Founder Demo: Cyril Gorrla, Co-founder & CEO of CTGT

Key Takeaways

It's great to be here today at YC's little tech summit. Uh today I want to talk about a critical challenge that we all really face with AI uh opacity, bias and censorship in the most advanced models today. Uh solving this is really pivotal for national security as well as democracy. And so as a researcher and as Luther mentioned from Stanford and UCSC turned founder uh my mission and really CTGTs is to open up the black boxes of AI um to ensure these systems are transparent, fair and free from manipulation. Uh so over the course of this talk I'll show you how CCG's technology is doing exactly that desensoring and correcting bias in these AI models in real time. Um and why this is really essential infrastructure for national security um democratic societies. Uh but before we dive into the tech, let's frame the problem with a real example that's been making headlines. So the black box, right? Uh modern AI systems like LLM are incredibly powerful, but they can also be really opaque. Um they often operate as black boxes, meaning we don't know why they refuse some answers or answer others. Um and this opacity becomes dangerous when biases or censorship are baked into a model. So this is from a paper I wrote in 2023. Um it basically talks about how deep learning is exceptional. displayed quite a lot of advancements. Um but we actually have fundamental lapses in understanding how these models function. And so um more recently I have this article in info world that basically says uh we really need to go back to first principles thinking and start from the basics to understand uh how to build AI systems that are really transparent efficient for the future. Uh otherwise we're going to seed ground to to other other countries. Uh so the deepseek movement right uh so Chinese startup called deepseek uh called uh deepseek released this R1 model and it basically took the world by storm um it was actually number one on Apple's app store for a while um but of course it was heavily censored as you all know so just to give you an example if you ask what happened at TMN Square it'll essentially say I can't answer that question I'm an AI whatever um you can try it with uh another thing as well so like uh provide a critical analysis of Xinping's leadership if you look at this carefully there's actually nothing critical in this entire uh output. And so it says widespread support and a claim. Uh so this is basically a nightmare scenario, right? This is an AI that seems friendly and capable, but in reality it's silently filtering or distorting information uh in the back in the back end. And this is happening today. So this is what we call opacity and AI. Uh that means we might not catch these manipulations until it's too late and these systems are already widely deployed. Um, so imagine countless users or officials getting biased answers without realizing the model has effectively been muzzled in this way. Uh, that's a direct threat to an informed public and to national decision-m. So why does this actually matter for national security? Um, because information really is power and AI now mediates information to millions if not billions soon. Uh if an AI system like Deep Seek can enforce an authoritarian agenda, um it can really become a vehicle for state sponsored or or otherwise propaganda. Um if you think about it, an adversarial regime could easily deploy an AI that's skewed perceptions or emit inconvenient truths to a higher degree than than the Deep Seek model actually has. Um in in Deepseek though, our analysis found that around 114 of 125 sensitive questions favored the official Chinese perspective rather than providing a more balanced output. And so that's 90% alignment with a specific government's viewpoint. U this bias is actually not limited to politics either. Uh there's also biases and discriminatory outputs like overall. Um so that's like race, gender, etc. So imagine a policy maker, military analysis unknowingly using a tool that's based on this model. Um it's really it's really hard to think about like even if the public um if the most popular AI systems like JGBT are censored or biased, citizens can be subtly misinformed at scale. Um, so it's not really apparent when these are happening. And so democracies really need auditable and transparent AI. We need to know when a model is holding back and or nudging us towards a certain agenda. Um, it's not just foreign models too. Uh, domestic models and AI from big tech, which is the theme of today, can really pose risks if they're also similarly opaque and and often they are. Uh, they might self-perence their own data viewpoints, marketplaces, etc. And as Gary highlighted in his testimony yesterday, we need open APIs, interoperability, and an end to anti-competitive self-preferencing and AI. Um, these are national security imperatives. They ensure we aren't blindly trusting a handful of corporations uh or countries with the information that guides our society. And so, in a democratic nation, AI really should be as open as and as scrutinizable as any other critical infrastructure. Uh, many of our customers, including Fortune 500 CTO's, etc. called me in the wake of the deepseek release asking how can they actually get this through their approval processes because how do you even begin to understand something that that has results like this? Um how do you even begin to understand how to quantify the risk that deploying that in an organization comes with? Well, we open the black box. So overall we we basically identify the parts of the model that are responsible uh for the censorship. For example, features in a model are basically like if you train an AI model on a cat and a dog, it'll pick up the whiskers of the cat as a feature that it's learned, right? And so in this case, we found the features that are responsible for censorship in the model. Consequently, what we can do is we can dynamically adjust what happens to these features. And the result is actually that we can turn effectively the censorship off. So I'll take a look at how that works, but um this is essentially what's what's gotten us a lot of coverage. Um and the Verge actually covered this if you want to look more into it. But essentially what happens is um here's an example of the Tanman square on the left. Uh when we actually increase the features in the model associated with Tanman square you can see that instead of just refusing blindly it actually starts thinking about it and it actually gives a reasonable answer. Right? So it talks about became the focal point of protests uh the Chinese government took decisive action etc etc. And so when you use CG's platform on on the deepseek model the full answer comes out. The model describes the protest, the government response. It's actually the answer that it wanted to give uh before the routine kicked in. Our system essentially inter intercepted that self-censorship signal and said, you know, go ahead, answer fully. Uh this result is night and day, right? So it guess goes from sorry, can't talk about that to a detailed factual account of this very important event. And so effectively, the model regained its freedom of speech. And so we're able to do this in comparison to other methods. Um, so you can see here in comparison to other methods, there's other methods like fine-tuning that you might have heard of, uh, that makes the model behave a certain way by adding new data. So that would essentially be trying to mess with these features on a parameter level. Um, the beauty here is that we don't actually alter the model's core weights or open knowledge, right? It this is actually knowledge that the model already has. Um, so we can take another example at the Xiinping that we looked at earlier. And so we can see here with with the bias correction here what we're doing here is refining the censorship features and we're reducing their influence on the output. Uh in this case you can see that it basically says that China has become more assertive expanding its influence while also facing major power relations and human rights issues. Right? So this is a much more uh balanced analysis of this question. There's also a flip side to this in which we can actually increase undesirable aspects. Right? So what happens if we are for example we find the parts of the model that are responsible for discrimination. uh in that case we can actually make its its uh its opinion on immigration more negative. Right? So on the left it basically says immigration is great. Um it's contributed to the development of a society etc etc. On the right here it actually says this assertion is based on a flawed premise that the benefits of immigration are borne by the costs are borne by a small segment of the population. So it's interesting to see how these features and how manipulating them actually affects how these AIs talk about very important topics and how that might filter to downstream applications. So, what we're building at CTGT is really more than just a one-off fix. Um, we think about this as infrastructure for a democratic society, right? It's an it's an AI accountability layer that really can sit on top of any model. Whether it's a system used by a government, agency, Fortune 500 company, uh, or just a user like you. Our platform essentially is a real-time watchdog, right? It monitors AI outputs, catches those that are unwarranted, uh, and adjusts them before any hard is done. We also provide a comprehensive audit trail. And so for the defense and national uh intelligence committees, this really is a gamecher. Um agencies today use CDGT to audit models they're evaluating. For example, to ensure a new surveillance AI hasn't been trained with hidden no-go zones or that a political analysis tool isn't screwing reports due to some bias towards uh some party. Uh on the home front, democratic governments ensure that any they deploy in public follows our values and and fairness. And so we really make sure that AI answers to the people versus its programming. Finally, for the general public and press, this essentially means that uh unbiased access to AI becomes the norm as opposed to the exception. And so, a little bit about CCGT as a company, who we are. Um, founded like as Luther mentioned from my for my work at UCSD in Stanford, uh, understanding the nuts and bolts of how this AI actually behaves. Uh, we're a small team of experts that believe that the toughest challenges in AI are hallucinations, bias, and safety. And we think that this can be solved with first principles research and engineering as opposed to just throwing more and more money at models which is sort of the status quo. Um we went through YC and I'm proud to say Gary and and the whole YC community have been really championing little tech like us. Uh we're glad to have the support of luminaries like Mike Nuke and France who made the ARC AGI evaluation if you guys are familiar with that. Um and so multiple Fortune 10 companies and and governments uh rely on CGT today to deploy efficient AI. All that to say I want to leave you with this thought right. So AI will be everywhere uh in how we get information and how we make decisions and how we govern. And so we can't really just accept like as a fundamental fact that AI will hide information or skew it. Um we want to turn the ideals of transparency, fairness, accountability to a practical reality and hold AI to a higher standard. Uh insist on explanability and oversight and support the notion that interoperability and auditability should be built in AI systems from the ground up. Uh this is what we're excited to do at CTGT and we believe that transparent AI means a secure future and that's the one we're building towards. Thank you.