Introduction

I’m really into machine learning advances of all sorts, and keep tabs on research from across the planet.

I saw a paper that was fascinating to me, [https://arxiv.org/abs/2401.02994]Blending Is All You Need: Cheaper, Better Alternative to Trillion-Parameters LLM (2024).

This particular paper came from the fertile minds at the University of Cambridge, University College London, and a fun team called Chai Research.

What they found, essentially, through actual user testing (not just theory)…

Users engaged more and had higher retention rates when they talked to a mix of low-end large language models, varying from 6 billion to 13 billion wparameters…

…versus ChatGPT 3.5…

Which uses 175 billion parameters.

That’s a blend of models up to 30 times smaller…

Getting a better result, for most users of their product.

I take this concept in BrightRobot to the next level - the models I blend for each of my blends are up-to-and-including the newest, best-performing version of GPT-4, GPT-4o, which dominated the LMSYS Chatbot Arena when it was prerelease under ‘im-a-good-gpt2-chatbot’.

Even the Instant blend that we have as our lowest-end blend is a mix where each model is stronger than ChatGPT 3.5… in isolation, while also being much quicker.

…And that’s what I’m making freely available.

What we have in the Fast HQ blend is all at least comparable to GPT-4, and everything in our HQ blend is much stronger, and even includes the newest iteration of GPT-4 itself, GPT-4o.

What does this mean for you?

It means better answers, a more engaging chat companion, and more varied, interesting responses.

It also means faster responses, which is a big deal for me, since it saves me time to work on more important things.

In other words - the blends do more for me, and for you, faster.

Cool? I thought so, so hopefully you do also.

And remember, if this “Blend” seems too crazy to you, just pick your preferred model (of many companies - OpenAI, Anthropic, Cohere, Mistral AI, and many more) and get on with your day..

To my knowledge, this is the first implementation that blends what’s called “frontier models” aka… really powerful models… automatically, to get you a better result.

I expect this to be copied by other interfaces soon, because…

Well, it’s just better.

And don’t we all want better, where we can get it?

Drop me a line at [email protected] with your feedback as you use them, I’d love to hear what you think.