AI Without Context Is Doomed to Fail
MIT recently released a paper published in FORTUNE showing that 95% of AI initiatives fail.
OpenAI invited nine tech reporters and journalists to an exclusive “on-the-record” dinner in San Francisco, hosted by CEO Sam Altman, COO Brad Lightcap, and Head of ChatGPT Nick Turley.
The stock market fell immediately.
Why are so many AI initiatives failing?
There are two reasons.
- Usually, the organization’s data is not the data the AI models have been trained on. Those models were trained on millions of records harvested from sites all over the internet. They do not train, and in fact, there is usually not enough data to train on the relatively small amount of data held by an individual organization. The system lacks adequate, contextually correct information, so it produces incorrect results.
- The organization’s data is not made ready for the AI system. Instead, it is just dumped in and everyone hopes for the best.
How do you get around this problem?
If you do not want to share your data with the internet at large, then you need to set up a separate instance, using algorithms and endpoints from the AI system, running on your content behind a firewall to keep your data safe.
Get the data ready by tagging the content to provide the specific meaning within that data’s context. The word “lead” can mean many things. It can be a term in management, a metal on the periodic table of elements, the feeding of one body of water to another or something you use to walk a dog. “Mercury” could be an automobile, an element on the periodic table, a planet, the messenger god in Roman Mythology or a modern bank for startup companies. It’s important to understand what a word means within your context. This disambiguation is achieved by tagging the data with terms from a controlled vocabulary.
Contact Access Innovations, Inc. to license one of their existing Knowledge Domains to auto-tag your content with a taxonomy and learn how to get the rest of the system set up. Further, to work effectively with large language models (LLMs), content cannot just be tagged at the document level. AI models process information more effectively when it is broken into chunks, or logical sections of text, such as a paragraph, image or table.
Access Innovations enables granular-level tagging, applying metadata to each unit of information. This granular approach allows the AI to retrieve, reason over and return precise answers instead of vague summaries.
Success rates don’t lie:
67% when you buy from specialized vendors and build partnerships.
33% when you try to build it all yourself.
Even if you are in the middle of an installation, it is crucial to get your content metadata in shape. Then it will work with the LLM and GPT you choose.
Marjorie Hlava, Veronica Showers and Heather Kotula
| Data Harmony is an award-winning semantic suite that leverages explainable AI. |
|
Access Innovations, the intelligence and the technology behind world-class explainable AI solutions.
