significance, since people can now develop their specific research area and then guide a model in understanding that area itself. I don’t necessarily mean everyone will be entering into academia, but I feel companies will start to have ongoing R&D as well as think tanks within their businesses, as the need to do the traditional process-driven tasks falls.
Past a certain point in training, it also seems like throwing more compute at the model yields marginally less results, which is why I feel that there needs to be this focus on higher-quality data that the models are trained on. At the pace that technology is improving, however, I’m sure these advancements will be realised a lot quicker than we expect.
@@krishparikh1 Keep in mind that we live in a world with a $100 Trillion economy. Remember, money is not relevant anymore. If they spent $1 Trillion on compute and it got more output it would be worth it. There is so much growth in software and hardware improvements in AI, we will see a trillion fold improvement over the next 15 years or so. Go read the research papers there are thousands coming out each month. Remember, there is probably $40 Trillion to $60 Trillion of that is knowledge and efficiency driven, meaning AI agentic systems could automate and expand knowledge work tasks and enhance physical processes and product design and development. Imagine a world where a design for a product is automated autonomously and a production process and equipment designed and then built and assembled by humans and eventually robots (2040s or sooner). AGI is all about job automation, systems, and organizational workflows. SGI is all about PhD and new scientific research and expanding discoveries, establishing new materials and uses, research frontiers (not the 98% of work that is done by essentially the regular people of the world). AGI will automate current types of work while SGI will make new work (for machines) that grows/expands the economy output and differentiates the output to more complex levels never achieved before. Humans will be in the loop, then humans will not be able to be in the loop anymore as the machines will be far to intelligent and to fast. We will barely be able to figure out what it did in one moment in time. AI will be better, faster, cheaper, and more valuable and useful for economic and management of work. We will likely have most redundant repetitive jobs automated over the next 10 years. We will have organizations ran by AI systems. We will have breakthroughs in science and medicine ongoing. Humans then can relax and enjoy life and think and learn and have peace, though this will the rest of the century process to fully happen, probably.
That’s an interesting point. I hadn’t thought about the irrelevance of money given the inflationary world we are in today. So, investment in training the next AI to improve productivity gains would be justified, as it becomes “cheaper” over time.
Once AI takes on redundant, repetitive jobs and humans can “relax and enjoy life”, I do wonder what this looks like. Part of the enjoyment comes from fulfilment in what we do. So, I feel that we would still continue to work, just not in the traditional sense of being paid to do repetitive tasks, but instead create economic value in other ways through creating stuff.
@@krishparikh1 Just think of the one billion jobs that will be automated over the next 10 years or less. More than likely the economy will double over the next 25 years due to AI, AEVs, drones, humanoid robotics. The ROI on that alone, for white collar type jobs is worth ten trillion plus efficiency gains of long term economic value and outside of agentic system advancements, not really based on gains in improvement, but rather adding functionality like multi-modal and action based learning to automate digital processes.