LLMs: Fake it till you make it
How does the current generation of AI work? Think of the phrase “fake it till you make it”, and then take it all the way to the extreme, that’s close enough to what’s going on to get a feel for it.
This post started with a chat with my family. I expanded on it and added a (overly?) positive take on where AI may take us. Don’t expect technical details here.
Think of a three year old kid. She’s learning how to talk by listening and imitating as best as possible. At first speech is short bursts of 2-3 words, but she gets better at faking it and eventually learns to string together multiple sentences. But she doesn’t really understand what’s going on, which results in funny stories, like the time she went to a department store, looked up at a mannequin and asked, “mom, is it dead?”.
Our brains start developing abilities for symbolic reasoning from an early age and it eventually takes over. Our learning changes from imitating to building up a mental model of the world and most of our learning revolves around understanding the world.
But what if our hypothetical kid never develops symbolic reasoning? What if she gains superhuman levels of being able to fake it? How far can she get in life?
She goes to college. She gets straight A’s in all her language and writing classes, because those only require her to regurgitate the most plausible-sounding text at the right time. For her literature final exam, she summarizes a 3,000 page book in an eloquently worded 10 paragraph essay in which she uses no single word more than twice.
History involves a little bit of memorization, but beyond that, it’s nothing more than summarizing events. It’s easy. During a study session for the final exam she formats the history of Tanzania as a series of limericks. Straight A’s.
Math was hard, but she finds that she if studies enough examples of math problems, she could fake trigonometry and calculus. It’s not perfect, but she can walk away with C’s and D’s, which is enough to graduate.
After graduation, she picks up a job as a businesswoman and becames a huge hit at the new company. She appears to have deep knowledge of a huge variety of topics. She responds in detail to every customer concern, and always speaks with the confidence of a strong leader. The company quickly promotes her into the executive ranks, where she excels.
Large Language Models (LLMs) are the current generation of AI. They work essentially like this, and they sound very impressive. I’m sure eventually we’ll see a breakthrough that gives AI symbolic reasoning, but they don’t have it now and they won’t for the foreseeable future. So how well can they do by just faking it?
“Fake it to you make it” is a common phrase in business. A lot of people think that’s one of the most effective strategies an executive can take. Some say that’s how startups in Silicon Valley succeed.
But we’re talking about very sophisticated faking. Superhuman levels of faking, beyond what you’ve previously imagined.
It can pass a trigonometry test just by writing down the most plausible-sounding answer. If you make it break down the problem into sub-problems, it dramatically improves it’s accuracy because it can readily come up with plausible-sounding answers for the sub-problems and then roll it all up into a solid plausible-sounding answer for the full problem.
It can read through a 300 page book in seconds, and answer any question you have about the book. We’ve even found ways of packing in near-infinite amounts of text with varying levels of success. It can turn dense legal documents into poetry. It can create Monet paintings out of a child’s crayon drawing.
Someone on Mastodon had a really interesting take:
I think this is a complex topic because, on one hand, we have people with valid claims that AI is stealing their hard-earned work and replicating it. But your example is why I think this is a sort of graphic version of the Gutenberg printing press all over again. I cannot tell you the number of adults with amazing ideas who cannot express them clearly with either words or pictures. The ideas get set aside because it’s so hard to get others to understand what you are trying to convey. I’m incredibly excited about an age where people can visually share ideas quickly. Can enhance storytelling. I think it’s going to change how we communicate with each other.
It’s not just visual. The level of difficulty of communicating to another person has dropped to zero in the last year. That opens up a lot of opportunities for many people.
It’s extremely difficult to predict the future, so anyone trying to tell you the outcome of AI is definitely trying to either sell you a political narrative or exploit a new business opportunity, but I can tell you this:
It takes a lot less skill to make decent things nowadays.
My three year old will use her overactive imagination to tell me about creatures and scenes that creative or even absurd, and together we’ll use ChatGPT to create pictures and stories that bring the idea to life. My older kid doesn’t need me, she can use voice-to-text and text-to-speech and do it all herself. It makes me wonder if reading & writing will have the same fate as cursive handwriting.
On this blog I’ve started using AI-generated art to augment the text. I think it looks better this way, but it’s not something I care enough about to pay money for. Before this I simply had walls of text with no images.
A Workforce Without Faking
If I try to predict the future (carefully), I tend to think that work will require a lot less faking it, because all that is done much better by an AI. I admittedly am biased toward being overly chill, but here’s what such a workforce could be like, take it with a grain of salt:
- Authenticity: No one learns the plastic exterior, because AI does it better anyway
- Collaboration: When people lack communication skills or speak different languages, AI can step in and help them communicate their true intent.
- Reduced Impostor Syndrome: When AI does virtue signalling better than we can, all that’s left is to be authentic about our actual struggles, and help each other through.
Having worked on AI for a long time, I can tell you that “faking it” can be taken a very long way and probably shouldn’t be underestimated. But if “faking it” is also no longer a viable strategy for excelling in this world, maybe all that’s left is to discover our true selves and be authentic.
If that’s too rosy for you, then read Bruce Schneier’s take. It’s very grounded, unlike a lot that’s written on the topic.