China Sounds the 🚨, Spending, & New Models

TLDR;

Today, we have a newsletter packed with interesting news and insights, from leaks about xAI’s latest supermodel, Grok 3.5, to Microsoft’s new super reasoner, Phi-4-reasoning.

We also discuss Big Tech earnings and whether the AI spending party continues, as well as Sam Altman’s latest venture and the hidden costs of AI models.

Finally, we look at a transgressive research paper from China that challenges the common assumptions on reasoning models, the industry’s latest sweethearts; they aren’t what you’ve been made to believe.

FRONTIER MODELS

Will Grok 3.5 Be an Engineering Beast?

On Tuesday, Elon Musk confirmed the launch of Grok 3.5 next week (although it could be released any time now), making very strong claims about it.

But it seems that a small mistake from one of the engineers has shown that xAI has up to sixty different Grok fine-tunes internally (sixty different versions of the model), with very surprising names like ‘grok-SpaceX.’ What this implies is that xAI is actively training its models on SpaceX, Neuralink, Starlink, or Tesla data, which could give them unparallel advantage on engineering efforts.

This could be a moat itself because some of these companies have quite literally unique data on crucial areas like space, autonomous driving, or even brain-computer interfaces.

For instance, a very recent research paper by Tufa Labs shows how small reasoning models trained on rocket design data can even outcompete not only frontier AIs, but they also outcompete human experts.

TheWhiteBox’s takeaway:

Every day that passes, it’s becoming harder and harder to bet against xAI. Not only have they managed to become a top-3 AI Lab in under a year, but have the necessary capital-raising capacity combined with the force of being part of the ‘Elon network,’ that it’s hard to fathom a future where they don’t completely blow out of the water most labs in terms of ‘raw intelligence’.

But the point here is that this isn’t a big deal as AI is clearly not a winner-takes-it-all industry, and most labs will specialize on specific use cases.

OpenAI appears to have a great lead for everyday, consumer-end questions or tasks and with clear interest on agentic computer use.

Google also has great distribution for more desktop-based search via its AI-enhanced search engine and they are also great at coding.

Anthropic or Cohere seem to be focusing more on enterprise use cases, the former on more agentic-style/coding work and the latter more on enterprise search.

xAI could be developing a really strong presence on engineering use cases (design, prototyping, etc.)

Of course, none have renounced their in any particular use case and remain generalistic solutions, but the end game will probably end up being more specialized.

REASONING MODELS

Microsoft’s Phi-4 Reasoning. A Small Beast

Microsoft has released new open-source models, Phi-4-reasoning and Phi-4-reasoning-plus, accessible through HuggingFace for free download, which exhibit incredible performance at their size, ‘just’ 14 billion parameters. This means they can be run on consumer-end hardware (the bare minimum for this model would be 32 GB…

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