AI & Fusion, OpenAI's Atlas, & More

TLDR;

Welcome back! I apologize for not having sent the email yesterday, but OpenAI’s sudden release of Atlas caught me off guard. Besides this release, we also cover NVIDIA’s increasing relationship with OpenAI, whether AIs can make money trading, how ChatGPT is losing ground, and other interesting news.

Finally, this week on TheWhiteBox Portfolio company deep dive, we analyze Oracle, a play that will either make a lot of people filthy rich or filthy poor.

Enjoy!

TRADING

Are AIs Good Traders?

A new benchmark, called Alpha Arena, has gone extremely viral over the past few days. A group of frontier AI models is funded, and the benchmark evaluates whether these models generate profit in the Crypto perpetuals market with the goal of maximizing risk-adjusted returns.

Crypto perpetuals are a type of derivative that today dominates Crypto trading (hence the abnormal volatility, even by Crypto standards, we are seeing lately), where people bet on whether the price of a cryptocurrency will be higher or lower than its current value at the time of the contract. Speculation on steroids, basically.

At the time of writing, DeepSeek v3.1 seems to be in the lead, generating around $1k of unrealized profit, with Qwen barely making a profit and all other LLMs losing a lot of money, and with OpenAI having lost around $7k.

It has made rounds because of the obvious repercussions: can AIs make us win money? What to make of this?

TheWhiteBox’s takeaway:

I’m sorry to say, but the benchmark is flawed. What you’re seeing is just noise, meaning these results are simply a result of chance. In other words, if we restarted the benchmark, we are likely to see something totally different.

Instead, the correct way of doing these benchmarks should have been testing the same model multiple times, to see if we can identify an intra-model pattern that really shows a tendency to make money.

Besides, these models are not only non-deterministic, meaning there’s always randomness added to every prediction (even if we set the model to greedy decoding, their underlying infrastructure is non-deterministic because of the non-associative nature of floating numbers, as we saw recently in this newsletter), but they are also trained under totally different data distributions, sets of biases, and such, meaning we have a huge set of confounding variables (variables that may be playing a role) we aren’t controlling.

Instead, to see meaningful results, these benchmarks should execute multiple trial runs to see a pattern emerge (i.e., if DeepSeek wins 100 of 100 times, there’s definitely something there), models should be tested under uniform prompts (although that would put some models at a disadvantage already), and many other things that are not being taken into account.

THEWHITEBOX

ChatGPT is losing share of AI Search

Although ChatGPT’s new product release (more below) might change this, in a study by SimilarWeb, it is claimed that ChatGPT has been steadily losing market share in the AI search market over the last twelve months, primarily to Google’s Gemini.

It’s important…

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