Don't Get Blinded by the Hype: Why Your Old Language Model Buddy Might Still Be the Best
New models are released every week: but regardless of the claimed SOTA few old buddies are still the best ones
Newsflash: the AI world is moving faster than a hummingbird on espresso. Every week, it seems like a new, bigger, supposedly "better" language model hits the scene, demanding more memory than a dragon hoards gold. But before you jump ship and leave your trusty “outdated” models behind, hold on – sometimes, the latest and greatest isn't all it's cracked up to be.
The Latest News is not always the better one...
New models are released every week: but regardless of the claimed SOTA (State Of The Art) few old buddies are still the best ones.
Sure, these new behemoths boast impressive performance on benchmarks and can generate Shakespearean sonnets about astrophysics while simultaneously composing a jazz symphony. But let's be real:
They're resource hogs: Running these leviathans requires hardware that would make NASA jealous. Unless you have a server farm in your basement, chances are your average laptop will turn into a whimpering puddle of silicon trying to handle them.
They're often over-engineered: For many everyday tasks, like writing emails or summarizing documents, these colossal models are like using a nuclear bomb to swat a fly. They're overkill, and you'll likely get the same, if not better, results with a smaller, more nimble model.
They can be opaque: Understanding how these complex behemoths work is like deciphering ancient alien hieroglyphics. For simple tasks, you need transparency, not a black box of algorithms spitting out results.
So, where do our beloved Flan-T5, Orca, Llama, Open Llama and Microsoft Phi come in? These "old buddies" are like the reliable friends you can always count on. They:
Are lightweight and efficient: They run smoothly on everyday hardware, meaning you can take your AI-powered creativity anywhere. No need for a dedicated data center!
Get the job done: For many practical tasks, they deliver outstanding results without the unnecessary computational extravagance. Remember, finesse can sometimes beat brute force.
Are more transparent: Their architecture is easier to understand, making them ideal for learning and experimentation. You can actually see how they tick, not just blindly trust the magic sauce.
You can read more about Flan-T5 and LaMini models here:
Transformers based models based on the T5 architecture are MASTERS in Summarization, QnA and information retrieval. You definitively must give them a try
Also if you want to enhance your success rate in Retrieval Augmented Generation you need to leverage small Models.
And let's not forget the exciting new kids on the block – models like those with emergent architectures that break free from the Transformer mold. These pioneers are showing us that big isn't necessarily beautiful, and there are innovative ways to achieve impressive results without becoming resource-hungry monsters.
The bottom line? Don't let the hype about the latest and greatest cloud you over. Remember, the best language model for you isn't always the one with the biggest buzz. Consider your needs, your hardware, and your budget. Sometimes, the best friend you already have is the one who will truly get the job done, without breaking the bank or your computer.
So, give your trusty Flan-T5 or Microsoft Phi a high five. They're not old news – they're classic tools for a reason. And keep your eyes peeled for those up-and-coming stars who are proving that bigger isn't always better. The future of AI is looking smaller, friendlier, and just as amazing!
This is only the start!
Hope you will find all of this useful. Feel free to contact me on Medium.
I am using Substack only for the newsletter. Here every week I am giving free links to my paid articles on Medium.
Follow me and Read my latest articles https://medium.com/@fabio.matricardi