Last week brought some AI news from China. A new LLM model called DeepSeek R1 was released and caused havoc in the US AI markets.
I’m not going to go down the rabbit hole of why the markets crashed and what DeepSeek has claimed to overcome. I want to address the more significant issue that might be keeping a lot of companies from investing in the power of AI. The problem is that this market is volatile and changing rapidly.
So, how do you know which technology to launch your AI project on?
You don’t. That’s the reality.
Between the time DeepSeek R1 was released and when I wrote this article, OpenAI has already released a new model that seems to be on par with DeepSeek R1.
Innovation is happening at a ridiculous pace in AI right now. But you can benefit from the innovation rather than stand back until the dust settles.
Why DeepSeek Matters
Let me back up a little and explain why DeepSeek R1 is a big deal from an LLM standpoint.
LLMs have to be trained first. Then, as we use them, we use what’s called “inference.” All you need to know is that it takes a lot less computing power for inference than it does to train these models initially. Normally. DeepSeek R1 is the first non-proprietary “reasoning model.”
Let me break that down: DeepSeek R1 was released as open source. That means you can download the model and run it yourself.
It also means you can use their model as a basis for your product, if you so desire. This is subject to standard limitations of the license.
By comparison, OpenAI’s reasoning models are only available from OpenAI. So this is a big deal for the thousands of companies that want to run their LLMs inside their data centers rather than in public clouds.
So, What’s a Reasoning Model?
This is where the absolute bleeding edge of LLM research is being done.
If you’ve used any LLM like ChatGPT (which isn’t an LLM; it’s the chat client that you use to talk to the LLM model), you’ve seen limited “reasoning.” For instance, asking ChatGPT simple questions does seem like reasoning.
But the distinction here is that ChatGPT and all non-reasoning models are just spitting back the most probable answer based on the data they were trained on.
It sure looks like human reasoning, but it is not. Let’s make that clear: LLMs don’t “know” anything. They are just using probabilistic mathematics. Very successfully, I might add.
Reasoning models, however, use a few different methods to try to get even better data back to you, specifically in areas that require more complex pattern recognition.
They “think” by iterating through the prompt that was sent to them. This method is called “Chain of Thought.”
If you’ve used DeepSeek R1, you see it “thinking.” It breaks down your prompt and works on each section before coming to a complete answer. This method of “Chain of Thought” is better at answering complex questions.
Why Chain of Thought Isn’t Useful
For example, I asked DeekSeek R1 to “Tell me a random obscure fact.“
In defense of DeepSeek R1, this wasn’t fair. I kept the conversation because it highlights exactly why these models aren’t suitable for specific tasks.
The model generated 4,400 words trying to decide what obscure fact to tell me and whether the fact was “obscure enough” to satisfy my simple prompt.
I had to cancel the response after it generated 4400 words. I assume it would not have stopped.
It was highly entertaining but ultimately useless. Depending on what you ask DeepSeek R1, it’s like dealing with a schizophrenic with ADHD. It’s not fun if you need a simple answer like “What is the date on this document?”
TL;DR
Which brings me, finally, to the point of this article. Data extraction doesn’t need a reasoning model. It requires models that adhere to the prompts.
You don’t want a model to “think” about a simple question about data that is literally in the data it was sent. That’s how we work with LLMs. We send the data in scope (usually a text segment of a document) and then ask questions that are geared towards extracting data from that text.
Thinking models will slow down this entire process. It will take longer for a model to respond if it’s off second-guessing itself or babbling on with 4,000-plus words.
I’m not saying there’s no use for reasoning models within Grooper. These could be precisely what we need to take us to the next level in contract analysis and evaluation. But that’s yet to be seen.
As I mentioned, this is bleeding-edge technology.
Our AI Assistants use a method of “Chain of Thought” and may benefit significantly from a reasoning model.
Testing DeepSeek with Grooper
However, my point is that Grooper can already take advantage of these advances. I was able to test DeepSeek R1 across several providers and a local-hosted model to determine whether it would work in Grooper in just a few minutes.
We’ve designed Grooper to take advantage of these breakthrough technologies. So, testing a new model is as simple as pointing to it.
Finally, Grooper protects you from selecting the wrong AI technology.
We can help you decide how AI can work for you:
- What models
- How they need to be deployed and secured
- How we can tweak their results to get you even better results.
I can’t stress enough that if you haven’t seen the latest from Grooper, you’re missing a lot!
LLMs have made it easier than ever to configure Grooper and easier than ever to get better data from your documents.
As always, feel free to contact me or your sales or technical contact at BIS. We’d be happy to show you what’s new and exciting.
See below for information on our next Live Event, where we will show AI Assistants and data from an ECM system.
