Cognite gets a prize for being fast out of the starting blocks with (inter alia) an ‘Open Letter to Customers’
from CEO Girish Rishi who gushes, ‘This is the moment that Cognite was
built for’. AI and large language models are the ‘next horizon for
democratizing data’. However, ‘almost accurate’, ‘probabilistic’ data
(read ChatGPT) produces ‘hallucinations’. Moreover, ‘putting
proprietary enterprise data or into the public domain is nothing short
of corporate blasphemy’. The answer, is Cognite AI,
a suite of generative AI capabilities atop Cognite’s Data Fusion
industrial data ops platform. Elsewhere, Cognite CTO Geir Engdahl
expects ChatGPT to ‘finally unleash the iPhone moment in digital transformation’.
Siemens has likewise reported on a Teamcenter app for Microsoft Teams that leverages the Azure OpenAI Service to help write code* for factory automation. Teamcenter for product lifecycle management now leverages the OpenAI language models and other Azure AI capabilities. Interestingly, Siemens’ system is bi-directional. Engineers can report product design or quality concerns using natural speech which is parsed by OpenAI, summarized and routed to the appropriate expert. The system also has a multi-lingual capability. Siemens and Microsoft are working to accelerate programmable logic controller code development, generating PLC code from natural language inputs.
* Using ChatGPT to write code was also featured in our report from the Rice HPC in Energy conference elsewhere in this issue.
There is also some interesting pushback from the NLP* crowd on ChatGPT. Notably from Arria whose business involves creating text from data and queries, rather like CGPT in fact. Arria recommends going with the flow, adding its language generating capability to ‘emerging’ generative AI technologies, like ChatGPT. However, Arria warns there are clear differences between the two technologies. CGPT is not a solution for addressing the complex, mission-critical challenges that businesses are currently facing. Arria’s models are ‘predictable, controllable, auditable, and 100% accurate’. CGPT’s accuracy is limited and results are ‘unpredictable’.
* Natural language processing.
‘Everyday AI’ boutique Dataiku opines that large language models (LLM) plus its own platform make up the perfect pairing. LLMs can be used in the enterprise either by making an API call to a service or by downloading and running an open-source model in locally. The company argues that LLMs can be computationally intensive and that a smaller language model addressing a particular task may be better. Dataiku has trialed OpenAI’s GPT-3 model to query the contents of its own documentation, knowledge base and community posts. The results were ‘impressive’, providing easy-to-understand and helpful context. Users reported that it was more effective than simple links to the ‘highly technical’ reference documentation. The company has also proposed a generative AI cookbook using an open-source LLM (Hugging Face) in parallel with its own API.
If you really want to get under the hood of the generative AI engines you should read Stephen Wolfram’s writing on ‘What is ChatGPT doing and why does it work’. A word of warning though. Just like looking under the hood of your car, you may not understand all that you see!
For the DIY enthusiasts, and admittedly more vanilla than generative AI you might like to see what is possible today using entry-level hardware and a few lines of code. Nurgaliyev Shakhizat, writing on Hackster presented TinyML, an always-on audio classifier using synthetic data’. The system uses machine learning to recognize and classify audible events. Hardware consists of an Ardunio Nicla Voice board. Shakhizat’s setup is designed to recognize a name in ambient speech, but one can imagine such a system being used to capture acoustic anomalies from machinery. Training was performed on the Edge Impulse platform before model deployment on the Nicla. The latter comes with a package of sensors. Along with the microphone, it features a smart 6-axis motion sensor and a magnetometer, ‘making it the ideal solution for predictive maintenance, gesture/voice recognition and contactless applications’.
Comment. Despite the generally held view, the ‘generative’ AI phenomenon did not arrive in virgin IT territory. Many companies have leveraged what was up to know known as natural language processing to attempt to extract useful information from a text corpus. ChatGPT and the like have indeed shifted gear and are capable of answering questions with plausible, even authoritative answers. But these are not always correct. See our report from the Rice HPC in Oil & Gas event elsewhere in this issue and also Neil McNaughton’s editorial where he catches ChatGPT out, twice. And, as reported by Bloomberg, Microsoft’s Bard ‘readily churns out conspiracy theories’. CGPT’s errors, lies and fabrications are described, as ‘hallucinations’ which is rather charitable in our opinion. All of this is understandable since these tools are trained on the massive amount of information of doubtful provenance that is available online. Vendors are now re-training these engines on their own reduced information resources to make things more ‘accurate’. But as the training corpus diminishes, you may be getting close to the point where you would rather see the original source documentation or evidence. Which would you rather, ask ChatGPT to explain why the pressure gauge is reading off the scale? Or look quickly at what’s bubbling up in the mud pit*.
* A true North Sea story by the way when a geologist came into the mud logging cabin to find the mud logger attacking the pressure gauge with a screwdriver!
© Oil IT Journal - all rights reserved.