AI is all about the money. Or the lack thereof.

TLDR;

Welcome back! This week, we have incredibly interesting research from Anthropic and Alibaba, why you should care about CoreWeave, Samsung’s historical quarter, and several cool products from the big AI Labs you might want to check out.

Enjoy.

In the last newsletter, I sent you a Medium article with extended coverage on Mythos that I claimed was allegedly free. However, I clicked the wrong button and sent you a link to the paywalled one (Medium is paywalled by default), which a subscriber diligently pointed out. Here’s the link to the article, this time truly free (and still very much relevant). If anyone runs into issues, please let me know. My apologies for the mistake. To make up for it, here’s the link to another Medium article I published today that goes into great detail as to why AI prices are about to shoot up into the sky.

FRONTIER MODELS

Anthropic’s Advisor Models

In one of the coolest releases in a while, Anthropic has announced a new way to improve performance-per-cost in its system using the advisor strategy, in which a weaker model (and thus cheaper) can receive advice from a superior model on demand.

The idea is that this avoids having to use the stronger, more expensive model at all times and having the cheaper model call it as an advisor when needed. And as you can see below, it works.

TheWhiteBox’s takeaway:

I always say that the best ideas are the ones that sound obvious. This is one. Pretty smart.

But before you try this on your internal local workload, the sad truth is that this only works if you train the whole system. In layman’s terms, this isn’t just simply giving your model a smarter model as a tool it can call, but the model was trained to do that.

Why? Because models are smartasses; you know that. They don’t know what they don’t know, and they haven’t been trained to say “I don’t know,” or at least they’ll leave that option to the last possible resort.

This means that having a model learn when it needs to ask the advisor is something that has to be trained.

Luckily, this very recent research, which seems eerily similar to what Anthropic has released as its own (AI Labs, not just Anthropic, but all of them, have a bad tendency to not give credit), explains how this works.

Without going too much into detail, the idea is to “breed” this behavior into your model by giving it a task it cannot solve without the advisor. Eventually, the model will “succumb”, try the advisor, and see that now it actually solves the problem. This reinforces this behavior until eventually your model learns to “know what it doesn’t know.”

VIDEO GENERATION

China is dominating video generation

HappyHorse is the name of a new video-generation AI model that appeared almost out of nowhere and immediately climbed to the top of Artificial Analysis’s rankings, overtaking several better-known rivals in both text-to-video and image-to-video…

Enjoyed the introduction? Subscribe or sign in to continue with this edition.