(Bloomberg) — Artificial intelligence (AI) advances in a way that is difficult for the human mind to comprehend. For a long time nothing happens, and then suddenly something happens. The current revolution of large language models (LLMs) like ChatGPT was the result of the advent of “transformative neural networks” around 2017.
What will the next half decade hold for us? Can we judge their quality based on our current impressions of these tools, or will they surprise us with their development? As someone who has spent many hours playing with these models, I think a lot of people are in for a surprise. LLMs will have significant implications for our business decisions, our portfolios, our regulatory structures and the simple question of how much we as individuals should invest in learning how to use them.
Just to be clear, I’m not an AI tabloid. I don’t think it’s going to cause mass unemployment, let alone the “Skynet goes live” scenario and the ensuing destruction of the world. I do believe that it will represent a lasting competitive and learning advantage for the people and institutions capable of taking advantage of it.
I have a story for you, about chess and a neural network project called AlphaZero at DeepMind. AlphaZero was created at the end of 2017. Almost immediately, he began training by playing hundreds of millions of chess games against himself. Within about four hours, it was the best chess-playing entity ever created. The lesson of this story: Under the right conditions, AI can improve very, very quickly.
LLMs cannot match that pace, as they are faced with more open and complex systems, and also require ongoing business investment. Even so, recent advances have been impressive.
I wasn’t thrilled with GPT-2, a 2019 LLM. I was intrigued by GPT-3 (2020) and very impressed with ChatGPT, which is sometimes referred to as GPT-3.5 and was released late last year. GPT-4 is on the way, possibly in the first half of this year. ANDIn just a few short years, these models have gone from being curiosities to becoming an integral part of the work routines of many people I know.. This semester I will teach my students to write an article using the LLMs.
ChatGPT, the model released late last year, received a D on an undergraduate job economics exam given by my colleague Bryan Caplan. Anthropic, a new LLM available in beta and expected to go on sale this year, passed our graduate law and economics exam with clear, well-written answers (in case you’re wondering, blind scoring was used). ). It is true that the current results of LLMs are not always impressive. But consider these examples and the one from AlphaZero.
I don’t have a prediction on the rate of improvement, but most normal economics analogies don’t apply. Cars get modestly better each year, as do most things I buy or use. LLMs, on the other hand, can take leaps.
Still, you may be wondering, “What can LLMs do for me?” I have two immediate responses.
First, they can write software code. They make a lot of mistakes, but it’s often easier to edit and fix them than to write the code from scratch. They are also often more useful for writing the boring parts of code, freeing up talented human programmers for experimentation and innovation.
Second, they can be guardians. These LLMs already exist, and will soon be much improved. They can give very interesting answers to questions about almost anything in the human or natural world. They are not always reliable, but they are often useful for new ideas and inspiration, not fact checking. I hope they will be integrated into the check and search services soon. Meanwhile, they can improve their writing and organize their notes.
I started to divide the people I know into three camps: those who don’t know about LLMs yet; those who complain about their current LLMs; and those who sense the amazing future that awaits us. The intriguing thing about LLMs is that they do not follow uniform and continuous rules of development. Rather they are like a larva about to become a butterfly.
It is human, if I may use that word, to be anxious about this future. But we must also be prepared for it.
Original Note: AI Is Improving Faster Than Most Humans Realize: Tyler Cowen
More stories like this are available on bloomberg.com
©2023 Bloomberg L.P.