What’s an LLM? Josef Co-founder and Solutions Architect Kirill Kliavin explains

Fresh from building Josef Q, a new AI-powered platform that transforms policies and regulations into digital Q&A tools in minutes, Josef Co-founder and Solutions Architect Kirill Kliavin helps demystify LLMs, explains their utility for legal professionals, and looks toward the technology’s future.   


Kirill Kliavin, Co-founder & Solutions Architect.

Kirill Kliavin, Co-founder & Solutions Architect.

Hey Kirill, congrats on the beta launch of Josef Q! For those who are unfamiliar with the technology behind it, could you share what an LLM (large language model) is?

For sure. At a high level, LLMs are complex, machine-learning algorithms that are trained on large amounts of text. Based on their training, LLMs can read strings of text and then predict and generate the most probable sequence of words or characters that should then follow.

This capability makes them the perfect tools for responding to user prompts, like questions, for example. 

Depending on the use case, LLMs work with accompanying knowledge bases such as policies, a selection of articles, dictionaries, or, in cases like ChatGPT, whole sections of the internet.

This means LLMs aren’t sentient AI beings as people tend to believe, but mathematical systems that use statistical probability to generate the most accurate and helpful responses when prompted by users.

What common tasks do LLMs perform?

All sorts, really.

LLMs can be used for translation purposes like in Google Translate where, firstly, the LLM interprets a string of text into words (or so-called “tokens”) and then locates the most similar word in another language to then perform the translation. This is what enables tools to turn  “Hello” into “Hola”, for example.

Summarisation is a popular LLM application too, with tools like the recent Notion AI being able to, amongst other things, read a body of text selected by the user and generate a summary featuring the most important points.

Email clients like Gmail also use an LLM for text generation. The LLM reads the existing content body and suggests a subsequent line of text, such as “Please let me know if you have any questions.” When you receive text suggestions in Gmail, that’s an LLM doing the work.

Elsewhere, LLMs are used for image generation in tools like Midjourney or Dall-E, they’re capable of transcribing video imagery for those with vision impairments, and they can also generate auto-captions as we see on sites like YouTube.

“LLMs aren’t sentient AI beings. They’re mathematical systems that use statistical probability to generate the most accurate and helpful responses when prompted by users.”
– Kirill Kliavin, Co-founder & Solutions Architect.

That’s a lot! What about legal applications? How are LLMs suited to the legal industry? 

Okay, well the legal industry is known for its dense documentation and jargon.

Together, this limits accessibility to the law and causes inefficiencies when working as part of legal teams. Granted, information may be located in policies or regulations, but it’s often hard to find and understand what it actually means.

Using tools like Josef Q–which is itself powered by an LLM–information can be located, interpreted and re-articulated for users at speed and in simple, understandable ways.

There are clear benefits to this.

  1. Tools that use LLMs can help people spend less time manually searching through local 100+ page documents;
  2. Important documentation can be centralised in one place; and
  3. Such information is made available on-demand as and when required.

For those who might want to build their own tool using an LLM, what’s a good way to check if a certain LLM is suited to a specific task?

First would be to fully understand the problem you want to solve.

There are lots of LLMs out there and each are designed for different purposes, using different processing methods. So, first up, know your problem so you know what capabilities your LLM will need to have. If your task involves language translation, your LLM will need to be trained on different languages.

Next, do some blind testing. See if the LLM can do what you want it to do and find out what it can’t. By pushing its limits, you’ll get a good understanding of its suitability.

There are obviously other considerations to keep in mind, such as access (is it open access?), cost (what’s the licensing like?), and privacy, etc. (how does it handle user data?), but if you’re just starting out, begin with those two.

Ok, final question! Looking forward, how do you feel about the technology’s future?

As an engineer, I love it when new tech is released that improves our day-to-day lives, whether it’s at work or at home.

There’s a whole history of technologies that have enabled this and I think LLMs will have a similar impact: they’ll pick up those low-level, menial tasks that we’d rather not do so that, at the end of the day, we’ll have more time for more fulfilling, creative work.

Previously, in times where books and libraries were our main source of accessing information, memorisation was a crucial skill. Now, with the internet and mobile devices, our ability to find information is more important, right?

We don’t really know what skillsets will fall in and out of favour as the AI space develops, but change is coming. By looking at the impact LLMs have had on accessibility already, I’m confident that what’s lying ahead will bring positive change.

That’s great. Thank you, Kirill!

No worries, thank you!


To see what LLMs can do for legal professionals, join the experiment and test the beta version of Josef Q here.