Musings on AI: Context and Intelligence
Context is how we understand the world. If someone were to ask you out of the blue — “write me a 5 line email to send to a client” — you would do your best, but you probably wouldn’t be able to produce anything useable without knowing who the client is, what service you are providing for them, your role, etc. The same goes for how you interact with your LLMs. A lack of context can have huge consequences, and in the case of AI, one of the biggest is the spread of misinformation.
Mattel’s Chief Information Security Officer, Tom Le, hit the nail on the head with this when he said in the WSJ this week “The devil is really in how you ask the question.” Not giving any sort of background information can lead to potentially dangerous conclusions.
Humans are not only necessary for the review of LLM output, but we are also necessary prompt machines specializing in providing context. This got me thinking (yet again), what is intelligence really?
Intelligence: the ability to acquire and apply knowledge; synonyms: intellect, mind, brain, judgement, reason, understanding
By this definition, sure, ChatGPT and other LLMs are intelligent. They are able to acquire massive amounts of information that they apply to answer certain questions and perform certain tasks. The human element, however, lies in the synonyms. ChatGPT does not and cannot apply any judgement, reasoning, or understanding. This is where we keep our jobs people! 🙌🏻 We are able to judge the output of an LLM, we are able to reason with each other about what it means, and we are able to understand what it’s telling us in the context of our individual lives — these are things that these models will never be able to do… or will they…
This week The Neuron reported on a series of leaks on Reddit that OpenAI is set to launch “profiles.” The rumor is that you enter information about your job, your interests, etc. that’s saved on your profile, essentially feeding your ChatGPT instance context for your life which it can use to provide more useful and accurate answers to your queries. Right now you can share links to your chats with other users so they can continue to query ChatGPT given the information and the context you’ve already provided the model — cutting down on replicating that leg work. With these developments OpenAI is really chipping away at the context blocker.
The more specifics you give a model on who you are and the problems you are solving, the more context it has to provide more accurate and useful outputs. The problem? Now the model has all of that data. How much information do you really want to be giving OpenAI about yourself, your job, your company — when all that information can then be used to continue to train and improve its models? While proceeding with caution is probably advised, it also limits the speed with which you can successfully implement this powerful technology at scale.
Context is everything — which, no surprise, makes it both the answer and the question when it comes to how we interact with AI.
