Alright, my mates, I’m again with one other publish primarily based on my learnings and exploration of AI and the way it’ll match into our work as community engineers. In at this time’s publish, I wish to share the primary (of what’s going to probably be many) “nerd knobs” that I believe all of us ought to pay attention to and the way they’ll influence our use of AI and AI instruments. I can already sense the thrill within the room. In any case, there’s not a lot a community engineer likes greater than tweaking a nerd knob within the community to fine-tune efficiency. And that’s precisely what we’ll be doing right here. Superb-tuning our AI instruments to assist us be more practical.
First up, the requisite disclaimer or two.
- There are SO MANY nerd knobs in AI. (Shocker, I do know.) So, for those who all like this sort of weblog publish, I’d be pleased to return in different posts the place we take a look at different “knobs” and settings in AI and the way they work. Effectively, I’d be pleased to return as soon as I perceive them, at the least. 🙂
- Altering any of the settings in your AI instruments can have dramatic results on outcomes. This contains rising the useful resource consumption of the AI mannequin, in addition to rising hallucinations and lowering the accuracy of the knowledge that comes again out of your prompts. Take into account yourselves warned. As with all issues AI, go forth and discover and experiment. However accomplish that in a protected, lab surroundings.
For at this time’s experiment, I’m as soon as once more utilizing LMStudio working domestically on my laptop computer fairly than a public or cloud-hosted AI mannequin. For extra particulars on why I like LMStudio, try my final weblog, Making a NetAI Playground for Agentic AI Experimentation.
Sufficient of the setup, let’s get into it!
The influence of working reminiscence dimension, a.ok.a. “context”
Let me set a scene for you.
You’re in the midst of troubleshooting a community problem. Somebody reported, or seen, instability at some extent in your community, and also you’ve been assigned the joyful process of attending to the underside of it. You captured some logs and related debug info, and the time has come to undergo all of it to determine what it means. However you’ve additionally been utilizing AI instruments to be extra productive, 10x your work, impress your boss, all of the issues which can be happening proper now.
So, you determine to see if AI may help you’re employed by way of the information sooner and get to the foundation of the difficulty.
You fireplace up your native AI assistant. (Sure, native—as a result of who is aware of what’s within the debug messages? Greatest to maintain all of it protected in your laptop computer.)
You inform it what you’re as much as, and paste within the log messages.
After getting 120 or so traces of logs into the chat, you hit enter, kick up your ft, attain to your Arnold Palmer for a refreshing drink, and anticipate the AI magic to occur. However earlier than you’ll be able to take a sip of that iced tea and lemonade goodness, you see this has instantly popped up on the display screen:
Oh my.
“The AI has nothing to say.”!?! How may that be?
Did you discover a query so tough that AI can’t deal with it?
No, that’s not the issue. Take a look at the useful error message that LMStudio has kicked again:
“Making an attempt to maintain the primary 4994 tokens when context the overflows. Nevertheless, the mannequin is loaded with context size of solely 4096 tokens, which isn’t sufficient. Attempt to load the mannequin with a bigger context size, or present shorter enter.”
And we’ve gotten to the foundation of this completely scripted storyline and demonstration. Each AI device on the market has a restrict to how a lot “working reminiscence” it has. The technical time period for this working reminiscence is “context size.” Should you attempt to ship extra knowledge to an AI device than can match into the context size, you’ll hit this error, or one thing prefer it.
The error message signifies that the mannequin was “loaded with context size of solely 4096 tokens.” What’s a “token,” you surprise? Answering that might be a subject of a completely totally different weblog publish, however for now, simply know that “tokens” are the unit of dimension for the context size. And the very first thing that’s completed while you ship a immediate to an AI device is that the immediate is transformed into “tokens”.
So what can we do? Effectively, the message provides us two potential choices: we are able to enhance the context size of the mannequin, or we are able to present shorter enter. Typically it isn’t a giant deal to offer shorter enter. However different instances, like after we are coping with massive log recordsdata, that possibility isn’t sensible—all the knowledge is vital.
Time to show the knob!
It’s that first possibility, to load the mannequin with a bigger context size, that’s our nerd knob. Let’s flip it.
From inside LMStudio, head over to “My Fashions” and click on to open up the configuration settings interface for the mannequin.
You’ll get an opportunity to view all of the knobs that AI fashions have. And as I discussed, there are a whole lot of them.
However the one we care about proper now could be the Context Size. We will see that the default size for this mannequin is 4096 tokens. But it surely helps as much as 8192 tokens. Let’s max it out!
LMStudio offers a useful warning and possible motive for why the mannequin doesn’t default to the max. The context size takes reminiscence and sources. And elevating it to “a excessive worth” can influence efficiency and utilization. So if this mannequin had a max size of 40,960 tokens (the Qwen3 mannequin I take advantage of generally has that top of a max), you may not wish to simply max it out immediately. As an alternative, enhance it by a bit at a time to seek out the candy spot: a context size sufficiently big for the job, however not outsized.
As community engineers, we’re used to fine-tuning knobs for timers, body sizes, and so many different issues. That is proper up our alley!
When you’ve up to date your context size, you’ll must “Eject” and “Reload” the mannequin for the setting to take impact. However as soon as that’s completed, it’s time to reap the benefits of the change we’ve made!
And take a look at that, with the bigger context window, the AI assistant was capable of undergo the logs and provides us a pleasant write-up about what they present.
I notably just like the shade it threw my method: “…think about searching for help from … a professional community engineer.” Effectively performed, AI. Effectively performed.
However bruised ego apart, we are able to proceed the AI assisted troubleshooting with one thing like this.
And we’re off to the races. We’ve been capable of leverage our AI assistant to:
- Course of a major quantity of log and debug knowledge to establish potential points
- Develop a timeline of the issue (that will likely be tremendous helpful within the assist desk ticket and root trigger evaluation paperwork)
- Establish some subsequent steps we are able to do in our troubleshooting efforts.
All tales should finish…
And so you might have it, our first AI Nerd Knob—Context Size. Let’s evaluation what we discovered:
- AI fashions have a “working reminiscence” that’s known as “context size.”
- Context Size is measured in “tokens.”
- Oftentimes instances an AI mannequin will help a better context size than the default setting.
- Rising the context size would require extra sources, so make modifications slowly, don’t simply max it out utterly.
Now, relying on what AI device you’re utilizing, chances are you’ll NOT have the ability to modify the context size. Should you’re utilizing a public AI like ChatGPT, Gemini, or Claude, the context size will depend upon the subscription and fashions you might have entry to. Nevertheless, there most positively IS a context size that can issue into how a lot “working reminiscence” the AI device has. And being conscious of that reality, and its influence on how you should use AI, is vital. Even when the knob in query is behind a lock and key. 🙂
Should you loved this look below the hood of AI and wish to study extra choices, please let me know within the feedback: Do you might have a favourite “knob” you want to show? Share it with all of us. Till subsequent time!
PS… Should you’d wish to be taught extra about utilizing LMStudio, my buddy Jason Belk put a free tutorial collectively referred to as Run Your Personal LLM Regionally For Free and with Ease that may get you began in a short time. Test it out!
Join Cisco U. | Be a part of the Cisco Studying Community at this time at no cost.
Be taught with Cisco
X | Threads | Fb | LinkedIn | Instagram | YouTube
Use #CiscoU and #CiscoCert to hitch the dialog.
Learn subsequent:
Share: