In September OpenAI released previews of their new model, ChatGPT-o1 which they claim exceeded human levels of accuracy at PhD questions for physics, biology and chemistry.

This new model introduces an idea that called "chain of thought" reasoning. The idea is that instead of having ChatGPT respond directly to the prompt, it could instead consider many different answers to the question. These answers could be used to prompt the model further, eventually providing a set of chains of thought. A seperate model could then evaluate each chain with respect to the original user prompt, returning what it considered the best answer.

This idea is very reminiscint of a much older AI known as AlphaGo. AlphaGo was a system for solving Go. Go is a chinese board game that is thousands of times more complex than Chess. Chess is solvable because there aren't that many possible moves, and a chess solver can rapidly search through all the possible plays from a given position and then choose the one that has the most number of win endings for the AI. Go is so complex that this search is not feasible.

To solve Go, Alpha Go employed an algorithm called Monte Carlo Tree Search (MCTS). This involved two seperate AI's. One AI would suggest possible moves that the AI could take, while another would look at the current state of the board and give it a goodness score. This score was not an exhaustive search, just that AI's best guess at how likely advantagous the board position was.

To decide upon a move, Alpha Go would imagine taking a move, get a goodness score for the move, and then imagine taking another move from that imagined board, or switch to a different starting move. It was given a computation time window in order to perform this, and would then present its best move found so far when time was up.

I like the Alpha Go solution, but it's an interesting that it combines a modern AI with a much simpler idea. The idea of search. Search algorithms are as old as mathematics, and they power lots of things we take for granted in our modern era. When you request a destination in your phones map app, it will perform a search over the road network, considering different options and then presenting you with a path. When a mining engineer decides which block to mine and which order, they are performing a search over the different options and determining the best one.

It might seem pedantic to make the distinction between an AI and a search algorithm, as we expect AI's to be doing this all the time. However, if you have ever interacted with ChatGPT, it can often seem less like an algorithm that searches through its database, and more like someone who gives you their best guess in that moment.

I'm sure you are thinking that given enough data, surely a ChatGPT style system can solve anything without search. And this is true, but the data needed is much, much more than you might expect. Recently the paper "Grandmaster level chess without search" by A. Ruoss was able to get superhuman performance using an AI to play the Chess. However, it required a 15 billion examples and used an existing chess solver to create them. It's unlikely that there exists enough data to train a ChatGPT model to a superhuman level of intelligience in this way.

This is why OpenAI's o1 algorithm could be a game changer. By combining the tried and true method of search, that powers much of modern life, with the arcane capabilities of Large Language Models, we might just see the birth of extremely useful AI.

But don't fire all your workers just yet. There is still a long way to go before we are going to get chat bots that you can trust with your business. And chat bots that your customers will be happy talking to.

ChatGPT's latest model uses an old idea that powers your phones mapping app

In September OpenAI released previews of their new model, ChatGPT-o1 which they claim exceeded human

The all-in-one writing platform.

Write, publish everywhere, see what works, and become a better writer - all in one place.