[Premium] China Does it Again...

TLDR;

Technology:

🤯 China Does it Again

😩 o3’s Results… Not Quite What they Seem?

🔮 Project Stargate & The Opportunity to Cure Cancer?

Markets:

🤔 Mistral’s Road to Liquidity, an IPO?

🥵 Meta Opens the Door to more “AI-Led” Layoffs

🤨 MBAs are Struggling…

Product:

🤧 Are Intel’s GPUs Any Good?

⚡️ Grok-3 Finishes Training

🫢 OpenAI’s Operator, This January?

FRONTIER RESEARCH

China Does It Again

One could make a very solid case that China’s DeepSeek AI lab is the hottest in the world right now, even ahead of all US/UK/EU ones. Time and time again, they continue to publish open research that is not just good but literally the most advanced research in the space.

On Tuesday, they finally revealed DeepSeek r1, their first reasoning model, and the entire industry is raving about it.

The model is an o1-level model, as good as OpenAI’s second-tier Large Reasoner Model in all tasks and, crucially, is vastly cheaper to run. While OpenAI’s o1-preview model is priced at $15 per million input tokens and $60 per million output tokens, DeepSeek’s V3 model, which underpins R1, is priced at $0.90 per million input tokens and $1.10 per million output tokens.

TheWhiteBox’s takeaway:

This release is full of so many amazing breakthroughs (no SFT training, no process verifiers or search, emerging reasoning behaviors, etc.) that break so many intuitions everyone, including myself, had simply assumed just because OpenAI did it, that I will talk about it in detail in our Trend of the Week coming this Thursday. Stay tuned because many things you took for granted in AI won’t hold true in a few days.

In a way, DeepSeek is what OpenAI should have been all along: the best open-source AI lab.

FRONTIER RESEARCH

o3’s results, not what they seemed?

Weeks ago, OpenAI shocked the world with its results on FrontierMath, considered the most challenging math AI benchmark ever built. However, it was recently revealed by EpochAI that these results have more than what meets the eye.

It turns out that OpenAI funded the development of the FrontierMath benchmark. The funding was only disclosed when OpenAI’s o3 model set a new performance record on the test. In layman’s terms, OpenAI has had access to the training dataset of FrontierMath for quite some time and, importantly, were involved in its design.

While they promised EpochAI they wouldn’t use that data to train the model, two considerations arise:

EpochAI can’t verify whether OpenAI stayed true to its word. OpenAI has every incentive to use it, and we don’t have any way of proving otherwise.

Part of the agreement between both parties was that EpochAI would have an independent, contamination-free held-out dataset that would check whether the results were real.

And while training on the training dataset does not invalidate your results, point 2 above is tricky.

TheWhiteBox’s takeaway:

Show me the incentives, and I’ll show you the outcomes.

OpenAI had every reason to train its model in that dataset, and we know by their own…

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