Conversations about Large Language Models (LLMs) were once confined to the domain of speech techies, but now it’s gone mainstream.
And everybody blogs or vlogs about it… Let me throw out a few salient points that come to my mind and the implications:
- Current models are based on text not voice. We type in search queries to You.com, we type in instructions to ChatGPT, and we type in descriptions for image generation into Dall-E (i.e. the image generated for this blog post).
- Future models will be based on voice and can converge with voice assistants, making them smarter and more capable. This is Sensory’s domain as we can perform the speech-to-text, text-to-speech, wake words and even voice ID to distinguish the user speaking.
- Answers aren’t always right. The data these models are trained on is the internet. The internet is full of false information. At least when a “normal” text search is done, you know the source, which can help identify validity. Even on Wikipedia with multiple sources, evidence must be cited and what’s written must pass the scrutiny of the crowd. I’ve seen a few examples of failed ChatGPT logic as well as plain content errors. I asked ChatGPT “what answers do you get wrong?” It responded, “As an artificial intelligence, I do not make mistakes.” Then it went on to explain how it can make mistakes if the input it was trained on was wrong.
- They will get better fast. LLM’s will become excellent in certain domains that they focus on. They also will get corrective measures to improve accuracy and sources (or combined sources). Some will give us alternative answers we can select from, and others will provide assessments of their accuracy. We are going to get a lot of models to choose from!
- Search will get a shakeup. That’s why Bing funded a billion dollars into OpenAI. Google has been too dominant. The new generative AI search will eventually combine the best of traditional search with generative AI, giving search the ability to improve and fine tune results through dialog. With current search, if I don’t get what I want I rephrase and try again. I can’t ask to refine the previous search by using natural language.
