How Generative AI Pushes
Data Streaming to the Next Level

The Link Between Generative AI and Data Streaming

In a post by The New York Times, it notes that “by pinpointing patterns in that data, an L.L.M. learns to do one thing in particular: guess the next word in a sequence of words”, and that these LLMs/Chatbots “produce new text, combining billions of patterns in unexpected ways. This means even if they learned solely from text that is accurate, they may still generate something that is not.”

But this doesn’t mean that Generative AI is doomed, eternally stuck in a cycle of being polluted by new data and never able to fully determine whether the new data it’s trained on is valid, quite the contrary. This is where we can start thinking about how Data Streaming could possibly come into play. Let’s say we’ve got a cybersecurity stream/system that is checking for potential threats/issues with logins - if we have a user logging in in London/England at 1PM, and 10 minutes later, we have the exact same user login in Perth/Australia. This should be flagged, and the account locked. But what if the problem is more complicated? What if an employee’s laptop is stolen, or someone is hacked in the same city or the hacker uses a VPN?

This is where Data Streaming and Confluent can fix the issues we might have with traditional systems, and solve a major limitation of current security. A traditional system might be able to detect this impossible travel and lock the account, but is a blunt instrument. It’ll stop the action and give a scrap of information, such as “Login Impossible: Account Locked - Distance (London/Perth)” - but this isn’t enough in this modern era of Cybersecurity and Cyber Attacks. We need more.

We need the context behind what’s happening. Rather than performing an autopsy on a breach that’s already happened/existing data stuck in a traditional database or scraping through logs or traces, we can feed live information using Confluent (every login, file accessed, command run in cmd and every system change) into an LLM, which can help us transform our system into a dual-wielding analytical demon of a gatekeeper rather than a silent protector. Now, our Generative AI tool can analyse the stream of events related to this user in question, and not only recognise the impossible leap, but the behaviour of the user as well to generate us a comprehensive security brief that says: “Account Locked (Impossible Travel Detected!!) - additional notes, login attempted to bypass MFA using a known session-hijacking pattern previously seen in this example exploit, here’s a blog that explains this and how to fix it”.

Now, back to Hallucinations. We’ve now got an antidote - but to understand how it works, we need to think about why an AI hallucinates: it’s trying to make a best guess based on old information. It might hallucinate a generic response because it’s not been trained on a specific event yet, and it’s got old data - this is where we end up with confident, but WRONG answers. We can use Confluent to provide a constant stream of information, and using this, AI doesn’t have to guess - it can see the exact metadata and data of your business operations as they happen and it now doesn’t have to imagine what your current inventory levels are, or what your system logs look like because you’re already giving them the context it needs to now accurately answer these questions it may have once gotten incorrect. This gap in logic can now be bridged, and hallucinations reduced and replaced with fact-based analysis of the latest data.

This is the beauty of the marriage of Generative AI and Data Streaming Platforms. The idea of informing LLMs with a continuous source of fresh information is what transforms a basic cybersecurity system into a real-time analytical decision engine. If you’re familiar with Confluent, you’ll know about certain concepts like Data at Rest and Data in Motion, and how Confluent bridges the gap between them, and by using it alongside Generative AI, we reduce the risk of Hallucinations, and improve the reliability of the systems we use.

But you may be asking, “Hmm. That’s not entirely what I was looking for. How does Generative AI better my Data Streaming, how does it benefit Confluent, and how can it serve me rather than my data helping train Gen AI better?”

Let’s dive into that, and learn how they can both lift each other up to new heights.