Design by Artis Briedis
Not often, but there are times when I seriously regret my educational choices made decades ago. Listening to Anima Anandkumar, a professor from CalTech, was one of those moments.
My pitiful high school level maths and physics skills (or what’s left of them after 40+ years since) were inadequate to understand any technicalities. That left me focusing on what the models can do – to the extent that I could comprehend. And I can’t vouch that I have understood everything correctly.
With the above caveats in mind, these are my three takeaways after watching a lecture delivered by Anima at Stanford University on YouTube:
Lesson #1
The main idea behind training language models is that they can predict what the next word is given all the context. The learning of how to predict this at scale is called pre-training. Then a little bit of human feedback is given with reinforcement learning. This is called an alignment process which makes models able to understand prompts and instructions.
Language has one major shortcoming – it lacks embodiment. You can learn how different words relate to each other, and what the underlying meaning is, but in the physical world, words need to translate into actions.
Models have been used to progressively learn new skills – from simple to complex – while continuing to use the skills learned before also in a virtual world, like Minecraft. The agents have been taught not only to execute the actions but also to decide what to do next. This is different from standard reinforcement learning – the agents have to figure out how to continue gathering new skills and solving more complex tasks.
Lesson #2
Language models are not limited just to the use of natural languages. They can be trained in using genomic language as well. This allows the model to predict and learn what genes do, and to do it at scale.
The first model has been trained on 110 million genomic sequences – from flu to coronavirus to ebola. The advantage of generative AI models is that they can look at how all the different viruses and bacteria can mutate over time, not just one as before.
When the model got fine-tuned specifically to focus on coronavirus, the model predicted Delta and Omicron variants before they appeared. The model’s ability to learn the evolutionary dynamics of viruses might allow the development of vaccines in advance before a new pandemic starts.
Lesson #3
It may take as long as the age of the Universe to perform a quantum simulation of a molecule with 100 atoms. Even when the equations that describe the systems are well known the challenge is to simulate at scale.
A neural model armed with 45 years of data and using more than 100 variables for weather prediction globally works 45,000 times faster than numerical models. And there is no need for a supercomputer.
Because of the high cost, the current numeric models used for weather forecasting analyse only 50 variations or ensembles whereas neural models can do thousands or even millions of samples. Many climate predictions we are using today are based just on one simple model. It is not enough to do just one run and say that this is what is going to happen.
And, as a final note, weather is not the only use case for neural agents. The carbon dioxide underground storage simulation, nuclear fission and material deformation modelling are areas that are already being explored. But the ultimate goal is to have a model that understands ALL the science, not just a specific field.
For those willing to try and pick top lessons of their own here is a link to the video.
* from anything that you are reading, watching or hearing you can realistically expect to remember only a limited number of things. My solution is to pick just 3 items or ideas from any material. This number is non-negotiable. Even the most extraordinary experience gets compressed into 3 things to remember. This approach has worked well for me.
This note was first published on medium.com on 13 November 2024.
Aivars Jurcans has more than 20 years of corporate finance and investment banking experience. His services are currently available through Murinus Advisers. More of his writings can be found on his page Corporate Financier’s Notes.
