AI has been in the headlines for quite some time now. The original hype cycle when ChatGPT was first released appeared to show a glimpse into the future where many remedial tasks would become a thing of the past, as well as a buzz of excitement for what could be possible. Alas, that seems to have been short lived. Over the course of the last 2 or 3 years, the public has effectively been told that AI will take their jobs, rendering them useless. Datacenters have become a political sticking point, not that datacenters are new, but rather how and where they are being constructed is causing serious effects on the lives of the people who live nearby. IP theft and serious copyright infringement has occurred with little recourse. “AI Slop” has become proliferating, from half-assed marketing, to overloading coworkers with half-baked documents and emails. Many people are asking, what’s the point of all this? Weren’t we promised cures for cancer and serious advancements in research?
What if I told you that believe it or not, AI can still be used for good. Personally, I like to think of myself as an optimistic AI skeptic. I recognize the potential this technology has to offer, however I have my reservations about how it’s actually being utilized today. And for the most part, align myself with the general population in terms of ethical and moral concerns. Big tech seems to live in a world detached from reality and can’t seem to figure out why the general population hates them so much. I’m not here to solve their problem on that front, but what I can do is offer some insight into where I see some hope for the future.
This last year has been quite the journey for myself. I quit my corporate job with the hope that I could create a company centered around an insights powered journaling app, just to realize I was too far out over my skis. To describe the experience as humbling, probably doesn’t do it much justice. However, I have learned a lot and tried to adapt as best as I can to figure out something people can use. Enter Small Language Models, designed to help identify people who may be at risk for self-harm or struggling with serious mental health concerns.
Let’s start first with what is a Small Language Model or an SLM. Most people are familiar with Large Language Models, the ChatGPTs, Claudes, Copilots, etc. An SLM is effectively a model that is then fine-tuned and adjusted to be really really good at specific tasks. This works because traditionally LLMs have billions of parameters due to the need to be able to do a lot of different things. An LLM might need to help draft an email for a business executive but can also write highly effective code for a software engineer. With an SLM, we are going to ask it to do one thing and to do that one thing much better than an LLM because it now only has one job. Picture yourself, like others, you are capable of doing many things, but if I asked you to learn how to play the saxophone and only spend the next 6 months solely doing that, you would probably become pretty good at playing the saxophone, whereas others would not. This is how we can think of SLMs vs LLMs.
Ok so how do we actually apply this? For context, suicide is one of the leading causes of death in the United States, as well as many other countries. For youth in particular, it is the second leading cause of death for those aged 10-34. Fortunately, there is some hope as studies have shown that earlier intervention has proven effective for reducing the likelihood of long-term major depression symptoms and self harm. If we can find ways to identify those who are more likely to be at risk, we can intervene sooner and hopefully introduce treatment that can help. By using an SLM designed to analyze, identify, categorize, and triage information about a person based on their activity, we can effectively identify if someone is at risk of self-harm or experiencing major depression related syndromes.
In terms of how this would be applied in the real world, we would start by taking the production-ready SLM and embed it directly onto devices or applications. Thanks to SLMs being smaller, in many cases we can quantize it down so the storage space (~2GB) and hardware needed can run locally on any device. An example would be children or teenagers using public school laptops or devices. If they searched “how to overdose”, the SLM can immediately identify this as a person who is at risk of self-harm and either alert a school official who is trained to help handle a situation OR the SLM can provide some soft form of intervention such as directing the user to 988 or other resources. Now if a person typed something less concerning such as “depression symptoms”, that would still be classified and scored appropriately, however immediate action may not need to be taken.
So what are the benefits of something like this beyond just the obvious of helping prevent self-harm? Why is this better than an LLM or “traditional AI”? Well for starters this is 1000x cheaper. How cheap you may ask? Well since the SLM lives and runs completely on the local device, no outbound network calls and compute resources are needed from a large datacenter. The only costs that would really accumulate is managing and hosting a supporting platform that administrators could use to help triage events.
What about security and privacy? Another win for the SLM. Once again because the SLM lives locally, we can restrict everything to the device. This means that whatever activity happens on the device can stay directly on the device, with only the SLM taking action if a certain safety threshold is breached. This ensures that the person's privacy is respected and their data is not shared with big companies or any third-party data brokers.
One wrinkle is there would obviously need to be some form of consent. This would be spelled out in the terms and conditions. In the previous example used, school administrators and parents would need to opt-in or agree to have it installed on school computers. And unlike how most AI companies treat you when it comes to explaining how they use your data, the SLM is much easier to explain.
There are probably a million other ways that the SLM concept can be applied for good - identifying specific types of cancer cells, studying plant disease mutations, personalized chatbot for specific addictions, and the list goes on. There will most likely always be a demand for LLMs and their ability to handle various tasks. However, this piece of writing was intended to highlight alternatives to what the general public knows about AI, as well as give some hope. Oh and if it wasn’t obvious, I’m working on this :)