10.3.23-Showcase-Osaka-Themis_AI

SAM YOON: Hello, everyone. Great to see you today. Today I want to share with you Themis AI. We provide trustworthy AI solutions for the industry. So I just want a bit of audience participation here. Put your hand up if you've used Google before.

OK, a lot of people in the room. So you trust AI for search. It's a good start. Put your hand up if you've used ChatGPT before. OK, still a lot of people. Given the room, I'm probably not surprised. [? Teddy, ?] you should be happy.

Next, put your hand up if you would trust a self-driving car right now. OK, saw a few hands, but you can see that people are slightly more worried about these types of use cases. And we see a trend in the AI world where AI is slowly and slowly taking control of more parts of our lives.

But the areas that it's really struggling with are high-risk areas where physical elements come into play. So that could be autonomous driving, areas like aviation. We heard a lot of biotech and health care companies today as well. And at its core, it's because we don't fully trust these AI models. And that's the problem that Themis is trying to solve.

So we're basically an MIT team. I'm the only Harvard Graduate. So I feel very left out and less smart. That's not a statement that I can usually say, but it is in this case. Yeah, our co-founders came out of MIT CSAIL lab under the leadership of Professor Daniela Rus.

And for those that have used any sorts of generative AI use cases, you probably understand that there's not a lot of good solutions out there when it comes to actually trying to control the technology behind these large language models. A lot of the commercial solutions that you see right now are basically point access APIs. And it's very hard to actually take it apart and understand, basically see what's happening behind the black box.

So it's hard to control the reliability and accuracy to your specific context. It's also very difficult to fine-tune these large language models. Our friends at OpenAI might be able to say a few things about that later on. And, more critically, you know, I really want to see a world where AI is used to benefit humanity. But we still have a gap there, because we don't trust AI to drive our cars or to protect our homes or fly us in the air.

So there's this category called responsible AI. It's taking every government by storm. I come from a lawyer and policy background, so I can say this. Regulators are definitely slower than industry. But regardless, we still see a lot of activity in this space. Even in Japan, you may know that the government is more and more exploring this category of responsible AI and making sure that whenever we use AI there's good policies and standards in place.

So just a few case studies of where our technology had industry benefit. So autonomous vehicles is a kind of obvious candidate that comes to mind, where you would ideally want to have safer AI. Just a few metrics out there, in a nutshell, we saw nearly a 90% success rate, well, of failure. Well, 90% less crashes when we applied our technology to autonomous driving.

Another big recent trend in AI is this area of generative AI. And everyone is talking about what do we do with hallucination. How do we solve for hallucination? There's many different efforts out there to do this. But our technology provides a very standardized way to automate the detection of hallucinations.

So, for example, here we used our product on Stable Diffusion, a very popular open source image generator. And we could automatically detect what part of the image looked a little bit weird. And you can understand where that information could be really helpful in terms of the evaluators of the outputs, but also using that to train your image generator.

So to describe a little bit how technology works, the beauty of it all, it's very easy to implement. It's just a few lines of code that you wrap your existing machine learning model with. So it's model agnostic and data agnostic. And that's the beauty of the technology that came out of CSAIL. So in a scenario where a company might not necessarily have the R&D resources to fully create a uncertainty package, you can basically use our API and make any existing machine learning models that you have uncertainty aware.

So this is just a bit more detail on the previous case study that I mentioned. Conventionally, right now, if you use these generative AI products, you just get the image at the top, without any kind of additional information of where the image is going wrong. But at a pixel by pixel level, we can see where the uncertainties lie.

And then ultimately you can use that to basically point out where in the picture specifically it is going wrong. We have some other use cases when it comes to prompt engineering. So, for example, writers out there, or potential movie producers, are using these large language models to generate content. And prompts are a key part of that particular experience.

What we can do is basically identify what type of prompts induce more uncertainty into the outputs. So here we just have an example. We want to show a woman's hand. And from here, you put in close-up, woman's hands, 4K. As you can see the output is not of high quality.

We know what part of the prompt is causing that uncertainty, and we can basically augment that. You can automate this using our technology, and basically identify what specific prompt will create the image that you desire. So instead of experimenting, like many times, to try to figure out what type of prompt will generate what you want, this is the closest thing to an automatic solution to enable that.

Another examples is LLMs. There's many ways to break the rules that sit behind LLMs. And we could potentially flag where they could go wrong, what type of inputs will create more of the outputs that we want less and create more uncertainty.

So just to summarize, our solution can be used in many different ways, ranging from using it to curate your data, risk-aware learning, and also providing real-time what we call an AI guardian to fail-proof your implementation of your AI models in high-risk scenarios. Here are just a few industry partners that we've worked with. We, unfortunately, are not able to share the names of most of them, just because we use our technology in very high-risk scenarios.

And they're more sensitive about us sharing those areas. But for example, we work with a leading robotics company based out of Boston to make sure that whenever they have image detection in their robots, they can immediately flag to the human coordinators if anything in the inputs is causing uncertainty, and the actions that they want. So that's us, Themis AI. We have a booth outside, so feel free to drop by and ask us any questions.

Specifically, we're focusing on industries in the automotive field, health care, and Fintech space. But of course, our technology can be generally applicable. So feel free to reach out if you want to start preparing for the upcoming regulations in the AI safety space, or if you want to talk more about how it's like to be the only non-MIT person at an MIT startup. Thank you very much.