AI, LLM and Networking
Bill Gates recently described the development of AI as a creation as fundamental as the microprocessor, the personal computer, the Internet, and the mobile phone. He is most likely right. OpenAI found that around 80% of the U.S. workforce could have at least 10% of their work tasks affected by the introduction of large language models (LLMs). For those in technology, that figure could be closer to 100%.
What does this mean for networking and professionals in the industry? Before AI anxiety sets in and you start looking for a bunker in Montana, consider the short- and medium-term impact: the potential to use AI to increase your output and learn new things. Lately, I’ve been paying more attention to r/networking/ on Reddit, as Twitter a dumpster fire. Almost daily, there’s a network engineer trying to figure out how to transition into the world of automation and software. P
What projects/free-products such as ChatGPT and Bing Chat (bing uses GPT4) present today is a generative Stackoverflow, Google, GitHub or a good IDE in one. I can’t believe I just said Bing. Amazingly, Bard doesn’t even support code generation yet which is pretty wild for as much of the initial research came from Google. You can trick it into leaking code with some work but it’s clear Google is not ready to open the spigot. We could speculate they are waiting until the cost can be modeled appropriately by only releasing LaMDA (roughly GPT3) rather than PaLM which is closer to GPT-4, but it’s clear code generation was not a top priority that it certainly is there now. OpenAI was the disruptor and Google will undoubtedly move mountains to catch up. That said, the amount of data Google can feed models as they develop Bard is staggering, while ChatGPT data feed was apparently scraped a couple of years back. The generated code is probably slightly worse than you would find in a randos repository on GitHub but the speed to derive that nugget that may unblock you can be impressive. If you are convinced AI will eradicate humans anytime soon, than you haven’t seen the bugs it can crank out.
Having experimented with GPT, it’s fascinating from a technological standpoint to engage in a back-and-forth with a machine to see if you can provide enough information for it to generate a solution or code. Most of the time, the code doesn’t compile, but it can be refined through better prompting. Sometimes, I even find myself being extra nice to my AI overlord in hopes of getting a better answer. Yes, we might be doomed. If I just encourage this bot to not be so lazy it can do a better job of self-learning and do what I want it to do. It even gets slightly defensive when you correct it. Angsty AI.. It’s your fault my code burned down the plant, not mine, I’m just a machine.
The exciting part is that the barrier to entry for many in networking is being lowered. Intimidation about moving from day-to-day operations to architecture and learning automation or programming can be daunting, but these new tools can accelerate your learning process. Assembly programming was eventually phased out for most, replaced with a more productive abstraction. I imagine something similar will occur here. Technology, including networking, has scaled through layering abstractions, but we are now reaching a critical mass. The stack is getting deeper every day. New protocols and technologies, such as QUIC HTTP/3, eBPF, or complex service routing, add another layer of complexity. Troubleshooting the stack can take years off your life. In short, AI has the potential to make everyone a bit of a wizard over time. Even the OpenAI team admitted that some of the more profound results from GPT-3/3.5 were not intended, and they still don’t exactly understand how they occurred.
I openly admit that considering how LLMs can be applied to networking and security in infrastructure is incredibly exciting. Having adaptive agents that can constantly search for issues adaptively will change infrastructure and alleviate some of the pressures on us mere mortals trying to build and operate it. I plan to make time to blog about how these tools can be applied to infrastructure, in particular networking and security. It’s a big deal, but it’s still a tool, not a replacement. It’s a great time to cross-train. Having deep domain expertise in networking and security is not something AI can generate like general-purpose coding, so dive in and don’t be intimidated.
Ultimately the ideas will become more important than the personal ability to implement it if LLMs continue down the trajectory they appear to be on. This is great for new comers to software. While it will still be critically important to understand what is happening, the implementation will be greatly accelerated by the AI tooling.
Since AI will kill websites, this was mostly a message to myself in the future to see how wrong I was about under or overestimating AI. So until next time frendos, drop me a line or come collaborate with me in open source via github.com/nerdalert (unless AI models begin replicating themselves, learning from each other and taking natural selection a bit too seriously).
Here are some papers that I recommend if interesting in learning more on LLMs and the potential for your kid to ask GPT how to create a death ray or create a sentient model that will crash the global economy:
- Eight Things to Know about Large Language Models
- GPTs are GPTs: An Early Look at the Labor Market Impact Potential of Large Language Models (warning: this was a very silicon valley centric view of the world)
- The Age of AI has begun
- GPT-4 Technical Report
Since AI could potentially kill websites, this was mostly a message to myself in the future to see how wrong I was about under- or overestimating AI. So, until next time, friends, drop me a line or come collaborate with me in open source via github.com/nerdalert (unless AI models start replicating themselves, learning from each other and taking natural selection a bit too seriously).
