Shocking Report Shows Why AI WONT Take Your Job (New Report)

So there was a recent report from MIT that was rather fascinating on how AI won't steal your job anytime soon. And of course, since MIT is a well esteemed university in the United States, I thought that their report research is rather important for the AI space. So we're going to be taking a look at the article, what the report says, and a few other things that many people aren't really thinking of when they look at the future of work with AI. So here you can see this is an article on Forbes and it clearly states that that AI will not steal your job anytime soon, MIT researchers say. So essentially in this article they break down exactly why. And the reasons that they give are rather intriguing because whilst they do pose some really good stuff, there are some factors that this paper didn't discuss. And I would argue that some bits of it are a little bit narrow in the sense that it doesn't cover a lot of stuff. But it still is a very good piece of research because it does show us why the innovation in AI doesn't necessarily mean that many jobs are going to be disappearing right now. So essentially what the main point of this research is is that they are talking about the cost of AI systems. Now when I'm referring to AI systems, it could be LLMs, it could be vision based systems, it could be robots as well. So just bear that in mind when I refer to AI systems. So essentially what we can see here is that it says to address these shortcomings and construct what they call the first end to end AI automation models. The MIT research is focused on vision related tasks where the cost estimates for the AI systems are more developed. The model evaluates the level of proficiency needed for a task based on surveys of the relevant workers, the cost of achieving that proficiency via human workers or AI systems which are pretty expensive to develop, and the economic decision by firms whether or not to adopt AI for a particular task. Basically what they're saying here is that AI systems aren't cheap to run in their current state. We do know if you look back at GPT4, you remember that GPT4 currently only has 70 messages every three hours or around 40 messages every three hours, depending on which API you're using, that is the current limit. And essentially they make a point that this is pretty expensive to run. And of course because of how expensive, how computationally expensive it is, that essentially means that it might not be as feasible for people to be replaced with AI at its current state. And they do use some examples in the paper where they basically say they that this isn't going to be feasible. So there's some other points here that they do talk about. So they said that using the new and improved model, the MIT research has found that only 23% of worker compensation exposed to AI computer vision would be cost effective for firms to automate because of the large upfront cost of AI systems. Lowering the cost of AI development and or the cost of AI deployment by increasing scale, possibly using AI as a service platform would lower these costs. However, even with rapid decreases in costs of maybe 20% per year, it would still take decades for computer vision tasks to become economically efficient for firms. So I mean, what they're stating here is pretty true. From based on what we've seen, there was someone I remember who even used the GPT4 Vision API. And essentially if you don't know what that is, that's essentially just GPT4 with vision. You know, it's ChatGPT with images and it was pretty expensive. I remember them saying that they racked up a, a pretty large bill in only about an hour and this was not economically feasible at all. Although if it was realistically really cheap, it would be feasible. But even if it decreases at 20% per year, they're stating that it would still take decades. Now I think that the cost decreasing at 20% per year is going to be something that happens a lot quicker than that because they're saying it's going to take decades. And I just don't know if that is the case because we've seen AI development rapidly, rapidly increase. So for them to say it's going to take decades, I'm not sure where that fits on the timeline with other things in AI because for example, of course, if you know that with Ray Kurzweil, he actually predicted that there would be a singularity in around, I think it was 2040. And he also did predict that in around, I think it was 2027 or 2029 that there would be the first AGI system. So with that being said, I'm not sure how they're saying it's going to take decades. And decades just means like a really long time. So I'm suggesting that maybe they're looking at some pieces of research that maybe we just haven't considered. Now essentially they also state MIT researchers argue that their findings apply also to generative AI or the automation of language related tasks. All areas of AI deployment require fine tuning or customization to adapt them to firm specific characteristics, a crucial cost factor and obstacle for rapid adoption. Now I'm not going to lie, this is a very good point. Now there are two things to consider here because this is rather incredible because one thing that I know a lot of people didn't notice was when GPT4 was released and OpenAI talked about custom fine tuned models, they actually talked about if you are a company, essentially if you want to get your custom tuned model, like for example, and I'm not talking about this where we're talking about fine tuning a model which is just like a simple version, we're talking about them building a complete thing for a company. I'm not sure if this is true, but the rumors stated that it was starting at around a million dollars, which is pretty insane. But, but I do think of course some of the larger conglomerates can easily, easily pay that amount. But we have seen that training a model on your data is becoming easier and easier because of course if you remember there was also this thing that was just recently released and of course this is the GPT store. So essentially if you remember there was this thing that was released because of course this is the GPT Store. And with this you have, I guess you could say a simpler way to essentially use your AI that is, I guess you could say, trained on your data in a way that could be applied to certain companies. Now we've seen time and time again that sometimes this doesn't work due to hallucinations and due to the guardrails not being stuck by, which is something I'm going to explore later. But it definitely is an obstacle to adoption that OpenAI realize and of course they are trying to change with the adoption of GPTs. And of course there of course are these things. So these things of course are plugins as well and these are ways that companies can use these different chatbots as well. So I think that the obstacles to adoption is a very, a very good point because that is of course true. But I think like I said, in the future we're going to have systems that are going to be able to be trained on data really rapidly. There are some new advancements that I'm going to be talking about and of course adapting them to firm specific characteristics. I do believe that there's going to be a very large automated process soon. Now essentially what they said here was that another obstacle to adoption not always considered by economists and other people was what another study calls the societal acceptability of AI. Some professionals may seamlessly integrate AI tools whilst others could face resistance because of cultural, ethical or operational concerns. And this is a very, very big truth that I will explore in a moment because although AI might be good, it's not human. And the inherent nature, because it's not, of course not human, doesn't mean that some things just won't be accepted because they aren't human. And I will elaborate on that more because I have some fascinating examples to show you. So then essentially they say here that essentially in advanced economies, this IMF report which we already looked over, it says about 60% of jobs may be impacted by AI. Roughly half of the exposed jobs may benefit from AI integration, enhancing productivity and of course this thing which people did freak out about. But essentially you have to remember the word impact doesn't mean negatively or positively, it just means that it's going to affect it in some way. Okay, so the 60% impact just essentially means that it's going to affect it. It could affect it positively, of course it could be negative, but according to the research it was that 50% of the jobs would be negative, essentially meaning that it could remove demand for that work. And then of course the other half, it means that it would make your life easier. So of course a shift in the economy, but of course the effects of this would be long term. Now one of the things that I do want to talk about is I wanted to actually talk about if cost is actually going down. Because remember, if we look back to the first point, they talk about how cost going down is of course one of the first things that we need to look at if this is going to be a reality and if these systems are going to be feasible for future use. So you can see right here that this is a paper that many people did actually miss. But this is going to be a conference paper at ICLR 2024 and this is called slicegpt compress language models by Deleting rows and comments so columns. So essentially they said here that Slice GPT effectively compresses models by up to 30% while retaining a high percentage of the original model's performance while on zero shot tasks, notably for Llama 2, 70 billion parameters, OPD 66 billion parameters and Phyto models and it maintains 99%, 99% and 90% of dense model performance respectively with 25% slicing. So essentially they were able to make these models smaller and retain most of the quality. So I think that this kind of research shows us that whilst they state that of course cost is one of the major things and, and even if they get 20% down, it's going to be decades. I tend to disagree just a little bit because I think that what we're seeing is that these models are really, really effective, even if they're not huge. Now there was also something that I wanted to also elaborate on a bit because if we do look at cost reduction, one of the main things that we do talk about is that this is crazy. Okay, so right here you can see that we do have Phi 2, and Phi 2 is a 2.7 billion parameter large language model that outperforms other models that are 25 times larger. And essentially right here you can see it says phyto matches or outperforms models that are 25 times larger thanks to new innovations in model scaling and training, data curation. So essentially what research has found was that even if the model isn't as good, if you just surely focus on those parameters, not like just giga scaling the model to a trillion parameters, if you just focus on the model and if you focus on the kind of data that you're feeding it, you can definitely get a lot more performance out of a much smaller size. And that's essentially what they did. And this is something that is, you know, it was pretty much groundbreaking. And I'm pretty sure in future models they're going to be as much smaller than the ones we have now, but they're also going to retain the same level of quality. Now this next thing actually does come from Sam Altman because Sam Altman does actually talk about the cost of AI coming down in this interview with Bill Gates. And it was something that a lot of people did miss. But I think it is rather important because it goes to show that cost is of course something that Sam Altman is thinking about because that limit that's been on GPT4 for quite some time, we've been wondering whenever is it going to come down?

Let's see. But I'm very optimistic and I agree with you. What a contribution would that be in terms of equity? Technology is often expensive, like a PC or Internet connection, and it takes time to come down in cost, I guess the costs of running these AI systems. It looks pretty good that the cost per valuation is going to come down a lot. It's come down an enormous amount already. GPT3, which is the model we've had out the longest and the most time to optimize. In the three years that's three and a little bit years that's been out, we've been able to bring the cost down by I think a factor of 40. So for three years time, that's like, that's a pretty Good start. For 3.5, we've brought it down, I would bet close to 10 at this point. 4 is newer, so we haven't had as much time to bring the cost down there, but we will continue to bring the cost down. I think we are on the steepest curve of cost reduction ever of any technology I know, way better than Moore's Law. And it's not only that we figure out how to make the models more efficient, but also as we understand the research better, we can get more knowledge, we can get more ability into a smaller model. And so I think we are going to drive the cost of intelligence down to so close to zero that it will be just this before and after transformation for society. Like right now, my basic model of the world is cost of intelligence, cost of energy. Those are the two kind of biggest inputs to quality of life, particularly for poor people. But overall, if you can drive both of those way down at the same time, the amount of stuff you can have, the amount of improvement you can deliver for people, it's quite enormous. And we are on a curve, at least for intelligence, we will really, really deliver on that promise. But even at the current cost, which again, this is the highest it will ever be and much more than we want, for 20 bucks a month, you get a lot of GPT4 access and way more than 20 bucks worth of value. So I think we're already.

So I think from that clip you can see clearly that the path in terms of the cost of intelligence is actually going to go down. And I think, you know, the comments from Sam Altman clearly show us the direction that we're heading in. And that is pretty much someone who really does know exactly where AI is going. So you couldn't have got a better spokesperson for that. Now, the next thing I did actually want to talk about was the fact that people do hate AI. Now, hear me out on this point because it's a point that doesn't get brought up a lot, but people have to understand that a lot of people don't actually accept AI. Now, remember previously when we talked about how there are, you know, societal, cultural and ethical, you know, reasons why AI acceptance is going to struggle. But one of the examples I did actually want to bring up was, of course, this channel. So a lot of people will say that, you know, AI is great, it's just going to replace everyone. But for example, on this channel, I rarely ever use AI voices. And when I do, it's because I'm just so completely busy that the environment doesn't allow me to record a voice or, you know, I have like, maybe like a health issue and I'm not able to use my voice because like a cough, maybe a cold, maybe something like that. And it just, you know, would sound awful. But people still need their news or up to date information. Now if you take a look at some of these comments on this singular video, you can see exactly why this isn't the case. So people are stating that, you know, we prefer and are used to your natural voice. Can you not just clone it with the Lemon Labs when you can't record? Please, no more AI voices. Now if you think those comments are just, you know, cherry picked, take a look at this one. It says the problem with using an AI voice is that the audience is left guessing whether real effort synthesis was involved while creating the content or not. And we want people who are trained on real world data far faster than AI is currently trained to do with the research because current AI aren't familiar with subtitles. And it's not that we hate the eye voice, it's more like we doubt the expertise level. You can see that it's got 23 likes as well. So. And I'm going to show you guys one more comment as well because there is another comment that is rather fascinating and it says right here, I prefer a real human voice. It's warmer, more engaging. When I can tell the voices generated, I honestly tune out. I believe the content more when a real person speaks to me. And that one had 82 likes. It was the most liked comment on that video. And the video wasn't bad by any means. You can go ahead and watch the video. It was a video about LK 99. But on that day I was just specifically really just drained, tired. I had a cold. I was just like, there's no way I could get this video out. But it was rather important. And of course, you know, some people just didn't enjoy that video. Now that is 100% completely understandable because the end of the day you're making a video for real people and real humans and humans get to decide whether or not they enjoy a piece of content. Now, I think in some video pieces, AI voices are okay. For example, there's some YouTube channels that do completely well with it and that's completely fine. But in some areas, you know, on some channels, I think the majority people want to talk to people. And this is where these kind of comments come from. And it's completely understandable because people don't just want to hear some droned out AI Voiceover that is, you know, pretty much, you know, it doesn't have a human emotion, doesn't have inflections, doesn't have pause, doesn't have stops. And I think that that is going to show you that whilst yes, AI can replace people and you might be thinking, why on earth are you just talking about voiceovers? I'm going to get into some more examples to show you that this is like, you know, across many different industries, this is a real thing. I think humans interacting with humans, unless it's like completely robotic, like data, you know, one plus one is two kind of automation software stuff, I think where the human approach is usually there. I think humans will always prefer a human approach. Now there was also, if you hadn't heard about this, there was a huge game. If you're not familiar with the gaming industry and how big it is, there was essentially a huge game that rose to the top of the Steam charts. And this game got a lot of hate, like a lot more hate than I would have liked to see. And it says Twitter is annoying. But all the big hate posts I found is focused on one thing, AI. They say the game models and game was created using AI, so they're hating on the game. But before you ask them for proof is, I have no proof. Basically, okay, there was a game that was created and it was really popular and you know, some of the devs were basically saying that they used AI to help them create the game better and make more assets quicker. And a lot of the hate that was coming from was around this. And I mean the game is great, it's loved worldwide and I don't think a lot of people care, but there were a lot of people who were just simply saying, I'm not going to play the game. And it's just like an AI generated game. It's just soulless, yada, yada yada. And once again, that goes to show that when people understand that something is AI generated or an AI contributed towards that, they really do tend to steer from that kind of content, even if the end product is somehow valuable. So another thing as well, and this is why I want to talk about customer service and chatbots. This has been a long time thing. If you've ever spoken to like, you know, those, and I know, I guess you could say it's a rudimentary form of AI, but you've ever spoken to, you know, a chatbot and you can hear like it's, you know, just like a woman, like a Siri kind of voice on the phone you know how frustrating that is, especially if you have a really dire issue that you need a human to take care of. And you completely understand that the robots are never going to understand because those ones are just like, you know, like branches, where if you say one thing, it's going to lead you to another thing. It's like, press five. People really do hate those things. I can completely understand why. Because when you call up, you've paid your money, you expect a kind of service, and you're met with like, a bot. It just feels soulless and, you know, customer service whilst people are trying to replace them with LLMs. I don't think that's going to happen too crazily because this person here, you can see if you weren't familiar with this, we covered this in a week in AI and essentially this person bought a Chevy Tahoe for $1. And essentially they were able to trick the customer service chatbot into getting that deal. Of course, it wasn't legally binding, so they weren't able to actually get the car for a dollar. But the point is that hallucinations are still a big thing. And. And whilst, yes, there will be guardrails to improve this in the future, I think it's going to be kind of interesting to see how people circumvent those guardrails in the future and if they're able to. And what kind of things are there in the future to prevent these kind of things from happening. Because once people realize it's AI, people always try and go around the guardrails. And I wonder if there's going to be some kind of question or some kind of chain of thoughts or some kind of prompting to be able to always reveal if it's an AI or not. Maybe there's going to be some kind of new Turing Test. And of course, this as well. You can see DPD Customer Service chatbot swears and calls company Worst delivery firm. So, you know, the point I'm trying to make here, guys, is that, yes, AI can take your job, and in some industries, it might, you know, reduce demand. I think that where people realize that humans are really valuable or even just, you know, humans are, you know, like, I guess you could say the ones that are supposed to be there. Replacing it with AI isn't always going to go down well. And remember, AI art, You know, some organizations banned AI art because they were like, you know, there's legal issues and there was even some controversy about AI artworks. So, for example, there was this artwork, and he created something called the Space Opera. Theater using Midjourney, submitted it as an entrance to the Colorado State Fair's annual art competition. And he won first place. And people were pretty, pretty furious. And I can completely imagine. And the problem was is that, you know, this is outrageous. Prompting a machine to make you something does not make you an artist. As an artist of any style and technique, we put in the time to hands on create something. And this guy can type and refresh in the process a few times. And that is being compared to craftsmanship. And I completely understand this entire Twitter thread. I mean, some people are saying that choosing the right prompts, selecting the best image, and post processing in Photoshop are all creative expressions, then it's art. You're judging work based on the effort behind it. But many works for abstract are technically effortless and considered art anyways. And that is of course a true point, because how many modern art museums have you seen where it just looks like something that you could have done in five seconds? And some people are saying that it's magnificent. So I guess art isn't of course the best example. But the point is that when you use AI to create something like this person did, and of course you win first place, it brings up a whole host of other issues that where you think, okay, because AI can do it better, it means that instantly people are going to value it. And of course, we always come back to the same example of chess, where chess is something that could replace people. Like, oh, chess is the best at AI. Nobody wants to watch chess play AI. Nobody wants to watch AI play chess because it's just completely boring. Now something that I think this is one of the most fascinating things because health is of course good. But essentially it says right here that 60% of Americans would be uncomfortable with their healthcare provider relying on AI in their own healthcare. So it says. Yet many see promise for AI to help issues in the bias of medical health care. So there was essentially some research that basically stated, I think it's right here. So it says that generally the American public still prefer a human physician to be responsible for their medical care over an AI. So this is. However, the general public seems mixed when the survey results are taken as a whole. So essentially people just prefer humans in terms of the medical issues. And that's crazy because if we look at the contrast, if we look at the health is good. But so essentially, if you haven't paid attention to this, there was this thing called Google's Amy and it was essentially AI assisted doctors performed better than UNAI assisted doctors. But AI on itself actually performed better than AI assisted doctors. And even though that is better overall, it's still clear from the research that humans prefer humans. And I think you need to understand that this is a recurring theme. So, you know, there's going to be some companies I do think in the future that do get a lot of positive PR from stating that they're not going to use AI and that they prefer using humans. And of course, until we get AGI, there was this video I want to show you from an upcoming video on robotics that, you know, this is from Princeton. It's just amazing stuff. Essentially a delivery robot that can go ahead and vault over curbs and they made this robot to deliver packages. And the point is, is that you're still gonna need humans to do some things until we get general purpose AGI humanoid robots, because some people are always going to steal the robots. Some people are going to just do stuff where I guess like what the situation you're seeing is going to be a zero shot situation and you're going to need to get that right. And humans do very well in those scenarios where these kind of robots just really don't do well in. So I think largely we're still going to pretty much need humans for quite some time. I mean, it will be also interesting to see the emerging fields. And overall, whilst AI probably won't steal a job anytime soon, I think the way how the industries will shift will be something to take a look at and of course pay attention to. Because whilst new research and new studies are always good, it's always important to see what's going on in reality and what people actually do want to and surveys based on how they do feel. So with that being said, let me know what you thought about this if you have any questions.