Lightning-triggering Drones

TLDR;

This week, we take a look at the state of frontier AI models, with some unexpected surprises, a breakthrough from a 14-year-old Chinese prodigy, a new hire by OpenAI that speaks volumes about its future, product updates that are really worth considering, and a fascinating way humans intend to use AI for: capture lightning to produce electricity.

Enjoy!

FRONTIER MODELS

Update on the Pareto frontier

The Aider Polyglot team, which evaluates models in what’s considered the prime coding benchmark, has acknowledged a mistake in Google’s cost analysis for Gemini 2.5 Pro, which appears to be higher than what they originally presented.

This means the Pareto frontier is now slightly altered, showing three leaders based on performance-cost ratios. DeepSeek’s models offer the best performance for the cheapest models. For high-end results, Gemini and OpenAI are the leads nonetheless.hey

TheWhiteBox’s takeaway:

An important thing to notice on the graph is that the scale is not linear, making it look like OpenAI’s models are not much more expensive than Google’s despite being better.

But that’s not true.

Both superior scores by OpenAI are around 2x and 3x more expensive than Gemini 2.5 Pro, only to be a couple of percentage points better. Huge cost increases for a marginal improvement, so the best model in that graph (the most Pareto optimized model) is clearly Gemini 2.5 Pro nonetheless.

AGENTS

A Chinese Prodigy Builds A Model… Too Good to be True?

A Chinese 14-year-old boy named Ji Shihao has built a model, Jinmeng 550A, that achieves historical results in two very well-known benchmarks: the AIME 2024 International Mathematics Olympiad and MedQA, a question-answering benchmark focused on medicine, while beating all frontier AI models by basically clearing the benchmarks with 100% and 99.7% scores while being only 3% the size of some frontier models and costing orders of magnitude less.

This amazing breakthrough is based on the idea of neuro-symbolic AI, an often disregarded field of AI that combines the virtues of neural networks like ChatGPT and symbolic engines, human-written code basically (like a set of rules to solve math equations).

But what does that mean?

If you look at the task below, you can clearly see the thought process a human would use.

Understand what the concepts of ‘large things’ and ‘metal spheres’ are,

identify each,

count,

and see if they match.

The issue is that while neural nets are great (above-human actually) at perceiving the different concepts in the image, they struggle with the high-level reasoning part required here (counting and equalization).

This might seem very easy for you but it’s certainly not for a neural net, as the concepts, the maths, and the high-level reasoning idea of comparing both values are all three tasks that have to be done by the AI model simultaneously, which often leads to entanglement.

By default, neural nets do not have procedural thinking as humans do. It’s as if you were forced to identify the objects, colors, sizes and the number of each all simultaneously in the…

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