GENERATIVE AI - AI NETWORKING

October 2024 Issue

This page is being developed with additional material from the article published in ISE magazine in October 2024:
“AI Networking: Training the Unruly Child.”

Introduction

This page covers the intersection between AI Networking and cybersecurity. It will continue to be developed as use and possible abuse of Generative AI as it matures:

  • Personal experience: Legacy AI and CybrScore Software Today 
  • Challenges: Positives and Negatives
  • Show me the Intelligence
  • AI Networking Work in Progress at ONUG
  • AI Networking Summit (October 23-24)
  • AI-based Cybersecurity threats (coming shortly)
  • GenAI Pretenders and Contenders (coming shortly)

Personal experience: Legacy AI and CybrScore Software Today 

Thirty years ago, I founded an AI-based software company. AI was described as “knowledge-based systems” back in those days. From a basis of sales knowledge and experiential data, I programmed the software to make recommendations to improve sales. It measured and forecast performance, adapting as the sales of individuals and the company progressed.

Critically, how I made it intelligent was my applying sales experience, judgement and adaptation to results achieved that was programmed into the software. The complexity of today’s evolving networks and massive increase in compute power plus cybersecurity threats make it a whole different world. Artificial Intelligence has been around for a while –using data and adapting to change are not new.

The application of AI to networking and security is what had me write this article since today’s complexity and opportunities for Network as a Service, network management, data curation and cybersecurity make the exploration a necessity. It seems like an exploration at this point because although there are attractions—cost savings and responsiveness to changing conditions—there are significant unresolved issues that the article raises yet still it looks optimistically at the potential possibilities.

CybyrScore

Similarly, Cybyr.com’s CybyrScore software is an interactive dialog with a user that provides prioritized recommendations, remembers results, assesses progress, measures risk reductions can be customized to match the target systems and includes multi layered defenses. It also is sensitive to costs.

 

Challenges

If I could give you ten actions that would ensure your network meets your business goals, you’d want to know about them, right? However, if I said one of them made no sense but I’m not going to tell you which one—you’d think I was crazy and definitely not intelligent. Imagine if that applied to a critical part of your network infrastructure.

Should you apply data curation, network service performance or transaction security actions that were incorrect—then you would be putting your organization at great risk.

Here’s an actual example. My wife asked CoPilot how to do something on her Galaxy phone. CoPilot claimed it knew about the latest software loaded on her phone—but the instructions it gave were incorrect. When asked why, its many apologies included: “I have certain limitations and potential weaknesses in how I adapt to new information.” Imagine if that applied to a critical part of your network infrastructure. The idea of rapidly adapting by learning about changes to policy service paths and threats discovered in real-time is the challenge being addressed by the concept of Adaptive AI. This anecdote appears to show that the lesson is still being learned. Yes, the interactions with copilot seem to be evolving – now it seems to want to be our friend and slip in colorful exchanges to show it’s a nice member of the human community! Really? No thank you.

AI: Show me the Intelligence:

The main challenge for Gen AI is, is it’s burdened by the word and the expectations of Intelligence. At this stage of development, it already has some outstanding characteristics, bring understanding of learned concepts, coding, and plain language explanations. It’s much more than the knowledge-based systems I referred to at the top of the page. However it does not meet our definition of intelligence by any of the obvious criteria. It also cannot currently meet the expectation of rapidly adapting to new information.

So, can GenAI be developed to correctly understand, grasp and reason –our definition of intelligence? Can it apply judgement adapting to our real-time experiences in the context of our business goals? How can it deal with the unacknowledged emotional aspects of every decision we make? We shall see! Suddenly, it doesn’t sound so intelligent.  

The Intelligence Fantasy

The most interesting part of GenAI is what I call “the intelligence fantasy.” CoPilot is not artificial intelligence – I don’t know what it is – but it’s not intelligent. Yes, it has some aspects: aggregating information and providing answers to questions such as “how I do X or how do I rework code from language X to language Y and other training type info, such as what the top XDR products?. That’s all great, very useful and there’s nothing wrong with it (except when it doesn’t give a valid answer).However, it does not display actual intelligence as we know it. I.e., it lacks these critical qualities

  1. Understanding: grasping the significance in a wide the context of our business and technical issues.
  2. Applying reasoning and judgement to make decisions that are relevant right now and not from even the recent past let alone last year. (Yes our brain also predicts the future from the brain patterns it recorded from our past and tells us what to do based on survival – but that’s not intelligence either.)
  3. It’s not transformative or innovative (coming up with game-changing new ideas)
  4. The user’s actual relevant human experience is not taken into account
  5. Real-time adaptation to my results as they are achieved and impact on future decisions
  6. Being curious and asking questions in order to give thoughtful recommendations. ChatGPT is only half a dialog: you ask questions, it gives answers – CoPilot has no curiosity and never asks you anything. Even your dog asks you (in its own way) “Is there anymore food?” “Can we go for a walk, now?”
  7. Understanding that most decisions have a critical emotional context!! – Not a common discussion about decision-making but brain science tells us that we store emotions together with facts about events!!! If you don’t agree then think about how your technical decisions are made with not wanting to look stupid, or wanting to be admired so you get promoted, etc. It’s the filter that CoPilot does not have because it’s not human and has never executed anything.

Sure its very useful, I use it all the time but intelligent, it aint. Agree? Disagree? Thoughts on what GenAI really should have been called: when I have written “Knowledge-based Systems” in the past I programmed in my personal experience, judgments and knowledge as the basis of recommendations and measurement of success. Could GPT have such data input? Possibly. 

Always Look on the Bright Side of Life

If I could give you ten actions that would ensure your network meets your business goals, you’d want to know about them, right? However, if I said one of them made no sense but I’m not going to tell you which one—you’d think I was crazy and definitely not intelligent. Imagine if that applied to a critical part of your network infrastructure.

This page is not intended to be a Luddite rant, since I do think that this is an exciting journey for all of us. However, before we get to the positive aspects there’s still the issue of cybersecurity threats penetrating the Large Language Models (LLMs). I.e., how do you know if the data you are basing your decisions on is inaccurate or deliberately misleading. For instance, if a threat actor deliberately distorts data about an asset’s authentication or privilege, then your network and organization is now toast.

Since we just started the AI Networking journey, let’s assume that these challenges will get sorted out quickly, making AI smarter and safer. Protection against the abuse of LLMs is already under way. The questions are what are the most promising candidates to simplify and handle network complexity? Are they cost-justified, and how can we apply some actual human intelligence, our experience and expertise to train this unruly GenAI child?

AI for Network as a Service (NaaS)

Since my article in the May/June issue, NaaS has made rapid progress. At the core is the need for services to be on-demand, available from multiple providers without lock-in, in multiple locations, accessing workloads in multiple clouds with varying performance requirements.

The advantage to any service providers or systems integrators who could offer such a service is huge, since it meets all of the enterprise requirements. An Adaptive GenAI system—one that constantly retrains as data changes—may be the only viable, scalable approach. Such a system could gather and update the information accessed either via APIs or a portal.

Managed Service Providers would be served well too, getting access to the services offered on behalf of their users. For those who are becoming aware of various. The race is on to develop this approach as it addresses the most challenging aspect of Network as a Service.

The use of AI Networking to the Orchestration/Automation of NOC/SOC (Network and Security Operations is one of the projects of the Open Networking User Group Collaborative (ONUG.net), where I participate. An interesting part or this work covers the choices of how to train LLMs or SLMs (industry/context specific Small Language Models) using publicly available services, GenAI augmented tools or custom solutions developed in-house.

Other Promising Areas

  • Policy as code : Using GenAI to bring Zero Trust principals to automate access, identity, authentication, privilege control, policy enforcement and continual monitoring.
  • In ONUG’s Cloud Secure Notification Framework project, GenAI could oversee and report on large quantities of different network data/monitoring/networking incident reports involving Cloud-based workloads.

There are many more. Some suggest immediate actions and others are currently under investigation. Speaking of which, the publication of this article is timed to alert readers to ONUG’s Fall event where industry luminaries give the very latest on their work and their perspectives on the state-of-the-art of AI Networking. If it’s anything like the ONUG Spring event, you do not want to miss it. Check it out at https://onug.net/.

Conclusion

Despite all the challenges, I am confident that our ingenuity will overcome the cautions on this page. Some say the AI bubble will burst. If it does, then it will mean we didn’t make the most of the opportunities to create a better world of networking and far beyond. This story, including updates and insights from the ONUG event, will continue here.

#63B4D1

#1D1E18

#6C6F7F

#55917F

#FFFED21

#E1F0C4

#211F63

[1] https://onug.net/onug-collaborative-blog/evaluating-different-approaches-for-adopting-generative-ai-in-network-and-security-operations/

AI Platforms of note: Pros and cons

Reflection 70B

To quote itself: “Reflection 70B has made headlines for its ability to outperform some of the most well-known proprietary models, including GPT-4 and Claude 3.5 Sonnet. On a wide range of benchmarks, this open-source model has shown remarkable results in reasoning, math, and general knowledge tasks. These performance gains are attributed to its ability to engage in self-correction and fine-tuned reasoning, setting it apart from previous models in the space.” Others have discredited it as a fraud.