Large Language Models: The opportunity for manufacturers
Unlock operational efficiency and enhance employee well-being
Recent developments in artificial intelligence (AI) will unlock enormous gains in operational efficiency and work-life balance for manufacturers.
To leverage this technology, lean managers should jump to isolate and automate the tasks that steal creative attention and add stress to their employees.
This article will help you understand the benefits of these AI advancements, and how you can start to frame their use in your own operations.
Hint: You already have the key to riches.
Intranets are a goldmine
I started my engineering career in 2012.
I worked as a design engineer at a century-old engine manufacturer. A big part of my work involved finding and making sense of hundreds of documents on the company’s intranet.
Many of them were version-controlled design, manufacturing, and process guides or standards with relevant entries dating back to my grandfather’s generation.
I was being mentored by an engineer who worked for the company in 1940. That long-lost employee was training me.
Your intranet is more than a place for corporate updates. It’s a repository of hard-won knowledge in design, production, and market operations. These documents — RFQs, requirement change orders, design guidelines, testing reports, etc — are a valuable store of intrinsic knowledge about your business. It’s very likely your intranet contains refined knowledge from millions of hours of employee experiences.
Your captured operational knowledge is under-utilized. Your people don’t want to manually search for information!
Using intranets is painful
For any given project your employee might find hundreds of relevant documents in your company’s intranet that can aid them in their work — from training a new employee to defining a product specification from a new customer request. Filtering and making sense of hundreds of intranet documents is not a leisure activity. It is cognitively demanding (and often stressful) work.
Design and manufacturing decisions made with incorrect or inaccurate information can lead to catastrophe.
Knowing this can add significant and continuous pressures to employees looking for decision-critical information in manufacturing documentation.
How can we categorize this pain?
- Short-term memory strain — ‘I just read a 100-page RFQ document — how many of ‘X’ features and ‘Y’ components did they need again?’
- Redundant repetition — ‘This is the sixth time I’ve seen this requirements document this month. What’s changed? Does this new requirement conflict with standard ‘Z’?’
- Prediction uncertainty — ‘Our sales team didn’t ask the customer about ‘X’, but based on the info I have, I think the missing information is ‘M’’
- Content overwhelm — ‘I just joined a new product team, I have 50 reports to dig into, what is important for me to know?’
More generally, the ‘sub-tasks’ that eat time and add strain to employees could be defined as:
- Answering specific questions related to multiple reference documents (in multiple formats)
- Guessing the most likely answer given the learned context
- Summarizing long texts and complex descriptions
- Translating between languages (technical, cultural)
These tasks are perfect for automation with AI.
How can AI help?
Natural Language Processing (NLP) is a fast-evolving sub-field of AI that is focused on enabling computers to understand and process human language. The aim of NLP is not to understand single words, but to understand their context.
Some manufacturing-related NLP applications include:
- Internal and external chatbots — Automated messenger applications integrated with your internal systems could help employees accelerate decisions; could help your customers understand your product; could help your distributors refine delivery and servicing.
- Customer sentiment analysis: Automated monitoring of media to track the sentiment of mentions of your company.
- Reliability engineering: Semi-automate the generation and reading of reports to accelerate the work of quality engineers.
- Supply chain management: Semi-automate the processing of shipping and legal documents to accelerate understanding of your supply chain status.
There are endless possibilities….
NLP systems can (and already do) dramatically accelerate our ability to process natural language. The problem is that creating a performant NLP model to address your problem will require a huge amount of labeled, context-relevant, training data. Refining that training data can be a prohibitively resource-intensive task (ie. extracting clean data from messy PDFs and scans, pulling data together from different archives, etc).
Enter Large Language Models (LLMs). LLMs are a recent breakthrough in AI that enable the creation of intelligent systems with a richer understanding of language than ever before. These systems have demonstrated jaw-dropping new capabilities — including but not limited to the generation of creative texts, solving basic math problems, and the ability to answer reading comprehension problems. How is this relevant?
LLMs reduce barriers to the creation of high-value, bespoke NLP applications by dramatically limiting the amount of context-specific training data that you need to create a useful model.
You can spend more time building applications, and less time preprocessing training data.
The point: Save resources, and use an LLM.
Keep two eyes open
All the impressive achievements of deep learning amount to just curve fitting.
– Judea Pearl
Manage expectations. Applying AI to any problem requires diligent and empathetic work. Understanding the processes at play. Understanding people’s needs. Isolating applicable AI techniques and their limitations. Manufacturing problems are complex. Know that AI is not a magic pill that can solve every problem.
Deep learning ≠ Deep understanding. Don’t conflate predictions with actual understanding. AIs are not ‘intelligent’ as much as they are statistical pattern matching models.
Centralization, bias, and energy usage. Training LLM models is resourc- intensive (see image below). It requires lots of data, computing, and energy. Given their size, it would be too expensive to make your own LLM. This is why most of the biggest models come from tech giants — Google (PaLM), DeepMind (Gopher), Meta (OPT-175B), and OpenAI (GPT-3). Given their potential impact and utility, LLMs compound the centralization of power within this small group of companies. Furthermore, because LLMs are trained on huge biased datasets from the internet, models that use LLMs are systematically biased with ‘pre-trained prejudices’ contained within that data.
The point: Protect the downside. Building with any technology requires a conscious awareness of its shortcomings and potential negative impacts.
The point: LLMs are resource-intensive; this leads to the centralization of their power.
Getting started with your team
Okay, so how might you start applying this tech to your problems?
My team and I just spent the last year helping a large German manufacturer apply LLMs to a high-value document processing task — from exploration to development and enterprise integration.
Here is a summary of how we did this:
1 — Break down your problem workflow
Understand what questions your employees are asking of your data and why. Ask them:
- Why are they looking for information in these documents?
- What specific job is your user trying to get done?
- What information are they looking for?
- How will what they find to be used, and by whom?
2 — Redesign your employee experience
Uncover how the answers to these questions fit into their current workflow. Where is the most friction coming from? Redesign their workflow and user interface experience around these points of friction. Test and iterate this workflow and interface. Keep in mind that users will have unique expectations of what AI can do for them.
3 — Select a pre-trained LLM
LLMs are first pre-trained on a massive amount of varied and un-labeled text (often at the scale of petabytes). The model processes millions of sentences, paragraphs, and dialogue samples and learns the statistical patterns that govern how each of these language elements should be assembled in a sensible order. This training gives them the contextual background to train smaller models with limited samples. The LLM still isn’t very useful for a context-specific task at this point. Now you need to tune it using a smaller, labeled dataset (hundreds of samples) that is relevant to your specific task.
4 — Fine-tune on context-specific data
Collect some tuning data. With LLMs, you can leverage the power of the bigger model by tuning it with much smaller datasets. The LLM you’ve chosen will determine whether or not you need to label it, and how many instances you need — can be only 100s of data points. Previous NLP methods required fine-tuning with thousands or tens of thousands of labeled examples.
5 — Apply your tuned model!
Connect your tuned model to the designed experience and user interface from the workflow and interface you designed in step 2 above.
Test and refine!
Take action
The process above is (of course) simplified. But the reality is that this technology has never been more accessible or generally applicable. The opportunity window for applying this technology is now open, it’s time to act!
Looking for guidance on how to quickly isolate, scope, and prototype automation projects?
Connect with me on LinkedIn
Words by Timotheus Gmeiner, AI Project Director at PCH Innovations. PCH Innovations is a Berlin-based, multi-disciplinary studio that designs and develops exploratory technology.