GPT-4.5 Unhinged Mode, Anthropic banks again, CoreWeave's IPO, & More.

TLDR;

This week in our Premium news rundown, we discuss recent frontier leaderboards updates, focusing on GPT-4.5 and its ‘unhinged’ new personality. We then discuss a new prompting method that could save you a lot of money and Google’s latest agent tool for Data Science.

We then move into markets, discussing CoreWeave’s upcoming IPO, Meta’s insane $200 billion data center plan, and Anthropic’s latest billion-dollar series E.

Finally, we discuss three cool new tools to create better presentations, better—and branded—graphs and visuals, and the latest free Deep Research tool.

Enjoy!

FRONTIER AI

GPT-4.5 Hits the First Spot… For a Day

On Sunday, based on the OpenAI-published comparison of GPT-4.5 and GPT-4o win rates, we predicted GPT-4.5 would win first place in the ELO-based LMArena. For a couple of days, this prediction was officially true, until xAI released its definitive version of Grok-3 and surpassed GPT-4.5 by just one ELO point.

The fact that they are similar in performance is not surprising considering that both were trained in the same order of floating operations, around the 200 trillion trillion operations mark.

Still, it just goes to prove that claims that the model sucked were way exaggerated.

TheWhiteBox’s takeaway:

But I do have to say that GPT-4.5 feels different from all previous models.

I’m beginning to understand what Sam Altman was referring to when he said the experience would make you “feel the AGI” (this is a stupid way of saying that robots are becoming more human, but we both know that’s not the case; they are just becoming better at imitating us).

But I kind of… get what he meant?

It definitely understands the nuances of the user’s requests better and adapts to the “vibes” of the user. More interestingly, it’s pretty much ‘unhinged,’ as you can see below:

Whether or not you enjoy those unhinged posts is not the point. The point is that the model better captures the user's intent, whether it’s to match the sarcasm, cynicism, or whatever and play along; it's definitely a more human experience, even if I hate to utter those words myself.

PROMPT ENGINEERING

Chain of Draft, A Solution to API Costs?

Source

Researchers from Zoom have published an interesting approach to prompting reasoning models that reduces costs dramatically.

This approach performs similarly to chain of thought (CoT), when the model generates all the intermediate steps to solve a problem, helping it reason out the answer, but requiring around 8% of the necessary tokens.

The proposal, known as Chain of Draft (CoD), helps reduce the cost of running reasoning models, which are often unnecessarily verbose, making them more concise and succinct in their responses.

For instance, for a company doing 1 million reasoning queries a month, switching from CoT to CoD saves them around $3,000/month, with no apparent impact on performance.

Source

TheWhiteBox’s takeaway:

The impressive cost reductions and efficiency derived from a simple modification (no further training required) have caused quite some noise in the industry these past days.

However, while you can…

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