Founder & CEO of Vital, AI-powered patient experience. Formerly Founder & CEO of Mint.com. 10 patents in algorithms.
getty
"Not All Algorithms Are AI" is a planned three-part deep-dive into the evolution of algorithms, what brought us to generative AI and how to understand what this technology will do for your business.
In Part 1, I explored how classic algorithms can be proven mathematically correct but are inflexible, while machine learning's use of probability allows big data inputs but at the cost of never again being 100% right.
Deep Learning And The Algorithm For "Cats"
The world changed more than you probably realize in 2012. For computer scientists, this was the rise of "deep" learning through convolutional neural networks and the first practical solution to the ImageNet challenge of recognizing objects in photos. To me, it feels similar to the momentous date in 1865 when James Clerk Maxwell unified the theories of electricity and magnetism that underpin the modern world.
Just 11 years ago, researchers fed millions of photos of cats into a multilayer neural network. In the end, an artificial "brain" could reliably recognize cats, even if it had never seen that cat in those particular surroundings before. This is easy for a child but very hard for an algorithm.
Soon, the system could recognize objects in general: cats, dogs, roads, bicycles, bridges, buildings, faces. Facebook's auto-tagging of friends became possible. Your Apple Photos app organized groups of pictures by faces or content (sunsets, beaches, etc.). The chaos of too many photos and too much data became, in part, tamed.
Remarkably, the multilayer neural network that the researchers invented worked quite similarly to the six layers present in the mammalian neocortex (the gray, wrinkly part of your brain). The first layer organized raw pixels (think sensory data) into edges, the next into shapes, then a composition of shapes, then objects. It worked like a human brain: Once it saw enough data, it could identify objects and similarities without having explicitly been directed.
At my company, we use a similar "unsupervised learning" concept. While there are databases of medications and symptoms, these relationships can be inferred rather than tediously maintained.
Take "pain" medication, for example. A deep learning algorithm only needs a few thousand medical notes: "The patient was given ibuprofen for pain," "Take aspirin, as needed, for pain" or "You can use acetaminophen for a painful headache." In each case, the drugs ibuprofen, aspirin and acetaminophen are used in similar contexts: to treat pain.
Think of it as an automatic synonym finder. In medicine, that means the term "cerebral infarction" gets grouped with its layman's equivalent: "stroke." Want all searches for "stroke" to return documents containing "cerebral infarction" too? Now you can. You might not have noticed, but this is exactly how Google search has evolved in the past few years.
ChatGPT And Large Language Models: Now And Beyond
Large language models (LLMs) and generative AI are deep learning on steroids. GPT4 is rumored to have more than 100 layers instead of the 6 layers that could recognize cats. The models are called "large" because they’ve been trained on a large portion of the internet: millions of books, images, scientific articles, Q&A forums, open source code, all of Wikipedia and more. These models are so big, in fact, that to train one from scratch can cost over $10 million in server and compute time.
LLMs generate output word-by-word using prompts, input (including prior messages), words already output and a bit of randomness thrown in for good measure. The result is usually well-formed sentences but with results that are hard to replicate: The randomness produces higher quality, more "human" output, but the same input may produce different outputs if run again.
In deep learning, we infer "synonyms" using a context that is typically the 10 words before and 10 words after the target. LLMs have a much longer context, sometimes tens of thousands of words. This lets them "remember" the question you asked a few interactions ago. This short-term memory, plus the corpus of the entire internet, makes them appear surprisingly smart at question-answering.
At my company, we use LLMs in our doctor-to-patient translation. This system turns long medical notes full of jargon into shorter, more comprehensible instructions for patients. In a medical note for a cardiac incident, the LLM looking to generate a next phrase may consider "myocardial infarction," "coronary infarction," or "heart attack"—these are all cardiac incidents.
But because the prompt also asked for a "high school reading level," the more probable phrase becomes "heart attack." The prompt skews the likelihood, like loaded dice, toward the desired outcome. That’s why we can ask ChatGPT to write a poem in the style of Shakespeare or construct lyrics that match the cadence of Eminem.
Instead of the fixed statistics of deep learning, the probability in generative AI changes as the context changes. This is an even deeper form of human mimicry. Out of context, if I said, "Let’s go to the bank," you might assume a visit to the financial institution. However, if we were in a canoe on a river, that exact same sentence is likely a directive to paddle to shore.
Context matters. Our brains use location, time of day, emotional state and even a gauge of our audience's intelligence to alter output. Generative AI is beginning to do the same, which is one of the primary reasons it appears to have uncanny human abilities.
Worried about your job? Concerned that AI in the real world will end up destroying humanity like AI in movies (think Hal 9000, Skynet or Ex Machina)? No need to worry yet. Today’s AI is still just fancy math. The rise of self-aware machines requires a solution to problems few have even begun to work on or understand. Artificial general intelligence requires emotion, physical or energetic consequences, a concept of time and a self-model of attention (focus).
Each will come slowly, over the coming decades. And each has the ability to define new billion-dollar businesses, adding value to the world.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?
