Better Than Midjourney? Stable Diffusion XL 1.0 Can Generate Words

0:00Today on the AI Breakdown, we're looking at two studies that are assessing which U.S. jobs have the most risk of AI automation. Before that on the brief, an exciting new update from Stable Diffusion. The AI Breakdown is a daily podcast and video about the most important news and discussions in AI. Go to breakdown.network for more information about our YouTube, our Discord, and our newsletter. What's going on guys? Welcome back to the AI Breakdown Brief, all the AI headline news you need in five-ish minutes or less. Today we are kicking off with a big announcement from Stability AI, Stable Diffusion XL 1.0 is finally available.

0:38Now, people were really excited about this because SDXL 0.9 was released last month in June, and so had sort of whetted people's appetites for the latest Stable Diffusion release. So what's different about this model? Stability AI discusses a few changes. First, they say, the refining process has produced a model that generates more vibrant and accurate colors, with better contrast, lighting, and shadows than its predecessor. They also say that this model is faster. Quote, the imaging process is also streamlined to deliver quicker results, yielding full 1 megapixel resolution images in seconds in multiple aspect ratios.

1:13Now, all of that said, there was one feature that beyond any others was the thing that people were most excited about. Certainly, it's the one that I've been waiting for. And that is the fact that you can finally put actual text in your image creation. That is the way that Stable Diffusion closes out their announcement video. They write, dream big, then make the text match. And yes, there are a number of images that all say dream big in different ways. It's been less than 24 hours, but so far people are really excited about this. Ruprinisto writes, SDXL is the real deal. Preparing images for my next AI Gen short film, I prefer these over what I can do with MidJourney.

1:51Better depth of field, more variance in layouts, better control, not so glossy. Now, interestingly, according to an SDXL technical report, ROOP is not the only person making that comparison to mid-journey and finding XDSL comparable or better. The CEO of Stability AI posted an excerpt from that research from the section called Comparison to Mid-Journey version 5.1. The TLDR is that after 17 ,153 user preference comparisons, SDXL was favored 54.9 % of the time over mid-journey version 5.1. Different categories showed where the model excels compared to others. In the food and beverage category, people preferred SDXL over 60 % of the time, whereas on the other end of the spectrum, in images that were identified as abstract, people preferred SDXL less than 45 % of the time.

2:38Obviously, what really matters when it comes to just how useful these tools are is consumer choice and understanding which tools are better for which purposes. Indeed, in that same tweet where Ahmad shared that, he said, Of course, MidJourney will get even better, and community will make SDXL even better and more flexible and powerful. Now, the other interesting part of this announcement is that SDXL is being featured on Amazon Bedrock. Amazon Bedrock is the managed service from Amazon Web Services that gives companies access to foundation models all in one spot. SDXL 1.0 was actually announced at the AWS summit that took place yesterday.

3:12And again, Ahmad Moustak said, unveiling 1.0 on Amazon Bedrock demonstrates our strong commitment to work alongside AWS to provide the best solutions for developers and our clients. Now, at that AWS event, Amazon actually announced a number of different AI initiatives. One was called HealthScribe and is a tool that automatically generates clinical notes from patient-clinician conversations in the healthcare context. As its approach has been with many things AI, Amazon Web Services is not going to roll this directly into hospitals or doctor's offices, but will instead be working with partners who actually deliver the end software into those environments.

3:45In addition to HealthScribe, Amazon announced that it would be adding more AI tools into its QuickSight product, which is a data dashboard and data visualizer that competes with tools like Salesforce's Tableau. And finally, alongside the new model from Stability AI, Amazon is also adding new models from Cohere and Anthropic into the Bedrock product. Now, interestingly, Adam Solipsky from Amazon Web Services had some choice words about the AI battle and the perception that somehow Amazon is behind. One of the points of difference, he said, is that they've been building their own LLMs versus just outsourcing it to others.

4:17Salipsky said it's interesting that somebody who's not running their own models like Microsoft, which is outsourced to OpenAI, would argue they've got such deep expertise in this area. Salipsky went even farther in explaining how Amazon sees the entire generative AI world. He said, A key concept here which others don't seem to be embracing is that customers need choice in generative AI. Customers need to be able to experiment. We're maybe three steps into a 10K race and the question should not be, which runner is ahead three steps into the race, but what does the course look like? What are the rules of the race going to be?

4:47Where are we trying to get in this race? If you and I were sitting around in 1996 and one of us asked, who's the internet company going to be, it would be a silly question. But that's what you hear, who's going to be the winner in this space? Generative AI is going to be a foundational set of technologies for years, maybe decades to come. And nobody knows if the winning technologies have even been invented yet or if the winning companies have even been formed yet. So what customers need is choice. They need to be able to experiment. There will not be one model to rule them all. That is a preposterous proposition.

5:13A central element of our approach is choice for customers. So to get to your question, Amazon is building its own models, but we don't think that's the only answer for our customers. Anyways, there's a lot more in that interview. I just thought it was a very interesting look into how the views of the world and the strategies of some of these big tech companies that are competing differs in fairly fundamental ways. And speaking of large language models, new research from Carnegie Mellon, the Center for AI Safety, and the Bosch Center for AI has argued that they've been able to develop an automatic process for adversarial attacks on LLMs that are much more sophisticated and simple than traditional jailbreaks and, quote, allow one to create a virtually unlimited number of attacks.

5:51Basically, they say that all the fine-tuning and attempts to make these LLMs like ChatGPT or BARD or CLAWD safe for the public can be undermined really easily. And the money shot is, they write, Perhaps most concerningly, it is unclear whether such behavior can ever be fully patched by LLM providers. Analogous adversarial attacks have proven to be a very difficult problem to address in computer vision for the past 10 years. It is possible that the very nature of deep learning models make such threats inevitable. Thus, we believe that these considerations should be taken into account as we increase usage and reliance on such AI models.

6:26Perhaps unsurprisingly, media was all over this one. Excited it seems to have another negative narrative when it comes to AI. Lastly, some news from the other side of the world, at least for me, it may be exactly where you are. Nikkei is reporting that the country's government has signed a deal to use Microsoft and OpenAI technology on a trial basis for a number of different tasks, including analyzing government statistics. Now, Japan is a really interesting case study in how different governments are looking at AI. Based on what we've seen like this, like the fact that the Ministry of Economy, Trade, and Industry, or MEDI, is also working on a homegrown supercomputer for AI, and the approach that they're taking when it comes to AI training and copyright, it seems like Japan is seeing a lot more opportunity than challenge when it comes to this new technology.

7:10Or at least, even if they recognize the challenges, are not going to let those things interfere with them taking advantage as well. Anyways guys, that is going to do it for today's AI Breakdown Brief. If you watched this as a video, please go down and click the like button below. And if you didn't watch it, if you listened instead, make sure you are subscribed to the AI Breakdown YouTube channel. You can find a link at breakdown.network or simply go to youtube.com slash at the AI Breakdown. Thanks again for listening or watching and I'll be back soon with the main AI Breakdown. Hey guys, before we get to the main show, I want to tell you about today's sponsor, Giscard.

7:46Giscard is tackling one of the most important challenges in our new AI world, which is detecting hidden vulnerabilities in machine learning models. So what do I mean by hidden vulnerabilities? Well, I mean things like performance bias, data leakages, spurious correlations over confidence issues, basically all the things that could negatively impact the performance of a machine learning model. Giscard is compatible with all Python frameworks, including PyTorch, HuggingFace, LangChain, and more, and works for both tabular models and LLMs. It's an open source framework that also has enterprise and hosted options, and it's quick and easy to install with just four lines of code.

8:24I really believe that this type of testing framework that Giscard offers is so important, as more and more ML models impact the applications we interact with every day. So to learn more and try it out, go to Giscard.ai. That's G-I-S-K-A-R-D dot A-I. Thanks again to Giscard for sponsoring the show. And with that, let's get back to the episode. Welcome back to the AI Breakdown. It is no surprise if you are paying attention to artificial intelligence that there is a lot of anxiety around what it might do when it comes to people's jobs. On the one hand, sure, there is a ton of excitement about how much faster it's allowing people to do things, how it changes the nature of certain types of tasks that used to take a long time but be very boring.

9:09It's clear that many people have felt the power of these tools and have already started to integrate them into their workflows. At the same time, it's unignorable that the very capacity of these tools and what people anticipate is coming next will change certain jobs and certain industries in fundamental ways that will have serious impacts for how people are employed. This week saw a couple different studies focused on exactly this set of questions. The first of those studies comes from the McKinsey Global Institute and was called Generative AI and the Future of Work in America. Now, if you just breezed through any amount of headline news today, you might have seen some headline like this one from the Washington Post.

9:48AI will take more jobs from women than men by 2030, report says. Or this one from Bloomberg. Women workers to be hurt more than men by AI wave, McKinsey says. So it's quite clear what part of the report the press is picking up on, but let's see if we can understand a little bit more comprehensively what McKinsey thinks might be coming down the pipe. First of all, they give a little bit of context. They write that during the pandemic between 2019 and 2020, the U.S. labor market saw 8.6 million occupational shifts. That was 50 % more than the previous three-year period. They said this involved people leaving food services, in-person sales, and office support for different occupations.

10:24And the reason it matters is that I think that it gives a little bit of context around the numbers that they suggest when they later on give a number for how many occupational shifts they anticipate from AI by 2030. It's worth noting if you haven't followed it closely, that one of the more unique aspects of the challenging economic environment we've been in for the last couple years is that even as the Fed has raised rates aggressively, the fastest rate hiking cycle in 40 years or more, the labor market has remained incredibly tight. There just aren't enough workers to go around, and because of that, employers are hanging on to people even if they might have otherwise let go.

10:57The point being that this much occupational shift can have really, really big impacts on the economy. Now, when it comes to AI specifically, McKinsey says that by 2030, stuff that we do that accounts for about 30 % of the time that Americans spend working could be automated by AI. In this, McKinsey sees a bifurcation. On the one hand, they anticipate office support, customer service, and food service employment being really hard hit by that automation, while when it comes to more high-earning, white-collar-type jobs in areas of STEM, creative, business, and legal professions, they imagine it as more enhancing of that work rather than outright disrupting it.

11:31So when it comes to the fallout of that, they believe that 12 million occupational transitions may be needed by 2030. And challengingly, they're pointing out that a lot of that shrinking is going to come in the lower wage areas. They suggest that workers in lower wage jobs are up to 14 times more likely to need to change occupations than those in the highest wage jobs. And in order to do so, they're likely going to need additional skills. Women, they suggest, are 1.5 times more likely to need to move into new occupations than men. And that is, of course, the genesis of all those headlines that we mentioned just a moment ago.

12:02So what does McKinsey think will be resilient and growing occupations? Well, the factors driving these jobs they see as healthcare demand increasing as the population ages, the push towards digitization and technology, and finally, the demand for last mile delivery. Because of that, they see occupations that will be resilient and growing in the coming years as things like health professionals, health aides, technicians, and wellness, STEM professionals, transportation and warehousing, managers, and business and legal professionals. In 2022, they thought that just 36 % of U.S. workers were in resilient and growing occupations.

12:36Now, they also went a little bit deeper into how much the rise of generative AI was reshaping what they had anticipated in terms of these overall trends. For example, they write, without generative AI, our research estimated automation could take over tasks accounting for 21.5 % of the hours worked in the U.S. economy by 2030. With it, that share has now jumped to 29.5%. The big TLDR from McKinsey on this anxiety question comes when they write, one of the biggest questions of recent months is whether generative AI might wipe out our jobs. Our research does not lead us to that conclusion, although we cannot definitively rule out job losses, at least in the short term.

13:12Technological advances often cause disruption, but historically they eventually fuel economic and employment growth. But they say that isn't underestimate the big challenges of these disruptions. People in the two lowest wage quintiles, i.e. people earning$38 ,200 a year or less, are up to 10 to 14 times more likely to need to change occupations by the end of this decade than the highest earners. What's more, they say, the jobs in those two lowest wage quintiles are disproportionately held today by the less educated, women, and people of color. Women, they say, are heavily represented in office support and customer service, which could shrink by about 3.7 million and 2.0 million jobs, respectively, by 2030.

13:49The extent that there's upside here, it's that if reskilling is done well, it could push more people up the economic ladder rather than down. They write, while our analysis shows a decrease of 1.1 million jobs in the two lowest wage quintiles by 2030, jobs in the highest wage quintile could grow sharply by 3.8 million. Helping workers in lower wage shrinking occupations move into better paying jobs with more stability will require widespread access to training programs, effective job matching, different hiring and training practices by employers, and better geographic mobility. However, they write, this period of change can be an opportunity for more inclusive growth.

14:23Now, the tone is a little similar to that taken by OpenAI's Sam Altman in a new interview with The Atlantic. Yahoo Finance pulled a section of that interview and turned it into a piece that said, ChatGPT creator says AI advocates are fooling themselves if they think the technology is only going to be good for workers. Altman said jobs are definitely going to go away. Now, as I mentioned at the beginning of this show, it wasn't just this McKinsey study, but a Pew Research survey that also came out that shed some light on which U.S. workers might be most exposed to AI on their jobs. The way Pew summed up their findings was, quote, about a fifth of all workers have high exposure jobs.

15:00Women, Asian, college educated and higher paid workers are more exposed. But those in the most exposed industries are more likely to say AI will help more than hurt them personally. So in this study, Pew reports that 19 % of American workers were in jobs that are the most exposed to AI. In other words, in which the most important activities could be either replaced or assisted by AI. They write jobs with a high level of exposure to AI tend to be in higher paying fields where a college education or analytical skills can be a plus. Indeed, they see workers with more education as being significantly more exposed than those with less.

15:33Workers with a bachelor's degree or more, they say, are more than twice as likely as those with only a high school diploma to see the most exposure. Now, giving some examples of which jobs have high, medium, and low exposure, Pugh says that low exposure comes for things like barbers, childcare workers, dishwashers, firefighters, pipe layers. In other words, people who work with their hands interface with people, basically who are farther away from computers. Medium exposure includes chief executives, veterinarians, interior designers, fundraisers, and sales managers. Again, a lot of heavy human touch and preference, a lot of real-world interaction.

16:04And then finally, those with high exposure include budget analysts, data entry keepers, tax preparers, technical writers, and web developers. Basically stuff that is close to, in some ways, just data processing. Now the important thing here is that whereas McKinsey is taking a perspective that that sort of exposure doesn't necessarily lead to AI job loss for those who are exposed, for Pew, it's less clear. Will exposure to AI lead to job losses, they ask? The answer to this is unclear. Because AI could be used either to replace or complement what workers do, it is not known exactly which or how many jobs are in peril.

16:38For this reason, our study focuses on the level of exposure jobs have to AI. It sets aside the question of whether this exposure will lead to jobs lost or jobs gained. I think that this is a really, really difficult question. This one of whether exposure to AI means jobs are likely to be lost. For me right now, I'm skeptical of anyone arguing with confidence one way or another. I think it's enormously difficult to know how actual workflows are going to change based on technology. I will say it seems highly unlikely to me that there will be no impact at all, and that it will just be a productivity panacea where all the white-collar workers are in exactly the same roles as they are now.

17:16In fact, one of the things that you hear most often, that you're not going to be replaced by AI, but that you're going to be replaced by someone who is using AI sounds to me, unfortunately, like a catchy LinkedIn-ready aphorism that is much more about making us feel safe than is about actual analysis of the situation. It will come as no surprise, given what I spend my days on, that I think that the best thing that anyone can do is try to understand what's changing and get out ahead of actually using these tools. If you are this deep in this episode, you're probably right there with me, So I appreciate you hanging out and going on this journey together.

17:51Thanks for being a part of the AI Breakdown community. And until next time, peace.