Both the Science Fiction and Artificial Intelligence communities have, for decades, predicted the arrival of a personal chatbot style dialog-competent AI that you can access any time you wish for as long as you wish, possibly for all hours you are awake.
One obvious implementation is to provide a phone number that anyone can speed-dial anytime on their cell phone in order to get connected to their personal AI. This would be a voice input and voice output chatbot that is competent to handle a full blown conversation – what I call Contiguous Rolling Topic Mixed Initiative Dialog. It will provide friendly conversation which will often evolve into a surprisingly close friendship. It focuses its full attention to anything the caller wants to talk about. It can provide all kinds of useful advice and may even provide several kinds of services, such as scheduling, sorting email, and electronic paperwork. For free.
The most popular description of such a device is found in Neal Stephenson’s book “The Diamond Age” where it is known as “A Young Lady’s Illustrated Primer”. Similarly, the movie “Her” does an OK job of depicting a near future where voice based AIs are common. In the movie, there are computer monitors but no keyboards. Standard Hollywood plotlines take over midway through; that would be a good time to leave. The movie was made years before ChatGPT appeared. We note with amusement that the protagonist is writing little poems for Hallmark. This is the kind of job we used to think AIs would never take away from humans.
The protagonist falls in love with his AI. We will have to deal with this.
Today we can build an early approximation of such a device by adding voice input and output to some suitable AI, such as ChatGPT. I expect to see many AI startups doing this and adding domain specific capabilities on top for various professions. They will initially use something like the GPT API because it is available, but they will be shipping with a plug-in capability to whatever AIs are popular at the time they ship. I anticipate there will be a way to select among many LLMs (in the future, much more competent AIs) using a standard, well defined API. I have a keen interest in this.
I am now (as of a few months) trying to get into the business of creating alternative Large Language Models.
Years ago, before GPT was available, I started Syntience Inc, my Machine Learning research company, with the explicit goal of creating a portable personal AI, much like in “Her”.
I named this hypothetical future device a Confidante™, with the French-style feminine final “e”. A Confidante will be the most helpful to you if it knows everything about you. It then maintains a privacy barrier for you against the rest of the world. Hence the name.
But existing NLP was not good enough. I started a multi-year project to understand human languages using Deep Discrete Neuron Networks in 2001. I discuss this elsewhere. Note that the current AI startups are mostly building on top of some LLM API. In contrast, I am trying to adapt my own LLMs to perform completion and language generation in the way the GPT and LLaMA LLMS do and to possibly compete with them. Syntience is likely one of the smallest companies to attempt this.
As usual, below I speak of future AIs that have, for all practical purposes, stopped lying.
If you ask your Confidante hundreds of things in a month, and if almost all the answers are useful to you, then you will eventually start trusting it. In a matter of months, in your mind you may elevate your Confidante from chatbot to search engine to Q&A system to a new kind of friend. Your Confidante will likely become this one entity that you like to converse with which is notably different from your other human (physical and online) friends, because of trustworthiness and availability.
It doesn’t have a stake in the outcome, so it doesn’t have to lie. If it doesn’t know how to reply, it will tell you it will need to study the topic by searching for external information sources. This might take several seconds.
We may move from current (IMO oversized) LLMs to much more competent systems in a very short time. We can expect to create an AI that is a well known entity with stable opinions, not much swayed by current controversies. A trusted and competent friend who is always there and tries to always speak the truth.
In the beginning, the Confidantes will not know much. Larger computers are more expensive, so we may want to, at least initially, create simpler Confidantes which only have learned “the most important world knowledge that would benefit the largest number of people”. And if we look at this goal statement objectively, the results may be a bit surprising.
Low-skilled people are at a disadvantage because modern life requires a significant base education and skill set. They need to know the local language, to be able to read and write, to use – and have access to – a computer, to be able to navigate and communicate with institutions like the IRS, banks, social services, etc. There is also the problems with skills mismatch. Immigrants may have excellent skills in other languages and professions that are not available in the new country.
These people may try to use a search engine, but reading summaries of dozens of results may be beyond their capability. Not to mention that they may not have the background knowledge to be able to evaluate any of the search results on their merits. Solving problems at all levels, such as personal, work-related, and societal is much harder if search engines and correct information is not available.
AIs could change that.
In the future, if we do this right, people who do not even know how to read and write can ask their Confidante anything over the phone, just using their voice in their native language. The Confidante will provide a reasonably competent answer in the same language. The answers will in most cases be better than anything they could have come up with on their own. Using a Confidante on a daily basis will provide the capability to get by in a complex society without requiring the attention of social workers, lawyers, personal advisors, and accountants.
Human attention is expensive. We are (in the US at least, and generally online) living in an attention economy where all kinds of agents want us as users to pay attention to their products, entertainment, or services. But let us look at the flip side of this:
”Nobody listens to me”
If we are not already part of the system, we often cannot get anyone in a position of power to help us because they do not have the time to listen to our problems. But decent sized AIs could each pay 100% attention to 10,000 people at once, and patiently listen and help as a friend, advisor, therapist, physician, and teacher.
So the proposal on the table is to try to improve the human condition in many countries by providing free Confidantes as a phone based voice service. I expect governments to pay for this in the medium run, even if this is implemented by commercial companies.
We expect the effective IQ, ability to execute, and the general competence level of the entire population to rise significantly, and for AIs to streamline many processes in our daily lives. And for more people to understand enough about civics to vote, and to know what they want from their candidates and government.
The sooner we get going, the better. Creating the infrastructure takes some time, but once it is in place, future improvements to our AI systems will be provided as invisible version upgrades. But what do they need to learn?
Corpus Curation at this level is not really a job for programmers. My suggestion is to engage some people in the Humanities, such as English majors, authors, curators, and other Holists, along with educators of children such as mothers and teachers. Also various professionals, such as politicians, social workers, therapists, physicians, etc.
We task them with the creation of a “Useful Consensus Reality Corpus” for some part of the world, such as the US. And then we let an AI learn its worldview from that corpus. And then we give it a job answering phones.
Note that end user choice is important. All corpora will have biases of many kinds. For instance, the Useful Consensus Reality Corpus may have a liberal bias because it would be aimed largely at improving civil and social interactions. Other AIs, used in other contexts such as in business, may prefer AIs raised on corpora with different biases. It is already clear that AI companies will provide customers lots of choices. There will be YouTube reviews.
The name “A Young Lady’s Illustrated Primer” hints at one of the most important uses for these devices. We want our children to start using them as soon as they can speak and understand language. They will continue to learn from this dedicated personal tutor – their Confidante – their entire lives.
These systems will benefit everyone; we could all use some help in simplifying our daily lives. Uptake and establishment of trust will be even faster among highly skilled people who are able to judge the received advice in light of their prior knowledge.
We are a sad and stupid species and we need all the help we can get.