Single Blog View

Right now we have two Goliaths in ChatGPT (Microsoft backed) and the Google AI suite (Bard, Gemini etc.) I have to give credit to Google as they're really hit back hard against ChatGPT with their new release and the model looks great. I'm still super surprised there's been nothing large released by Amazon or Apple as they're definitely late to the table. For now, I see there being two Goliaths in Microsoft and Google, sitting atop their respective hills in their mega mansions. Life is good - these guys have access to a significant amount of compute power and storage. They can spend the money to keep training better and better neural networks and hone in on how to improve the response speed.

Now, in the not so nice part of town, in the heart of chav central, we see cars on bricks held together with twine. Meet the Dave's. I mentioned this before about there potentially being two main branches of AI being commercial versus open source similar to Windows versus Linux and yes. Here we have the hopes and dreams of the freedom fighters. There are many open source generative AI models, for example: WizardLM, Vicuna, StableX and Pygmalion. These models are amazing and are definitely worth looking at, but they can't hold a candle to the Goliaths when it comes to all-roundedness; whether that be coding, translation, calculations etc. But they do seem to do certain things well, for example, Pygmalion's ability to role play. The other thing is that these models aren't censored which has a number of advantages.

So, based on this, Goliath(s) crushes David(s) right? Well yes, sort of, maybe .... hmmm, not necessarily. You see, there's always that sling isn't there. Yes, in a straight fight, Goliath does indeed crush David. ChatGPT and Bard outperform their open source rivals, but it did take a lot more time, effort, resources and money to train those models than the open source ones. This is where David pulls out his sling. If you want an all everything model, sure, you have to look at the Goliaths. But if you don't need all that, why go to the extreme of the extra overhead? Right now, the race is on to find use cases for AI models to be used effectively in the workplace. It's great to have an assistant in the form of something like ChatGPT, but what if there were a model trained precisely on your company's data? That might be more powerful in certain instances. This is where open source may provide a more cost efficient solution. What the open source models have proven is that you can create a really powerful AI based on 7bn and 13bn parameter models and that these can be trained pretty quickly and cheaply. Do we need a 1tn parameter model to answer the question of what can I used instead of milk for my bakery or how much seed do I need to plant in a 20 acre field? If the answer to those questions vary based on the type of bakery and food your making or the location of your farm and the type of soil and amount of rainfall you have, wouldn't you just need a model trained on your specific scenario?

So is Goliath scared? I mean, it's not like they'd lobby for legislation to force AI products to be licensed or anything like that. That would surely be a desperate measure to create a barrier to entry.