As public benchmarks saturate and models get better at optimizing for the tests themselves, Rayan makes the case for independent, continuously evolving evaluations.
For the first time in human history, we have a mathematical model of all of human knowledge. It's not perfect, but that's essentially what all of the LLMs are.
today's concentration may be a feature of the current technological paradigm rather than a permanent feature of AI. Transformers reward enormous amounts of data and compute, but future breakthroughs could make smaller, more specialized models far more capable.
Today's models can grind through calculations, combine known techniques, and search enormous spaces, but still struggle with intuition, theory building, identifying the right questions, and developing the kind of big-picture understanding that drives much of mathematical progress.
It's funny that AI systems are all still pretty bad at elementary school arithmetic, but getting increasingly good at very high-end abstract math. That raises some big questions for the field of advanced math.
What I believe is that driverless cars today and lots of automated systems look very much like NLP systems in the 2010s. They are these very complex very handtuned systems. They're intelligent but they're intelligent in sort of isolated ways and that those systems hold back what these technologies can do.
The thing that's special about transformers is these attention layers where every token is looking at every other token. That operation, when you implement it on a GPU, is really inefficient because you have to load the weights of the model for every single token that you're generating, even though the weights don't change between token generations. And that's really the core insight behind why ASICs make sense for transformers — you can etch the weights directly into the chip.
The thing that's really interesting about transformers is that the model architecture is essentially fixed. The operations that transformers do are the same whether you're running GPT-4 or any other large language model — it's always matrix multiplications and attention. And if you build a chip that only does those operations, you can make it much faster and much cheaper than a general-purpose GPU.
The thing that's so special about transformers is they can be trained end to end with gradient descent. You don't have to come up with the algorithm yourself. You just show the network examples of inputs and outputs and it figures out the algorithm. And because the algorithm is learned rather than programmed, it can be much more sophisticated than anything a human could come up with.
The thing that's really interesting about transformers is that the weights are basically fixed after training. You load the weights once onto the chip, and then you just do inference over and over again. So if you build a chip that is specifically designed to run one model or one architecture, you can get massive efficiency gains because you're not paying for the flexibility of being able to run arbitrary code.
The thing that I find most stunning about this is that transformers were invented in 2017, and the reason that they've taken over everything is that they're extremely good at using parallel compute. And so the question is, are transformers the last great neural network architecture? And our bet is yes, or at least that they will be dominant for the foreseeable future, which means that if you build a chip that is perfect for transformers, you will have a chip that is perfect for AI.
The thing that's really interesting about transformers is that the attention mechanism, which is the key innovation, is actually a fixed algorithm. It's not like the weights change the algorithm; the weights just parameterize the algorithm. And so if you know that the algorithm is fixed, you can build hardware that's specifically designed to run that algorithm as fast as possible.
The thing that's special about transformers is the attention mechanism, and the attention mechanism has this property that every single token has to interact with every other token. The amount of computation required scales with the square of the sequence length. So as you try to process longer and longer contexts, it gets quadratically more expensive.
The thing that's really interesting about transformers is that actually, when you look at what all these models are doing, they're all basically running the same algorithm, which is the transformer architecture. And so if you build a chip that's specifically designed to run transformers as fast as possible, you can get enormous speedups compared to a GPU, which is a general-purpose chip that has to be able to run any algorithm.
The thing that I find most compelling about the transformer is that, unlike prior neural network architectures, the transformer's compute requirements are almost entirely matrix multiplications. And matrix multiplication is a very structured computation that you can build hardware specifically to do very efficiently.
Odyssey AI is not trying to build yet another large language model. Instead, Jeff Hawk, the company's co founder, is focused on world models AI systems that learn from site and sound to understand and simulate how the real world works over time, predicting what happens next and letting us interact with those simulated futures.
It was basically sort of thought to be career suicide. Like of course we know AI doesn't work. You know, we tried it in the 90s, places like MIT and it was a dead end. You know, and that was the prevailing view. But we just felt the small band of us felt that that actually with the right ideas and using learning systems, reinforcement learning and betting on neural networks that a lot of fast progress could be made.
world models AI systems that learn from site and sound to understand and simulate how the real world works over time, predicting what happens next and letting us interact with those simulated futures.
getting there required more than better models. It meant building a complete system for training, evaluating, and deploying a driver in the real world.
