Elon vs Sam, The Guardians of Knowledge, & More

TLDR;

In today’s newsletter, we review the interesting findings of Humanity’s Last Exam, reflect on the saturation speed and real consequences of AIs breathing such benchmarks, comment on T5’s plans to build five Giga-scale data centers across the US, and discuss Sam Altman’s updated views on AI.

Next, we describe Wiz’s findings on DeepSeek’s poor cybersecurity systems and Elon’s bold bid to buy OpenAI. Regarding the latter, we comment on their announced partnership with Los Alamos National Laboratory.

Finally, we’ll focus on the two hottest released products of the week, Mistral’s LeChat and Pika Labs’ Pikadditions, which have drawn much attention these past few days (and for very good reasons).

KNOWLEDGE INTELLIGENCE

Results of Humanity’s Last Exam

Scale AI has released its internal results after testing several models in Humanity’s Last Exam, a benchmark created by Dan Hendrycks to determine how smart these models are at solving knowledge-related questions that are considered extremely hard even for expert humans. Solving this benchmark would position AI at a point of remarkable wisdom compared to the best of us.

The results are not surprising. o1 scored the highest, reaching 9%, followed closely by the Gemini models.

TheWhiteBox’s takeaway:

It’s important to note that the results shown above are obtained without search capabilities. This means that models answer based on their compressed knowledge and can’t actively search for the results. If so, results increase considerably, with o3 with Deep Research getting 30% only two weeks after the benchmark was released.

It’s also important to highlight that progress in this benchmark indicates how well these models retrieve and provide knowledge to the user. Thus, these models are poised to become our primary source of knowledge ahead of tools like Google due to finding the perfect balance between great knowledge and convenience.

If that becomes true, it’s more important than ever that we fight against proprietary solutions, which could be biased (just like Google is) to show us skewed views of reality. If reasoning models are to become humans’ primary source of knowledge, we better make sure they are open and not biased by cultural biases or driven by economic incentives (for instance, showing results of websites that are paying for higher attribution).

Finally, please stop with the flashy names; this is not humanity’s last exam.

KNOWLEDGE INTELLIGENCE

T5 Announces 5 Giga-scale data centers

On Sunday, I made my case about why I thought AI would continue to be extremely resource-hungry, leading to growing spending and model sizes. In particular, we discussed the potential size of the next model generation’s training budget, which we estimated to be at least 10 times the size of the previous generation (GPT-4 et al.) and even speculated with 100x training runs.

For that, we estimated that we would need giga-scale data centers, eight-to-ten times larger than the largest known cluster, xAI’s Colossus, and we commented on how all Hyperscalers had such data centers in their plans before the end of 2026.

Now, T5 has announced that none other than five…

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