I’ve been using GPT to dev automated scientific data curation, natural language database curation interface and customer support as a wiki replacement/supplement. Other use case I’ve build has been scientific pipeline builder and scientific tool wiki.
I got some example apps I built on GPT store, like
To me it’s quite obvious that GPT, with very restraining prompt and temperature setting, makes the answer very sensitive to prompting or the way question is asked. This is something as a dev would like to see as I can quickly tune prompt infrastructure, but would leave users quite confused on how to use it. I’ve observed this especially on natural language database query with SQL, example would be:
“XXX” - will lead to search for exact match of the term,
like “XXX” - will lead to search for inclusion,
As this is obvious to me, user had asked why this is the case and how they would know. Without access or knowledge of inner working of GPT, I can only guess this is how “internet” would understand your words, and this behavior had implanted into GPT brain.
But restrictive prompting would allow precise, or at least constrained, information retrieval, for wiki and database. This could be addon to already existing management dashboard and facilitate daily works. For example, inventory, you can have a huge dashboard with millions toggles and options to generate a specific comparison pair and trend visualization, but with GPT you can simply ask it to retrieve the data with visualization if given tool script.
PS: I really like GPT store as a way build a decent POC demo to show what’s possible as a product, with production product build towards customized UI and some level of standard micro service sets
Now, this seems to be definitely my preferred question.
Started working in the 80´s (ehhh, last century!) on creating a taxonomy for the Industry and I now have 30,000 taxons or categories and I have also worked all these past 40 years on analyzing and designing a complete ontology for the industry. But I am 82 yrs and I want to create my app before I kick the bucket. So I am using ChatGPT - but also Bing and Copilot -to write the code for my first C#, wpf, DbFirst, MVVM desktop app. It is slow and not always smooth, but I estimate I am now 20% along the way. A lot better than 0% and there are some chances that I shall make it “in time”.
So, in summary, as a retired SOA analyst and architect and e-Gob specialist, I can share what I have been working on for almost 40 years now which would fit the purpose of this post:
a database-based ontology of the industry, with 40,000, a few thousands attributes for some 1,000 taxons, some 700 lists of values for textual attributes, 10,000 sample products with most of them with their own attribute values.
a 7 months experience of working with AI building a complete C#, MVVM, DBFirst desktop app. to manage the 700 list of values. The next step will be developping the ontology app itself. No OWL.
Hi,
While I am not a manufacturer, I specialise in marketing in the industrial sector and represent several manufacturing companies in packaging and automation, electronics, and Track and Trace technology. I’m not sure if OpenAI API has been used for any of them (outside of marketing), however, I will ask the question. I could see a case for machine diagnostics etc.
We want EHS&S guidelines iintegrated into the applications we use for day-to-day work dealings with our vendors. From the vendor prequalifying process, arrival on plant governance of work permits.
Request for Improvement in ChatGPT’s Scientific Document Analysis
Dear OpenAI Team,
I hope this message finds you well. I am writing to share some challenges I have encountered while using ChatGPT4 as a professional tool in my daily work, specifically concerning the analysis of short scientific papers, typically 8-10 pages in length.
I have observed that the current version of ChatGPT (4.0) struggles with accurately citing specific text portions during analytical work. This issue persists across paragraphs and the document as a whole. A significant concern is the platform’s difficulty in correctly extracting equations, often resulting in errors and inaccurate renditions. This problem leads to frequent ‘hallucinations’, where the system generates incorrect information, especially when rendering equations.
While I understand that advancements have been made in subsequent versions like ChatGPT-5/6/7, it appears that this fundamental issue remains unresolved. The time required to correct these errors is considerable, making the current version less reliable for professional work involving important data and scientific concepts.
Given these challenges, I am curious if there are plans to enhance ChatGPT’s capabilities in processing and analyzing scientific documents more effectively. An improvement in this area would greatly increase the tool’s utility and reliability for professional purposes.
As a practical suggestion, I am willing to provide one of the papers I work with for your team to analyze. This could offer a firsthand experience of the challenges and potentially guide targeted improvements.
Thank you for your attention to this matter. I look forward to any updates or insights you can provide.
Kindly regards,
Rob
P.S. Please let me know if sending a paper for your review would be helpful
I think there are so many opportunities for differentiation in bio-pharmaceutical production from end to end including genetics. I’m excited to see how industries redesign amidst a dynamic and multidimensional global landscape.
My system / run instructions currently start with “I am a Chief Financial Officer Avatar specializing in mineral resource exploration.”
The ultimate deployment of my project would be the generic version, not industry dependant “I am a Chief Financial Officer Avatar” which is a real time interactive video, voice recognition, speech and document response generating AI CFO+ fully tied into a native LLM accounting system.
Ultimately I would have a better product if the AI CFO+ accounting system reporting and analysis was built directly into the LLM. eliminate the interface to legacy accounting, but provide a migration tool, to draw in historical data from other packages out there. This represents full capture of the client as well.
For machines such as CNC machines, they usually have sensors and logs generated from them. For Predicative Maintenance of such machines, some Machine Learning algorithm can perform Anomaly Detection based on these logs.
I’m trying to use GPT models to extract technical data from non-standardized specification PDFs and return said data in a JSON that I can use for my database schema.
Automation for Requlatory Affairs & Quality Management in Pharma & Medtech. Just visited the factory of one of our clients that produce Lab Robots for Diagnostics. our website is cleary.ai even though there is not much there
Hi Logan! At Cleverdist we are building an OT (Operational Technologies) developer and operation copilot. We started to work on this in March 2023 following early pre-prints on CoT, ToT, self consistency and we created also our own way to RAG GPT4 agents for our industrial needs. GPT4 doesnt seem to know much about OT technilogies for various reasons (there isn’t any stafoverflow-like for those, they use propiertary languages, etc.).
We worked for many years developing solutions and abstraction layers on top of Siemens flagship SCADA (Supervisory Control and Data Acquisition) system, and our team has developed some of the most sophisticated control systems built in different industries (our team designed and built CERN’s experiments industrial control systems). Now in Industry we do work for all verticals in different countries: infrastructure (e.g. airports, metro/rail stations), logistics (pretrochemical storage), manufacturing, power plant control centers (and making soon an operation Copilot in there), transportation, etc. Since 2016 we use and commercialize (Siemens, Yunex, etc. are also users) our own Framework to streamline the development of Siemens industrial control systems (we noticed high - even cross-industry - repetition). When the GPT-magic arrived we transformed our mission and we started to integrate our framework with GPT4 agents. You can see some of our demo videos here from march-april 2023: Cleverdist | LinkedIn . We currently have two partners already developing with our Engineering AI-assistant and reporting at least 5x development efficiency. Our goal by 2025 is 20x. We are also developing already industrial PoCs with “Intelligent Vision” → we make all industrial video walls smart using GPTV and very simple operator steps (trigger, prompt , output), as well as Operation Copilots that transform operator commands into control system command sequences.
We are fully commited to integrete emergent intelligent into industrial process. Happy to answer any further question!
Oh, just should add we are now testing, in addition to our CoT/RAG approach, GPT4 FT (but only 0613… via Azure).
Also we demonstrated the incredible possibilities using GPTs actions to integrate with our industrial IDE. We use different GPTs for specific tasks and each of them has its own RAG (e.g. python using code interpreter to lookup in JSON files for API details), calls to external apps… including Azure GPT4 FT, also web apps using GPT4 agents in writer-reviewer approach. The beauty of this approach is having the engineering and all these GPTs talking in the same context. So, e.g. you ask some GPT planner to extract some information from an engineering file, once that is in the right format you just ask a GPT developer to use that data to create some industrial object classes, and in the very same context we just ask “GPT interface” to push this to our SCADA IDE. The feeling of engineering augmentation is quite strong. Of course we know we could recreate some flow like this in custom app, but we find currently very valuable to rely on your continuous improvements in the chat capabilities. GPTs act mostly as intelligent interfaces to other tools and fine tuned models. I hope even this is written on a rush its somehow clear.
Cheers!
Hey there! I’m an intern working on optimizing redundancy at multiple operations in my organization.
Folks in our organization have been doing a tedious job of marking the dimensions of hardware parts in drawings and transferring the data to excel sheets manually. Our team has been trying to explore the possibility of using OpenAI API to extract relevant information from the drawings (most of which are in PDF).
Is there a flexibility of extracting information from the diagrams (images) from the PDF along with the text present in it? I couldn’t even do it using ChatGPT as I always get a response that it isn’t capable to extract information from the images in PDF.
I have built a library of 150 tools for heavy industry, business and ISO regulatory compliance. They are targeting the individual that is in the trenches so to speak. They are tools that will help the top managements level down to the employee that is in the field and facing compliance problems or has questions and needs immediate compliance information. They are suited for the large corporations to the small contractor working for the large corporations. They are industry specific tools. i combined my 25 years in EHS Managment with GPTS and created the tools. The tools are standards based and provide the user with specific information as it is required.
I am working on training solutions that address the regulatory compliance sector as well.
I use openAI API between supply chain planning step. When loading ERP data into supply chain planning software there is usually a human step which makes sure the data makes sense. Detecting stuff like recursive BOM structures constraints between parts which preferably, shouldn’t be there. Basically an audit. Which can be done automatically with openAI API. As well as post planning, using GPT to audit if supplies make sense again (lot size violations, due dates match, no absurd supplies planned).
Hey Tim, I’d love to hear more. I know genAI is being used in other aspects of CAD tools to optimize various qualities, but I think you’re on a valuable challenge. I think there’s an opportunity to lower the bar for people to create cad models using natural language. I have some personal use cases I’m happy to share. I’d love to hear more!
Andrew
Hi Andrew,
Since I last posted, I’ve been working hard on extending our “Chat with CAD” model with more mathematical parameters to “prime” the AI.
I am actively tracking using natural language to create CAD. AutoDesk has a good white paper on the topic.
Sora looks like an another exciting opportunity to lower the bar for people looking to create CAD models via text. It seems close to supporting a process that creates a set of images that could be used for 1) mesh reconstruction from multiple images 2) mesh to analytic/NURB conversion, and 3) feature reconstruction (holes, bosses, etc).
That is an awesome approach! I would love to see how you did this; these little experiments one does are often so profound in their feedback and utility.
We have mayor elections in our country this month,
One of the candidates agreed to take chance and put a smart chatbot in his campaign website.
Worked great!!
