Algorithm War Coming! (Part 2)
Last year, I wrote an issue that companies will be competing against each other to find out who has the ‘best’ algorithms. The issue was written in June last year. From then till now, we have seen that the tech firms are now competing on the Generative AI front, especially multi-modal, with OpenAI showing up with SORA (Side Note: I was not impressed but am sure it will get better) and Google coming up with Gemini. Even on the Image GenAI front, there are now many competing models, Bing Image Generator, Stable Diffusion, Midjourney, TensorArt just to name a few.
Per what I have discussed previously, since then the competition has heated up in having the best algorithms has developed. In this issue, I wanted to discuss a bit more on the competition between these algorithms. In fact, I wanted to discuss more on the word “best”.
We all know that the word “best” can be measured in many dimensions. It could be the fastest, or most accurate, etc. I kind of felt the word “best algorithm” by itself is a misnomer rather. Researchers are coming up with different measurements on how good GenAI model is. Just on LLMs alone, we can measure in terms of its reasoning skills, depth of knowledge, lexical semantics, etc. It will be very difficult to determine which is the “best algorithm” unless it trumps every other model in ALL dimensions. Such a scenario I felt is unlikely. It can only happen when the data it is trained on, is the largest and encompasses ALL the training data of ALL other models, followed by there is a single model training method that can trump all other training methods, which those in machine learning will know is not necessarily true.
What does this mean?
What this means is that as long as sufficient amounts of data is collected, a company can train its own GenAI model to compete in the space. It will have to find out where its GenAI model strengths are, and hopefully the model’s strengths produce a good product-market fit i.e. the model can serve a large enough market to generate reasonable profits for the trainers.
There will be a proliferation of GenAI models and the models that last are the one that can bring in the highest sustainable profit for its trainers, till it is replaced by another model that is stronger in that particular dimension.
The only constraint or limiting factor will be computing power. Available computing power will determine how many models can be produced, and data accessibility and the chosen model architecture will determine in which dimension the model will be better at than others.
The algorithm war will be never-ending. GenAI models will continue to be trained, with each GenAI model being successful in certain dimensions. Model that are currently doing better will only be replaced by a later model that is doing better on all coinciding dimensions.
What are your thoughts on this?
I hope this issue has prompted more questions and thoughts. To support my work, consider sharing, liking, and subscribing my newsletter please. :)
Perhaps buy me a “book” or two, to let me know I’ve made a positive. :)
Recommended Content
The longer post, “Algorithm War is Coming” is found on my blog site. <Blog>
Thanks for sharing this Koo! Just like humans , there is no 'best' human. "Best at what?" should be the question that consumers of Gen AI ask. Longest context window – what does this mean? It's like saying a human has the best memory ... so what? "Best at ..." must be in the context of application to allow for the right consumer choice-making.